Well, that became my own personal odyssey this week. And since The Odyssey is a cultural phenomenon right now, Greek mythology references are in vogue again. I had to jump on that bandwagon.

(And yes, I know the “thousand ships” line was not Homer. It was Christopher Marlowe.)

This piece explains that odyssey, in more detail than I expected to have. Preparing for the R2T launch turned into a guided tour of the landmines you have to find before any official backtest gets run.

Why do this? I want this publication to be honest and rigorous, and rigor doesn’t stop at the strategy math. It applies to the data underneath, or none of the rest means anything.

A manager of mine once called work like this “breaking rocks.” I call it “clearing the data mud.” This work is not sexy. It is not fun. But it is absolutely vital for credibility.

Share

Refresher: Survivorship Bias

A quick refresher, since everything downstream depends on it. Survivorship bias is what happens when you evaluate today’s index members historically. You have quietly excluded every company that died, merged, or fell out along the way, and the exclusion isn’t neutral.

Survivorship bias flatters results in exactly the direction that gets strategies published, because the names that survived are, by construction, the names that did well enough to survive. Fixing it requires point-in-time membership records, a bar-by-bar account of who was actually in the index, delisted names included. That requirement is where all the complexity in this piece comes from.

Removing survivorship bias doesn’t remove problems. It trades easy problems for honest ones. That trade is this piece.

Working with historical data: the price of admission

Mid- and small-cap daily history is a new dataset for Alpha on the Edge, and I knew from experience what was coming. This data is messy, and that’s nobody’s fault. Managing complex market datasets is extremely difficult. I know because I’ve had to manage small corners of them, and the corners alone taught me humility.

So you will not read any vendor or index-provider complaints here. Imperfect data is the normal course of business. What a practitioner gets is a choice, made over and over. Handle each issue with a judgment call, or pass it through untouched and make it the next person’s problem. What follows is a tour of those judgment calls.

The most obvious landmine is the price spike, a print wildly out of line with its neighbors that makes volatility explode. Spikes are dangerous for backtests because volatility feeds everything downstream. Position sizing, stops, and Sharpe ratios all inherit whatever the vol estimate believes. The trap is that some spikes are real.

As many of you will remember, GameStop in January 2021 was a genuine roller coaster, and you could genuinely have traded it. Twenty-fold in under a month, then most of it given back within a week, all while sitting inside the S&P 600.

Others are stranger. The chart below shows a single day in early 2000 where the adjusted price series prints a +270% return while the actual tape recorded an 80% loss. The event underneath was real, a spinoff distributed to shareholders and worth real money. But the +270% never traded. Nobody bought at the open and tripled their money by the close. It’s what happens when a corporate action gets folded into the back-adjustment seam at the wrong scale, and a pipeline that consumes adjusted prices without ever checking them against the unadjusted tape will book that day as the trade of the decade.

A screen that deletes everything weird deletes GameStop. A screen that keeps everything weird keeps bars that never traded. You have to know which is which.

Onboarding this dataset meant a lot of scrubbing. Walking bar-level quality flags, deciding which bars were eligible to trade and which weren’t, building screens so the next bad bar gets caught by machinery instead of luck. I’ll spare you the plumbing details. The point is that this work happens before a single backtest is worth believing, and it never shows up in anyone’s results table.

It also happens at scale. The universe behind R2T covers 4,324 instruments across US large-, mid-, and small-cap stocks since the 1990s. A single configuration of a single mid-cap backtest run touches 6,550,068 daily bars, and each run carries ten walk-forward configurations. The full-history small-cap run alone executed 17.1 million simulations. Across every member’s quoted history, the panels hold 20,262,375 daily bars.

We’ve handled the scrubbing. So the data’s ready, right?

Right?

The chart that launched a thousand ships

Universe-aware backtesting adds a second layer of decisions on top of price hygiene. Cleaning the bars gets you time series, but it doesn’t get you the time periods to actually use those time series.

So you have to draw borders, and I designed R2T v2 with a clean border definition. A universe member is an index member, by the point-in-time record.

Then, I made a mistake that is obvious in hindsight and was invisible until one chart made it visible.

The mistake lived in the walk-forward split geometry. I calibrated it the way R2T v1 did. It was blind to point-in-time universe constituency, tiling evaluation windows over each stock’s full quoted price history.

Meanwhile, I added universe membership as a gate to tradable bars. The idea was that bars outside membership windows were not allowed in the OOS splits.

Two rulers, measuring two different date ranges. The splits were measuring a range the strategy was never allowed to trade.

Every gate I built kept every promise it was given. The promise that evaluation windows would be bounded by membership was only carried out at trade simulation time, not at split-determination time.

When I ran the R2T v2 dress rehearsal, that gap surfaced as one flat line.

This is why you run dress rehearsals, and why you look at the results instead of filing them. When I first saw this chart, my first guesses were wrong twice. I suspected bad prices for the flat stock. I suspected warm-up leakage for the flat portfolio start.

The chart didn’t care what I suspected. Visualizations don’t lie. They just wait for you to catch up.

Here is what the hero chart shows. The membership gating worked exactly as designed. Vlasic was force-exited on removal day and never traded again. But the measurement didn't stop. The performance record kept the name on the books in cash for 222 more bars, and those bars flowed into the return stream like real trading days. All of this because the WFO splits were not configured on the same dates.

So I went digging, face in my palm most of the time. The problem was so obvious. Look at another manifestation. Return history in this dataset began before index inception, but the WFO splits didn’t know this. So we get an abomination like this:

That kind of exactness is a gift, because noise doesn't line up with anything. It tells you the problem is structural.

So what is the flat line actually showing? Every price on the chart is real, so it isn't a data error. No position was ever held past removal day, so it isn't a trading bug. It's a measurement bug. The simulation kept retired stocks on the books in cash, and those cash bars flowed into return streams, volatility denominators, and Sharpe ratios. An ugly stream of zeros, dressed as evidence.

Once you know the padding exists, the next question is how much of the run was padding. The answer deserved its own chart.

The census puts numbers on the damage. Of the 6,550,068 bars in the US mid-cap run, 60.8% were exactly zero, and 93.3% of those zeros fell outside membership. Only 41.7% of all bars were a member actually trading. One hundred forty-four instruments had zero overlap between their membership and their evaluation window, and still produced full return streams and full rows of metrics.

The distortion isn't hypothetical. Re-measure Vlasic over a membership-clipped window with the identical trades and its annualized return goes from 55.9% to 120.8%, its Sharpe from 1.99 to 2.69, without a single trade moving. Pure denominator distortion, and it runs in both directions across the portfolio. The aggregated equal-weight line was diluted roughly 2.4×.

About the "we checked twice" in that chart title. Before believing any of this, I re-executed a single symbol bit for bit outside the pipeline to prove the chart was the simulation and not a rendering artifact. It was the simulation. That run is now retired, kept as an exhibit of shame, and everything was re-run.

You’d think “in the index” would be simple

If you haven’t worked with membership data, you’d assume it’s easy. One day a stock joins, one day it leaves, two dates, done.

That is unfortunately not the case, which is the polite way of saying this part makes you want to pull your hair out.

None of what follows is a complaint about the data. The data faithfully records a complicated world. The complications are the world’s.

Consider Cleveland-Cliffs, which has been in the S&P 1500 for 35 years, provided you don’t ask which index. Seven separate stints across three different indices, plus four and a half years outside all of them. One ticker, and every one of those transitions is a decision a backtest has to get right.

Then there's flicker. The raw membership panel for one name blinks out for a day or three at a time, over a hundred times across its history.

A three-bar dropout is not an index exit in any economic sense, so gaps of five bars or fewer get coalesced into a single stint. That's one of the aforementioned judgment calls, made once and applied everywhere. It's also stable. Thresholds of 21 or 63 bars produce identical results, so nothing hinges on the “5 bar” measurement.

This type of judgement call is extremely important. Where did the earlier plague of 0s come from? From here:

An honest backtest has to know these things.

The median longest stint in the S&P 400 runs 5.9 years. Roughly a quarter of ever-members never complete three years. True rejoins, the Cleveland-Cliffs kind, affect 7.2% of S&P 400 members, and real rejoin gaps run five to eight years while flicker gaps run one to three bars, so the two are not remotely the same phenomenon.

Borders, careers, and revolving doors. That is what "in the index" means, and it's what the walk-forward engine had to learn.

Teaching walk-forward about borders

Walk-forward analysis assumes a stock is a continuous ribbon of history. Tile the ribbon into in-sample and out-of-sample windows, march forward, repeat. Membership turns the ribbon into segments, and every segment raises the same questions. Where may evaluation start? What funds the first in-sample window when a stint is young? What happens to a position when the border arrives mid-trade? What do you do with a stock that has four stints?

I went much deeper on all of this than this piece needs. Here, I’ll outline decision highlights, because they matter.

• Each qualifying membership stint is its own evaluation domain. OOS windows tile inside the stint, and the domain ends one bar past the last member bar, so a position that’s open at removal realizes its forced exit honestly.

• No OOS window may start before its stint does, and no OOS window exceeds 40% of the total split.

• Parameter selection may look back before the stint begins, using trailing, quality-screened tradable history. That replicates what a practitioner would actually do the day a stock joins an index.

• Every qualifying stint is evaluated, not just the longest. Evaluating only the dominant stint measured as flattering, +23.65 percentage points per year on the affected names, so that design died. Re-entry is never automatic. Each stint starts flat and waits for a fresh signal.

• Any benchmark lives under identical windows. Whatever discipline the strategy accepts, its comparison bar accepts too.

No ideal solution exists in this design space. These are the accepted trade-offs, stated plainly so you can disagree with them specifically.

What the new geometry does is easiest to show. Four stocks, four shapes of the problem.

• Vlasic has a single stint too young to fund itself, so the front of the stint pays for selection, 141 bars of it.

• Cleveland-Cliffs has four S&P 400 stints and gets a split budget of 3, 1, 1, and 2 in a 7-split WFO configuration.

• Celsius joins with enough trailing history to fund selection immediately, at a cost of exactly two bars of rounding.

• Apple was a member before the panel even starts, so its front is censored and roughly eleven years of history serve selection only.

The blue reaching back across Cleveland-Cliffs' gaps is trailing history funding selection, and it's annotated on the chart precisely so it can't be misread as look-ahead leakage.

What was all this for?

The fix is easiest to judge on the chart that started everything.

The rebuilt run holds Vlasic for 350 bars and ends where the position ends, on the forced-exit fill one bar past membership. No padding tail. It also starts three months earlier than the retired run, because parameter selection now draws on trailing history instead of burning the front of the stint. The flat line is gone in both directions.

A claim like that deserves a receipt, so the same census ran on both runs.

Exactly-zero bars fell from 60.8% to 12.6%. Membership exposure rose from 0.44 to 0.86. The run shrank from 6,550,068 bars to 3,164,547, and every bar that left was a bar that should never have been measured in the first place.

The thing about that flat line is that everything downstream of it was working perfectly. The math was right and the gates kept every promise they were given. The pipeline was even fast.

But the measurement was wrong, and only a dress rehearsal and a pair of eyes on a chart could have caught it. That is the whole argument for how this publication operates. Run the tests. Look at everything. Publish the screens and the costs alongside the results.

Data work is breaking rocks. It’s two hands in the mud, a dirty job of the 21st century. But it has to be done.

What launches next is the rebuilt, membership-aware R2T v2, and the first official results will ride on it.

A thousand ships have been launched.

Until next time, keep on the cutting edge, everyone.

Disclaimers

The content on this page is for educational and informational purposes only. Any views and opinions expressed belong only to the writer and do not represent views and opinions of people, institutions, or organizations that the writer may or may not be associated with.

No material in this page should be construed as buy/sell recommendations, investment advice, determinations of suitability, or solicitations. Securities investment and trading involve risks, and not all risks are disclosed or discussed here. Loss of principal is possible. You are encouraged to seek financial advice from a licensed professional prior to making transaction decisions.

Further, you should not assume that the future performance of any specific investment or investment strategy will be profitable or equal to corresponding past performance levels. Past performance does not guarantee future results.

Technology Note

This article was written with LLM assistance. My typical writing workflow starts with voice transcription via a service like Super Whisper or WisprFlow. Next steps involve grammar checks and line edits with assistance from services like Claude.ai or Grammarly. For sections that are difficult to word correctly, I will ask an LLM for a draft. I always re-write it because LLM prose is almost always mediocre and boring.

At the time, I wrote that this kind of content was at best misleading and at worst outright fraud.

I stand by that.

In fact, the problem has gotten worse since I wrote it, because the ability to run large-scale backtests keeps proliferating, and easier backtesting makes cherry-picking easier. When you can run a thousand backtests in an afternoon, the spectacular curve you inevitably find proves nothing beyond your patience and your ability to use a max function.

R2T v1 was my answer to that. Instead of showing one hand-picked curve, run millions of simulations across entire universes of securities and show the whole distribution of outcomes, including the ugly parts. I was and am proud of that. It planted a flag in both methodology and ideology.

However, I always knew R2T v1 had limits, and they bothered me. So I rebuilt the whole thing.

This post is a preview of R2T v2: what changed, why, and what’s coming. The short version is that v2 doubles down on the lane I want to occupy: honest, transparent results, published whether or not they flatter a strategy.

Share

From distributions to verdicts

V1 gave you distributions and let you judge. That was already a big step past cherry-picking, but it left the most important question hanging in the air: does this strategy actually work? I could show you a cloud of out-of-sample Sharpe ratios, and you still had to decide for yourself what it meant.

V2 answers the question directly. Every evaluation now produces a verdict: supported, mixed, not supported, or, when the evidence is too thin to say, insufficient evidence.

Each verdict attaches to a pre-registered claim rather than a vibe. The claim has a precise scope: this signal family, under this trading policy, across this universe, over this time period, under this version of the methodology. Change any piece of that, and you have a different claim with its own verdict. There is no “this strategy is good,” only “this specific claim, tested this specific way, earned this specific verdict.”

Why is this more than simple labeling? Two reasons:

1. The criteria are set before the results exist. The thresholds for “supported” are written down, versioned, and published. If I ever change them, the change gets its own version number, a changelog entry, and a side-by-side recompute. I cannot quietly move the goalposts after seeing a result I don’t like.

2. History is append-only. Nothing already published ever changes retroactively. Re-runs on new data or new criteria become new result sets that link back to what they supersede. This makes the entire trail transparent.

To be clear about what a verdict is not: verdicts never sort into “best strategy” lists, and no verdict is a recommendation. A verdict is evidence of effectiveness stated consistently, so that when you see “supported” or “not supported,” you know exactly what standard of research was applied.

Expect to see “not supported” a lot. As we saw with v1, Robust walk-forward testing makes most popular strategies look much worse than their proponents suggest, which means this is going to print a great deal of red.

I am perfectly okay with that. A methodology that can’t say no is just marketing.

The upgrade I’m proudest of: no more survivorship bias

V1 had one limitation that made me grit my teeth every time I disclosed it, and I disclosed it every time. I could only test current index constituents. That’s survivorship bias, and it’s not a small problem. When your universe consists only of companies that survived long enough to be in the index today, you’ve silently excluded every company that failed or got taken out along the way. Your backtest is auditioning on a stage where the failures were removed before the curtain went up.

V2 fixes this at the data layer. I now use constituents back to the 1990s: every company that was ever a member, their full price series, and the exact windows when they were in the index. Backtests only trade names during their actual membership windows. Delisted companies are in the data, trading right up until they stopped trading.

Every R2T result from v2 onward is free of survivorship bias. I have not found another publicly available resource doing consistent, large-scale, survivorship-bias-free strategy evaluation in this format for retail readers. This is the kind of data work that normally lives inside institutional research desks, and building it was most of the reason v2 took as long as it did.

And because I know some of you are wondering how much survivorship bias actually matters: a follow-up piece is coming where I run the same methodology on both datasets, current constituents only versus full historical membership, and show you exactly how the results move.

The data gremlins

A dataset this deep comes with a huge number of small horrors, and edge cases are exactly where backtests lie most easily. So a lot of v2’s substance is unglamorous consistency work. For example:

• Tradable versus anomalous bars. Old data is full of judgment calls. March 2020 and the meme-stock era produced unbelievable volatility spikes that were nonetheless real, exchange-printed prices, and those stay. A dollar stock with a random 400% wick on no volume in 1998 is a different story. After an exhaustive amount of exploratory analysis, v2 applies consistent data-quality rules that exclude untradable bars from entry eligibility, so performance reflects what could actually have been traded at the time.

• Windfalls that don’t certify. If a strategy happened to be holding a name that got acquired at a premium, that profit is real and stays in the portfolio results. But two lucky trades on one name don’t get to certify a claim about breadth. Trades that thin contribute to P&L, never to verdicts.

• Results that round to zero. In experimental runs, one test counted a universe-median Sharpe ratio of +0.000184 as a positive result, which is a generous name for zero wearing a costume. The registered criteria now require a minimum margin before a number that small counts as evidence for anything. I would rather report weaker-looking results than let a Sharpe ratio of 0.0002 quietly pad a strategy’s case.

• Short-side accounting. Unstopped short positions on collapsing names can produce mathematically undefined equity paths. The verdict-standard strategy wrapper includes a static stop for exactly this reason: a minimal guard that keeps the accounting well-defined, with nothing more claimed for it.

None of these choices is exciting on its own. Together, they make a claim far more robust and trustworthy.

What’s coming

The destination for all of this is a scoreboard: strategies as rows, universes as columns, verdicts in the cells. Small-, mid-, and large-cap US equities, FX majors, core ETFs, and commodities, all held to the same registered standard at once. At a glance: what works, where, and with how much evidence. Every cell carries its scope and provenance, so every verdict is transparent.

There are plenty of places to browse backtest results. This is meant to be something different: a consistent, versioned, survivorship-bias-free evaluation standard, applied at scale, that is allowed to tell you a strategy doesn’t work. Because most of them don’t, and you deserve a source that will say so.

The first verdicts are coming soon. Most of the cells are going to be red. The interesting part is finding out which ones aren’t.

Until next time, keep on the cutting edge, everyone.

“Yeah, but it’s all hype.”

“Yeah, but it can’t get enough right.”

I’ve spent the last several months finding out for myself, and the moment it crystallized was mundane: I was at a coffee shop, on my phone, using Claude Code remotely to add purged walk-forward cross-validation to a backtesting pipeline, with an interface flexible enough to swap protocols quickly. Done before I finished the coffee.

The yeah-buts aren’t entirely wrong. They’re just aimed at the wrong part of the work.

Share

The bigger build I’ve been iterating on is an agentic stock analysis workflow. Five pairs of analyst agents cover factor values, sector data, sentiment, technicals, and fundamentals, with each pair running competing bull and bear cases. An orchestrator coordinates them; a decision layer reconciles their outputs into buy, hold, or sell. Underneath it sits MARC, a research assistant with access to a 2,500-paper corpus embedded in a vector database, accessible through an MCP server. That infrastructure layer would have been a multi-month team project a few years ago. I built it in weeks.

The speed gain isn’t the interesting part. Where it comes from is.

The middle of the research workflow has compressed dramatically: writing code, wiring infrastructure, running sweeps. The two ends haven’t. Setting up the question, defining what a meaningful result looks like, deciding whether an apparent finding is real or an artifact — all of that takes the same time it always did. Sometimes more, because there’s now more output to sift.

So the researcher’s hours don’t disappear. They redistribute: less time in the middle, more at the start framing problems precisely enough for an agent to execute, more at the end scrutinizing what came back.

The end is where this matters most, because AI’s default disposition is generous. Here’s what that looks like in practice. Evaluating backtest results, a model would correctly label median Sharpe ratios as positive, but unless prompted, it wouldn’t exclude the ones that weren’t meaningfully different from zero. The result was a stream of presumptive positives, and when I looked under the hood, many were backed by median Sharpes indistinguishable from noise. A less suspicious reader gets a pipeline that manufactures “promising” strategies. Getting measured evaluation out of the model requires careful prompting and constant calibration. The judgment has to be supplied. The model executes against it.

That’s the pattern across the work. AI compresses tasks where the path from intent to output is mostly mechanical: syntax, scaffolding, well-specified transformations. It doesn’t compress knowing whether a sample is too small, whether a backtest is overfit, whether a result generalizes. Those calls take domain knowledge built over years, and they’re exactly the calls AI is least able to make on its own. If anything, that expertise is worth more now, because every error a researcher doesn’t catch compounds across far more output.

Fifteen years inside institutions also taught me what a solo setup can’t replicate, and why those things exist. Clean point-in-time data and reliable corporate actions are genuinely hard to assemble alone, and the gap shows up as caveats in any analysis. Risk review, peer challenge from colleagues who will push back hard on a thesis, capital to test ideas at scale: AI doesn’t remove those bottlenecks, and I wouldn’t want it to. What I’ve built is good for building, iterating, and learning.

Human-in-the-loop isn’t a token gesture. The human is more leveraged than before — one researcher can drive what used to take a small team — and the parts of the work that didn’t compress are the parts that determine whether the output is worth anything.

Until next time, stay on the cutting edge, everyone.

Instead of scrolling social media or reading the news, I spent those two hours coordinating a codebase reorganization from my phone, using Claude Code to direct Fable as it orchestrated the work. We’ve heard almost a year of hype about AI handling execution, about shipping features from coffee shops. This was different. An entire codebase reorganization is judgment-heavy, and I ran it from a jury holding room without supervising anything, because it wasn’t supervision. It was management.

That sounds like a small distinction and like overblown hype at the same time, depending on who’s reading it. In my experience, it was neither.

Share

The experiment

One rule governed the sprint: Fable never writes final code. It plays architect and orchestrator, and the implementation goes to Opus and Sonnet acting as the engineering team. I play the principal. Set direction, make the calls only I can make, review the work.

And I did nearly all of it by voice. No prompts refined over ten drafts (I’ve written enough of those for a lifetime), just spoken instructions, dictated the way you’d brief a senior colleague in a hallway. Rambling and imprecise, full of the shorthand you use when you trust the listener to fill in gaps.

Voice input is a minefield for LLMs. They fail in a specific way. When we speak, tone of voice carries part of the message, and we count on the listener to receive it. When I first started using voice input with AI, I learned fast that I had to compensate with declarative sentences, a whole lot of them. That’s prompting 101, but it’s not human voice 101, and I had to teach myself to do it.

When I question assumptions out loud, I ramble and hedge, which is natural for me when I’m thinking something through verbally. With LLMs, I had to learn to announce it: this is me working it through, this is an unstructured brain dump, help me think. What I found was that Fable needs those announcements less. Prompting 101 is still prompting 101, but Fable infers intent from text alone far better than anything before it, and for voice input, that’s an enormous benefit.

The moment I knew

Early in the sprint, I dictated a memo with an offhand claim about how a transformer architecture had behaved in my past research. I said “majority,” as in: this pattern held in the majority of cases I’d seen. Experienced judgment, not a citation.

I know what most models do with a sentence like that, because I’ve lived it. They flag it as unfounded, hedge everything downstream of it, and you spend the next three turns defending your own experience to your own tool. The skepticism is healthy. The loop is exhausting.

Fable skipped the loop. It didn’t challenge the claim and it didn’t swallow it either. It wrote tests to check the claim empirically, delegated them to the junior models, reviewed the results, folded the verified finding into the plan, and documented the whole thing so I could audit the call later.

A junior challenges your judgment. A sycophant accepts it. A senior tests it. Resolving doubt with evidence instead of rhetoric is what you hire senior people for, and I watched a model do it without being asked.

It kept happening

A few days later, same pattern, different setting. I was doing exploratory analysis on output from a model I’d just trained, my first look at results from an unfamiliar system.

One return number came back around negative 400 percent. Anyone knows that’s an impossible result, but not Opus. Opus, running the same analysis, reported the number faithfully and moved on. When the data turned out to be riddled with missing values, same thing: 40 percent null, stated as a finding.

Fable’s reaction was closer to mine: that looks wrong, let me dig. It treated the impossible number as a symptom of a bug and went hunting for the bug. The nulls got the same treatment.

Here is where I argue against my own headline for a moment, because the nuance is real. Everything Fable did there is promptable with models like Opus. I could teach Opus to sanity-check outputs and treat anomalies as symptoms, and part of learning these tools is learning to teach them. But the teaching is the supervision overhead. With previous models, competence is something you install, correction by correction. With Fable, the reflex that impossible numbers mean something is broken comes standard. The whole difference lives in the defaults, and defaults determine whether you’re operating a tool or briefing a colleague.

What actually got built

Six days produced two substantial things.

The first was a migration I’d been dodging for months. My personal research platform, years of accumulated code and data, needed a ground-up reorganization. Projects like that never happen. The activation energy is enormous and the payoff is invisible until the very end.

I opened by asking Fable the question I’d put to any senior consultant, which is not how to do the migration but whether to do it at all. What’s the payoff, and what breaks if we skip it? The case it made was concrete. It compared the legacy dependencies against updated ones and worked out what the updates would buy in speed and security. For speed, it used my own throughput scale: this upgrade lets you run this many more analyses every week, across parameter sweeps and model runs. That speedup was pretty tremendous.

And then, there was the benefit I hadn’t fully articulated until we discussed it: the migration would make the codebase AI-native. Not so I’d stop writing code, but so coding agents could crawl it and work on it alongside me. The legacy codebase had no structure for that, for it was never planned with AI help in mind. I think that was the biggest benefit of all.

Then, Fable ran the project. It planned the migration in phases, dispatched implementation to teams of junior models (sometimes several working in parallel on separate parts of the codebase without stepping on each other), and validated everything against the old system. The data side was the easy part: about 6,000 files, already well structured, came over cleanly.

The code was where things got questionable. I had unorganized Jupyter notebooks from years of prototyping, and migrating those was rocky, because the original code was a mad scientist’s lab. How do you ask a robot to sift through a mad scientist’s lab? Fable did a surprisingly good job, but it took a couple of turns to decide what came over and where it went, and it hitched when it took on too much at once. I couldn’t just say migrate all ten of these notebooks; context got garbled. So I worked out memory stores better suited to Fable than to other models: restructuring each repo’s ARCHITECTURE.md, using pointers that Opus and Sonnet could read. Where something was ambiguous, Fable made a documented judgment call and flagged it for my review instead of stopping to ask.

The second build was a full strategy engagement. Over interactive sessions, Fable helped me define what to pursue over the next six to ten months and in what order. Then it did the thing that actually impressed me: it broke next steps into roughly ten work products and told me which ones deserved my live attention and which could be fired off and collected later.

Attention. It was helping me manage attention. That’s huge. Every opinion piece about dystopian modern technology ruining our brains focus on that: technology destroys our attention.

But here was something different. Fable was helping manage my scarcest resource, and managed correctly. The high-stakes design decisions got interactive sessions where I could push back in real time. The well-specified deliverables ran on their own. Juniors ask for direction on everything. Seniors tell you where your direction is needed.

The claim I disagree with

The consensus on the Claude subreddits, especially among people using Fable heavily, is that its edge over previous models only shows up on very complex technical work. My week says otherwise.

The clearest gap I saw was in ordinary conversation. Planning sessions, design discussions, sessions that never touched code. Forty turns into a session covering several research projects, Fable still had my constraints, my tooling, and my earlier decisions woven into its answers. Other models are good at this for a while, then the threads quietly drop and you’re re-explaining yourself to something that was nodding along an hour ago.

For me, this is the headline. The code orchestration is downstream of the thinking partnership, and a thinking partner who doesn’t lose the plot is worth more than a faster coder.

The honest caveats

It wasn’t flawless, and some of the flaws were mine to fix. Early on, Fable repeatedly misread how I actually work. Wrong assumptions about my workflow and tooling, and I burned turns correcting the same misunderstanding. The fix was embarrassingly managerial: I wrote it a short onboarding memo about my operating style, and the misfires stopped. Even excellent senior hires need onboarding. The difference is that juniors need onboarding plus ongoing supervision, and seniors need onboarding and then room.

The other thing that didn’t work: Fable would start doing things before I asked it to. Without a proper harness, it often went too far. Nothing disruptive ever came of it, just tokens spent before I could talk it through. One session I said: here’s the goal, but before we work on it I’d like to discuss some of the designs, and then we can implement. Fable read the handover documents and decided to implement some of the changes itself. It taught me to treat it as a very eager coding assistant.

Success also created its own management problem. I now hold validated plans and specifications representing threads implementation work. The architect outran the builders. Plans that pile up faster than execution capacity are a liability unless you triage them deliberately, but it’s a very, very good problem to have.

What this means

Every previous model asked me to be a supervisor, checking and directing everything. This one let me be a manager. If you’ve only ever handed these models tasks, try handing one a role instead.

We’ve had teams that never sleep for almost a year now. What’s new is that the management doesn’t sleep either: senior-level competence and judgment sitting next to the execution. The execution was already there. The judgment is starting to catch up.

Until next time, stay on the cutting edge, everyone.

I remember thinking: that sounds nice.

The bottleneck in serious research has never really been access to information. It’s knowing what to pay attention to. A graduate student who understands your work, your questions, and your blind spots, and who surfaces the right papers at the right time, is genuinely valuable. Most researchers don’t have one.

So I built one. I call it MARC — Multimodal Academic Research Curator.

What follows is a set of lessons from building and living with MARC over the past year. Some surprised me. All of them have changed how I think about what it means to stay current in a fast-moving field.

Share

What MARC Does

Before the lessons, a brief orientation.

MARC ingests research from academic papers on arXiv and SSRN, academic journals, podcasts, YouTube, and selected social media. Each source is processed by a large language model that produces a structured summary: not a recap of findings, but key ideas, notable quotes, and open questions the research raises. Every source is also scored across sixteen relevance dimensions I’ve defined based on my own research interests, from quantitative finance to geopolitics to machine learning. Those scores allow filtering and retrieval that goes well beyond keyword search.

The corpus now contains well over a thousand documents, roughly three-quarters of which are academic papers. That balance matters. Social media aggregation is easy to build. Getting a system that can reason meaningfully over dense academic work is a different problem. The more interesting one.

I interact with MARC through a persistent server accessible from anywhere, including my phone. I can ask it to surface recent papers on a specific topic, review trends across the corpus, or pull the full text of something I want to examine closely. It also sends me a weekly digest, automatically curated by topic, of what’s come in over the past seven days.

With that context, here are the lessons.

1. Model Choice Is a First-Order Decision

When I started building MARC, I tested a range of language models for the summarization layer. The differences were not marginal.

The cheaper models produced summaries that were dressed-up abstracts. They could repeat findings. They could not synthesize. When I asked for key ideas from a complex paper, they surfaced what was already stated in the introduction, verbatim often. What I wanted was the kind of summary a smart reader produces after actually engaging with the work: what does this really argue, how do the pieces connect, what does this imply for adjacent problems?

Frontier models do this. Cheaper models don’t. The gap isn’t about polish. It’s a qualitative difference in reasoning. For a system whose value depends entirely on summary quality, using the wrong model would have made MARC close to useless.

There’s a broader observation here. A lot of products being sold today are wrappers around models that don’t perform at the frontier, because frontier models are expensive at scale. For large commercial deployments, that tradeoff may make sense. For an individual building a tool for their own research, the economics work out differently. The capability difference justifies the cost several times over.

2. Building a Tool Forces You to Articulate What You Care About

To build MARC’s relevance scoring system, I had to define sixteen dimensions, each specific enough that a model could score a paper on it and distinct enough that they weren’t redundant.

This was harder than I expected, and more useful.

It required honesty about what I actually research, as opposed to what I tell myself I research. Some dimensions I initially wanted to include turned out to overlap substantially with others. That forced a real question: was I distinguishing two things, or just naming the same thing twice? Others I thought were peripheral turned out to be central once I looked at my actual reading habits.

Defining the system is a form of research planning. It surfaces assumptions, clarifies what you actually care about, and produces an explicit map of your intellectual interests. That map has value well beyond the tool.

3. Multimodal Access Changes Your Relationship With Research

Early in building MARC, I added the ability to generate audio summaries and visual representations of research. I treated these as nice-to-have features. There were, in fact, more than that.

Being able to listen to a summary of recent papers while driving, or ask for a visual of how a set of ideas connect, has changed how often I engage with the system. Research is no longer something that happens only when I’m at a desk with time blocked out for it. It fits into the margins of the day.

The lesson isn’t really about audio or images. It’s that the format in which you consume information shapes whether you consume it at all. A text-only system is constrained to contexts where reading is possible. A multimodal one adapts to your life rather than requiring you to adapt to it.

4. Serendipity Can Be Engineered

The weekly digest has produced some of my most useful research leads. Not because it found what I was looking for, but because it surfaced things I wasn’t.

Recently it brought up a paper on transformer architectures using a methodology I hadn’t encountered. It came through because of its relevance score on a dimension I track, not because I searched for it. I ended up spending several hours with it, using MARC to interrogate the methodology and think through where it might apply to problems I’m working on in quantitative finance.

Cross-domain discovery, machine learning methodology finding its way into finance research, isn’t a new idea. But it’s easy to miss in practice when you’re focused on your immediate reading list. The digest creates a regular encounter with things outside your active search.

Serendipity has a reputation for being random. It doesn’t have to be.

5. Prompt Architecture Is a Design Discipline

Early versions of MARC asked the language model for a summary. The results were inconsistent, sometimes strong, sometimes shallow, and the shallow results were hard to diagnose.

The fix was decomposition. Rather than “summarize this paper,” the system now asks separately for key ideas, notable quotes, and open questions. These aren’t stylistic categories. They have different failure modes. A model that misses something in one field often surfaces it in another. The structure creates redundancy that improves reliability.

This generalizes. The way you ask shapes what you get, and careful prompt design is a form of systems design. A vague prompt produces a vague output not because the model is weak, but because the question was underspecified. That’s on the architect.

6. The Upstream Curator Still Matters

MARC is only as good as the decisions I made when building it. The sources I chose to ingest, the relevance dimensions I defined, the topics I weighted. These all reflect my priors. The system surfaces what I’ve implicitly told it to look for. It doesn’t challenge my framing from outside.

There’s a version of this technology that sounds like it replaces judgment: put the research in, get the insights out. That’s not how it works. What it does is extend judgment, amplifying the ability to act on existing research instincts across a much larger volume of material, much faster. The instincts still have to come from somewhere.

The graduate student analogy holds here too. A great research assistant extends what you can do. They don’t substitute for your understanding of what’s worth doing.

7. The Time Dimension Requires Deliberate Attention

This one took me the longest to notice.

MARC’s summaries are generally strong on content. What a piece argues, what evidence it marshals, what questions it leaves open. Where they tend to be weaker is temporal context: when something was written, what the market environment was at the time, whether a forecast proved out.

In many fields this matters less. In finance, it matters quite a bit. A paper on portfolio construction written during one rate environment carries different weight than one written in another. An interview with a practitioner making predictions requires knowing whether those predictions came before or after the relevant events.

I’ve partially addressed this through careful prompting, and the system does surface publication dates and peer-review status. But this remains an area where human judgment has to stay engaged. The model tells you what a document says. It doesn’t always situate it in time the way a careful reader would. In finance, those are sometimes very different things.

What This Adds Up To

MARC has changed how I work in ways I didn’t anticipate when I started. Literature review, historically slow and incomplete, is now systematic. The breadth of what I track has expanded without increasing the time I spend on it. The depth of engagement with individual papers has also increased. I have something to think with, not just read with.

The force multiplier framing is accurate but undersells it a little. The nature of the work has shifted. Less time managing information, more time using it.

That’s what I was jealous of when that professor mentioned his graduate student.

Until next time, stay on the cutting edge, everyone.

Appendix: A Brief Technical Note

For readers interested in how MARC is built.

The system ingests content from academic sources (arXiv, SSRN, academic journals), podcasts, YouTube transcripts, and curated social media. Each document is processed by a frontier language model that produces a structured summary with three defined fields: key ideas, notable quotes, and open questions. Every document is also scored on sixteen relevance dimensions (each a specific topic area, rated 0 to 1), which allows filtering and retrieval beyond keyword matching.

Summaries and metadata are stored in a vector database, which enables semantic search across the full corpus. Queries surface conceptually related documents even when exact terms don’t match.

The system runs through a persistent MCP (Model Context Protocol) server, accessible remotely including from a phone. This allows both reactive use, querying the corpus on demand, and proactive use in the form of the weekly digest.

The corpus is over a thousand documents, approximately 75% academic papers. The system distinguishes preprints from peer-reviewed publications, and that distinction is surfaced in responses.

A full backtesting pipeline. Data sourcing, walk-forward optimization, thousands of parameter combinations per security, out-of-sample validation built for rigor rather than convenience. I built it without AI coding assistance, which, at the time, wasn’t good enough to meet my standards anyway. I built the full pipeline by hand, while working full-time and in grad school, and when I finally finished the visualization suite that would have made it publishable, I hit a wall I couldn’t get through.

I put it down. I didn’t come back for a long time.

When I did, the pipeline was still there. Intact, documented, ready to run. However, working with it again felt stranger than I expected.

The tooling landscape looks nothing like it did when I built that pipeline.

In 2024, I tried using AI assistants to support my workflow and found them wanting. Not dramatically, for they could handle simple tasks, but they couldn’t keep up with the kind of thinking the research required. The judgment calls, the architectural decisions, the places where domain knowledge actually matters. I went back to building by hand.

That’s changed. The tools I’m using now are legitimate thinking partners. They accelerate implementation, catch things I miss, and handle the mechanical work that used to just cost time. The delta between then and now is not incremental.

But I want to be precise about what that means, because the force multiplier framing cuts both ways.

These tools amplify what’s already there. Someone with deep domain knowledge, good architectural instincts, and a precise relationship with language will get extraordinary results. Someone without those things will get something that looks like extraordinary results for a while. The difference shows up later, in maintenance, in edge cases, in the moment when the system needs to do something slightly outside what was originally built.

I’ve thought about this in terms of language as much as code. The words you use with these tools are not interchangeable. Telling a system to treat something as a guideline produces different behavior than telling it to treat something as an axiom. One invites interpretation. The other closes it down. That distinction matters in quantitative research, where the difference between a soft constraint and a hard one can be the difference between a valid backtest and a compromised one.

The baseline I have, knowing exactly what it cost to build this infrastructure without AI assistance, is the thing that lets me measure the tools accurately. I’m not guessing at the delta. I’ve lived both sides of it.

What I’m building here runs on two tracks, and I think they’re inseparable.

The first is original research. Rigorous methodology, honest results, no cherry-picked outcomes. The kind of work that doesn’t apologize for complexity or dumb itself down for accessibility.

The second is transparency about how that research gets made. The tools, the decisions, the judgment calls, the places where AI assistance helped and the places where it didn’t. To me, that’s intellectual honesty, not just a concession to the education market.

The case for combining them is simple. AI tools are only as good as the person directing them, and the only way to demonstrate that is to show the work. Results without methodology are just claims. Methodology without results is just theory.

This is also a moment that rewards engagement over consumption, and learning over absorbing. The speed of change in this field doesn’t make research and education less important. It makes them more important. Judgment still matters in trading and investing. Curiosity still matters. The tools have gotten extraordinarily powerful, but they don’t replace the person using them. They reveal that person more clearly than ever.

I’ve been on the other side of this, building serious infrastructure alone, without the tools that exist now, and hitting a wall I couldn’t get through. Coming back to that work with a different set of tools, and a clearer sense of what I’m trying to do with them, is what makes this feel like a beginning rather than a restart.

Until next time, keep on the cutting edge, everyone.

We even saw some new subscribers, which is very humbling. Welcome to our new members!

As a reader, it’s always hard to know if a lapsed publication is a going concern, so actually subscribing or following during a lapse is very generous leap of faith. Truly - thank you all!

We could talk about life getting in the way and other details as explanations, but there are more exciting things to share with you all. Projects have been moving forward. Useful research is on the way. This dispatch will outline these developments and give you a better idea of what to expect from Alpha on the Edge.

This dispatch will focus on three topics:

1. Extending R2T Backtests with robust walk-forward optimization

2. Making more room for bespoke AI research

3. Launching Market Driver Insights

Onward and upward!

Subscribe now

1. Extending R2T Backtests with robust walk-forward optimization

The Really, Really Thorough Backtests have covered a lot of very fruitful ground. We have looked at literally millions of simulations of different trading algorithms across hundreds of securities. This research is one of the best ways to answer the question “Does this trading strategy actually work?”

But there is always room for improvement. The research so far has mostly used randomized cross-validation, meaning that in-sample and out-of-sample subsets are selected randomly from historical time series. This works very well but has drawbacks. Many randomized splits select subsets where trading strategies have no trades or very few trades to build metrics from. (To clarify: The research published here filtered out those simulations that weren’t useful.)

There are a number of ways to address this - alternative randomized sampling strategies or extended data time series, for example.

But for daily frequency time series, walk-forward optimization is a very elegant and more interpretable framework for backtesting. It gives a higher percentage of useful simulations while maintaining robustness in results.

The R2T Backtests now have pipelines for both randomized cross-validation and walk-forward optimization. This gives us the ability to be flexible and gives more avenues for research. There won’t be any less transparency - the research published here will continue to disclose which method is used.

2. Making more room for bespoke AI research

A lot of our published research has focused on rules-based trading strategies using technical indicators. We have not left a lot of room for experimenting with strategies built from deep learning and neural networks.

There’s a good reason for this. Building, tuning, and testing neural networks is extremely difficult and time-consuming. Neural networks can be unstable and unpredictable, and deploying them in a standardized “production”-style pipeline requires a large amount of computational resources and judgement calls.

But here’s the thing: It’s also very, very rewarding. This is the area that I always want to focus on more because of how powerful these models are and how quickly the field is developing, especially in finance.

There’s never enough time in the day. I still feel very strongly that the ability to run millions and millions of simulations is a valuable service. It exposes the real truth behind a lot of trading strategies. The problem is that scaling neural nets to that level is simply not possible without far more advanced technology. (Hopefully someone can tell me when I can get access to quantum computing!)

So the solution is to compromise! While building a modular WFO pipeline, I built more standardized components. A lot of the past projects required custom builds because of very bespoke research questions. This gave less time for the AI models.

Now, we can still share useful results from R2T Backtests, and rest assured, that content will not become boilerplate charts. The framework is far too valuable to stop custom research.

And this is great because I really, really want to play with Temporal Fusion Transformers….

3. Launching Market Driver Insights

The bulk of our recent posts have gone in-depth with factor research, or what we call market drivers. Market drivers have many powerful insights for traders and investors, but good content is not easily available outside of large financial institutions.

To fix this, we launched Market Driver Insights. This is a weekly publication with extensive analysis of market drivers and macroeconomic factors. It covers US equity markets and cryptocurrencies. The content on Market Driver Insights goes well beyond what was published here.

Here’s an example of the new tools:

Share

So if you found market drivers a useful tool, consider subscribing to Market Driver Insights.

To wrap this up, thanks again for sticking with Alpha on the Edge! We’ll be back with more research and look forward to sharing our work with you all.

Until next time, keep on the cutting edge, everyone.

It’s best to stay aware of and ahead of macro trends. When we were looking at crypto market drivers, for example, we saw that sensitivity to the US treasury yield curve steepness was a key driver of market returns.

How do we measure these sensitivities and stay on top of them? In this post, we will look at our method of choice, a very simple application of linear regression.

Sensitivity and Regression

Our goal is to estimate how a change in one variable affects a stock’s returns. For example, let’s consider yields on 10-year U.S. treasury bonds. If yields increase 0.01%, can we say how much a stock’s return will change?

To do this, we look at a simple linear regression model:

(For a good explainer on simple linear regression, I recommend this guide.)

The theory and practical applications of linear regression are extensive and very elegant. This model helps quantify exactly what we are looking for - how a change in a variable (x) affects another variable (y). The beta term (coefficient, sensitivity, or exposure) the amount of change in y we would estimate would occur for a one unit change in x.

Back to our question of yields and stocks - say we estimate a linear regression model with a beta of 0.3. This would tell us that a 1% change in yields would see a 0.3% increase in a stock’s returns.

We can’t claim that yields cause the change in a stock’s returns, nor can we claim that such a model has unimpeachable predictive power (it doesn’t). Meaning, this is an inexact method simplifying complex statistical modeling techniques.

But here’s the good news - the observations this technique gives us can lead to some powerful insights, especially if we look at sensitivity in aggregate or over time.

Regression in Practice

One way to use regression is to measure a stock’s sensitivity over time. We can use a rolling window of returns and yields and estimate a linear regression model. Meaning, we estimate a model on one day with the previous n returns and yield changes, move to the next day, and do again. We would then have a time series of regression betas, one for each model we estimated.

Let’s look at the stock in the S&P500 with the highest sensitivity to 10 year treasury yields as of the end of March - Progressive Corp . We want to look at:

• Is the sensitivity to yields always high?

• How quickly does the sensitivity change?

• Is there a trend that we can identify?

So our estimated model tells us that at the end of March, a 1% change in US Treasury yields historically is associated with about a 5.5% increase in PGR’s returns. However, we can easily see that PGR’s sensitivity to yields is not always high - it was negative at the beginning of 2023. We also see that the sensitivity does not typically change so dramatically that we would expect negative betas in the next week. The sensitivity has also been trending upwards since January 2023 but might have reached a peak recently.

If I were a trader looking at PGR, in other words, I would be very interested in changes in the 10Y yields. Macroeconomics matters quite a bit for this stock right now.

This technique also helps us look at the market as a whole.

Share

There’s a lot going on in this chart. In the middle of 2022, when the US Federal Reserve’s open market activities were front and center in the news, the stock market had a very wide, very significant spread in sensitivities to the 10Y yield. This spread narrowed throughout 2022, and the median sensitivity dipped below 0. Here in 2024, we see a somewhat consistent range of sensitivities with changing medians.

So, traders likely are not looking at 10Y yields as much as they were in 2022. But when yields become a significant market driver, this matters.

Are all macro forces as significant as 10Y yields? Let’s look at sensitivities to changes in the price of oil:

Share

Outside of energy stocks, we see little sensitivity to oil price changes. Even though oil prices can drive inflation, oil prices themselves are not usually a significant driver. (But this can change, as we saw with our market driver framework.)

Hopefully this post helped clear up what we mean by “sensitivity” and how we use (or misuse) linear regression models. This powerful tool informs many trading algorithms and AI models, so we definitely will see the concept again.

Until next time, stay on the cutting edge, everyone.

Disclaimers

The content on this page is for educational and informational purposes only. Any views and opinions expressed belong only to the writer and do not represent views and opinions of people, institutions, or organizations that the writer may or may not be associated with.

No material in this page should be construed as buy/sell recommendations, investment advice, determinations of suitability, or solicitations. Securities investment and trading involve risks, and not all risks are disclosed or discussed here. Loss of principal is possible. You are encouraged to seek financial advice from a licensed professional prior to making transaction decisions.

Further, you should not assume that the future performance of any specific investment or investment strategy will be profitable or equal to corresponding past performance levels. Past performance does not guarantee future results.

There’s a musical analogue to this. At the start of Wagner’s epic Ring cycle, he creates a vision of what lies beneath the surface of the Rhine river. This musical fantasy takes us to a story where beautiful Rhinemaidens live underwater, part of a world under the waves. The prelude to Das Rheingold is a sweeping, undulating, breathtaking depiction of the magic and mystery just beyond what we can see.

Surely there’s magic and mystery to market segmentation. There are no Rhinemaidens, but there are still surprises and insights - treasures in their own way. So let’s have a look.

How to Segment the Stock Market

We are returning to our familiar framework of market drivers, which helps us explain performance. (Take a look here if you need a refresher.) In our previous set of research, we saw that specific drivers came up when looking at the entire US large cap market (the S&P 500 Index). For example, sensitivity to oil prices and price acceleration were major drivers for the last four weeks.

Today, we will use two ways to create subsets of a stock market - market capitalization and sector classification. Market capitalization is a measure of the “size” of a company, calculated by multiplying the number of shares traded by the market value of those shares. Theoretically, that is the cost of a company’s traded equity. Right now, the largest two companies on the S&P 500 Index by market cap are Microsoft (MSFT - $3.1 trillion) and Apple (AAPL - $2.7 trillion).

We will use the S&P indices to separate US stocks by market cap. The headline S&P 500 Index contains the 500 largest companies by market cap in the US. The S&P MidCap 400 contains the next 400, and similarly with the S&P SmallCap 600. These three indexes are useful way to identify large, mid, and small cap stocks, respectively.

Sector classifications help divide companies by the type of markets they operate in. Microsoft and Exxon Mobil, for example, obviously sell very different products, so having a way to separate them is very useful. We will use the GICS Sector classifications, which divide stocks into 11 sectors.

Why do this? Different drivers operate on different segments of the market. Microsoft is much less associated with the price of oil than Exxon Mobil. If we can identify the drivers of specific segments, we know what is signal and what is noise.

Drivers by Market Cap

To look at driver performance for stocks of different sizes, we calculated driver performance three times, once for large caps and again for mid caps and small caps. Each set of stocks is it’s own “universe” from which we create portfolios with the top and bottom quintiles for each driver. We used the same set of 50 drivers from previous analysis, 10 drivers from 5 categories.

Share

The first row in this chart shows results we’ve seen before - the same set of drivers for the S&P 500. The second and third rows are new.

Sensitivity to economic data are still significant drivers for all three market segments for similar reasons as we saw previously. The S&P 500 had a great year, and stocks with high beta to that saw the highest performance gains.

Interestingly, momentum and volatility drivers were less impactful on small caps than large caps. Small caps tended to see less herding (which happened in the large cap tech/AI companies), so the momentum effects make sense. Volatility is somewhat more complicated. Small caps tend to be more volatile than large caps, but that volatility tended to not be a significant driver for small caps.

And finally, volume activity was a more significant driver for mid caps. That might be worth looking into. Those drivers can be manipulated by large buyers/sellers, but somehow, the effects on mid caps did not translate to less liquid small caps.

Sector Drivers

In our last post, we briefly mentioned that the S&P 500’s performance in 2023 could be attributed in large part to the Information Technology and Communication Services sectors. Those are the sectors with Microsoft, Amazon, Google, Meta, and other giant tech companies. So what are the drivers of those sectors, and do they differ by market cap?

We have created three subsets again, but this time, we added the additional filter of sectors. Then, we measured driver performance again for these subsets. This explains the differences from the first set of results.

And indeed, there are quite a few differences. It is startling to see the effects of momentum on large caps in these sectors given the lackluster momentum effects on the full universe. So this means that concentrated momentum drove the full market less than it drove these two sectors.

The effects are different across market caps again, however. We see some similar patterns - less volatility for small caps, higher effects from volume activity for mid caps.

To drive home the differences among sectors, let’s look at one more set of results for market drivers. This time, we will focus purely on large caps and look at drivers for each sector individually. This means that we created 11 subsets and measured driver performance for each:

Share

High dispersions in IT and communication services are expected, but other stories show up elsewhere - volatility in real estate and technicals in energy and industrials are stand outs. There is a lot more going on under the surface of the market here.

We didn’t find any mythical creatures or fantasy worlds here, but we did see how much segments of the market are affected by different drivers. Market segments do not react similarly to different forces, which is a key insight for traders and investors. Our driver framework continues to give us useful information for further investigation and research!

Until next time, keep on the cutting edge, everyone.

Disclaimers

The content on this page is for educational and informational purposes only. Any views and opinions expressed belong only to the writer and do not represent views and opinions of people, institutions, or organizations that the writer may or may not be associated with.

No material in this page should be construed as buy/sell recommendations, investment advice, determinations of suitability, or solicitations. Securities investment and trading involve risks, and not all risks are disclosed or discussed here. Loss of principal is possible. You are encouraged to seek financial advice from a licensed professional prior to making transaction decisions.

Further, you should not assume that the future performance of any specific investment or investment strategy will be profitable or equal to corresponding past performance levels. Past performance does not guarantee future results.

We’re looking more at market drivers and pivoting from crypto to stocks in this post. Explaining the stock market is a different endeavor than explaining the crypto market, after all. We should not assume that the same drivers are propelling both asset classes.

In this post, we will take our set of drivers and see what has and has not been helping stock performance over the past year. Along the way, we will ruminate on why it has paid to be one with the market. It pays to be the house when the house wins, after all.

Stock Market Drivers

Let’s begin with the one year performance of the top/bottom 10 drivers from our set of 50:

Share

As a quick recap - We are looking at the performance of strategies that measure an attribute of all stocks in the S&P500, buy the top 100, short the top 100, and repeat this every week. We have moved from the top 50 cryptocurrencies into the equity markets but are using the same driver set.

(Sidebar: Research on drivers/factors in stocks is far more developed than that in the crypto markets. Our set leaves out significant drivers with mountains of research purely to create a throughline in our posts. That is, our analysis here is meant to compare performance drivers between asset classes in addition to performance drivers for a single asset class.)

The top performing strategy was what we call “sensitivity to the S&P500.” This is also known as “market beta.” It basically says if you’re a company with a stock price that moves with the market, you do well when the market does well.

It’s worth taking a moment to think about that. The S&P 500 Index is a market cap weighted index, which means that the largest companies dominate the index’s returns.

This is shown in this great post from Statista:

So if you are the market, you do well when the market does well. And if you owned IT and communication services stocks in 2023, you probably did very well.

But our driver framework reveals something more interesting. Recall that our strategies here are buying the top quintile of stocks and equal weighting them. This means that even when removing company size from the equation, a stock’s sensitivity to the market as a whole captured more upside than any other driver in this set.

The story doesn’t end there, however. The stock market’s value isn’t a monotonically increasing series, so simply owning high-sensitivity (high-beta) stocks won’t guarantee your fortune here. Indeed, looking at the last 4 weeks:

Share

In the last 4 weeks, the story has shifted more to oil and momentum.

So the winners were the plays on oil:

Thinking back to our last post, this was not the case with crypto. Oil sensitivity was in the top drivers of crypto for the same period, but the top driver was sensitivity to changes in the steepness of the US treasury yield curve. In other words, stock portfolios wanted oil exposure, crypto portfolios wanted yield curve exposure.

With the S&P500, we can see that high momentum stocks also did well, but only for certain kinds of momentum. 90D acceleration and 30D rate of change performed well, but 7D rate of change did not. Longer-term momentum was the high flier, in this case. At some point, I’d like to look more at momentum, because it has tons of research supporting it but isn’t the most popular investing style.

Finally, let’s look at our giant summary table:

Share

What stands out to me is the set of return correlations for the set of sensitivity drivers. A good rule of thumb is that an average correlation above 0.02 or below -0.02 is a significant signal. We have 0.04 for SP500 sensitivity and -0.06 for gold sensitivity, both significant readings.

We didn’t see this with crypto. With crypto, the significant return correlations came with technicals, volatility, and momentum drivers.

Another reason the driver framework is so powerful is it helps us give language to the differences between asset classes. Here, we see stocks connecting with certain “themes” - tech/AI, commodities, and credit markets. With crypto, the story has been more about price action. Even though sensitivity to the steepening yield curve was the highest performing driver for crypto, that had less correlation on average with asset returns than price action drivers did.

We have stayed with the market driver framework for a while now because of its power and versatility. What this framework lacks in mathematical rigor it makes up with its interpretability. It’s very easy to explain and reveals many insights for it.

In other words, given that we all learn from stories, this market driver framework is instructive because of the stories it tells.

Until next time, keep on the cutting edge, everyone.

Disclaimers

The content on this page is for educational and informational purposes only. Any views and opinions expressed belong only to the writer and do not represent views and opinions of people, institutions, or organizations that the writer may or may not be associated with.

No material in this page should be construed as buy/sell recommendations, investment advice, determinations of suitability, or solicitations. Securities investment and trading involve risks, and not all risks are disclosed or discussed here. Loss of principal is possible. You are encouraged to seek financial advice from a licensed professional prior to making transaction decisions.

Further, you should not assume that the future performance of any specific investment or investment strategy will be profitable or equal to corresponding past performance levels. Past performance does not guarantee future results.

1. Additional volume metrics

2. A new category called “Sensitivity”

3. Results for a huge set of 50 drivers

There’s a commonly cited criticism of market factor (what we call market driver) research - the huge number of them create a “factor zoo.” This term comes up a lot in research because so many factors are redundant, have little explanatory power, and lack robustness. In other words, why do we need so many?

In this post, we will discuss that question after briefly introducing measurements of volume activity and sensitivity. Let’s explore the zoo!

Volume and Sensitivity

If you need a recap of our market driver framework, see this post:

Introduction to Market Drivers

Traders have long found signals in trading volume, which is like the pulse of activity from buyers and sellers. Volume acts as a strong complement to price activity and can signal trends, reversals, and the intent of large investors. There are a few helpful volume indicators we are using here:

• On Balance Volume (OBV) - OBV measures the flow of volume in a positive or negative direction. This indicator cumulatively adds volume on days the price finishes up and subtracts volume on days it ends down, painting a picture of whether money is flowing in or out. A rising OBV suggests buyers are willing to step in and push prices higher, while a falling OBV indicates sellers are in control, possibly leading to lower prices.

• Accumulation/Distribution (AD) - AD takes a broader sweep, combining price movement and volume to determine whether a stock is being "accumulated" (bought heavily) or "distributed" (sold heavily). It’s akin to tracking footprints in the sand, showing where the market's big players are moving, often ahead of price changes.

• Klinger Volume Oscillator (KVO) - The KVO is the detective of the volume world, uncovering the story behind the price movements by comparing high-volume to low-volume days. It aims to predict the direction of the trend by detecting divergences between volume flow and price. The KVO oscillates around a zero line, with readings above zero possibly signaling bullish trends, and vice versa for readings below zero.

We would expect long/short portfolios with these metrics to have high correlation, right? Sounds like these would be redundant.

Another important category of measurements looks at how returns respond to changes in other data. For example, if the price of oil increases, what would happen to the price of bitcoin? Measuring the sensitivity of bitcoin to oil can be accomplished a number of ways. We are using “beta” - the coefficient from a single linear regression model - to measure this sensitivity. A higher beta means that a higher change in the price oil would be associated with a higher change in the price of bitcoin.

Here’s why this is useful. If you expect the U.S. Federal Reserve to lower the Federal Funds Rate and expect short-term yields to decrease, you could have information on what might happen to the price of a coin or stock. It’s not an exact prediction, but as we’ll see, the sensitivity can matter in times of change.

Crypto Top 50 Market Driver Performance

Now that we have described the additions to our zoo, let’s have a look around. We have a total of 50 drivers, and this giant bar chart show the performance of all 50 for the 4 weeks ending April 5, 2024:

Share

The 2s10s spread has been a huge driver in the last 4 weeks. This means that a portfolio concentrating on coins with historically high sensitivity to a steepening yield curve outperformed. It’s not just that we’ve seen a recently steepening yield curve, but that the coins with the highest sensitivity did very well.

Additionally, our volume drivers have done very well recently. But, they aren’t as correlated as one might think. The short-term KVO and OBV rates of change did well, but the short-term AD Oscillator did not.

Here’s a giant table to go with the giant bar chart:

Share

Different stories emerge as we expand the horizon to 52 weeks. Volatility was a more significant driver in the one year time period. However, sensitivity to the yield curve shows some predictive possibility even over 52 weeks - the 2s10s spread had an average correlation of 0.02 over the period.

Our volume drivers have mixed performance over the extended time period, and most do not show significant average correlations. However, they clearly constructed diverse portfolios, despite the on-face similar metrics used.

To drive this home, let’s look at the giant correlation heatmap:

Share

Is this zoo overgrown? Is this an information overload?

Yes. A 50x50 correlation matrix is just too much, as is a 50-row performance table.

But does that mean we have redundant drivers?

Not quite - the varying correlations and results suggest that it might be worth tracking these regularly if not looking at the results in depth. Meaning, we don’t have to look at a giant correlation matrix every time we run this analysis.

Even still, the fact remains that our market driver framework reveals some interesting insights. We will stick with this for a little while longer and look at the equity markets next time.

Until next time, keep on the cutting edge, everyone.

Disclaimers

The content on this page is for educational and informational purposes only. Any views and opinions expressed belong only to the writer and do not represent views and opinions of people, institutions, or organizations that the writer may or may not be associated with.

No material in this page should be construed as buy/sell recommendations, investment advice, determinations of suitability, or solicitations. Securities investment and trading involve risks, and not all risks are disclosed or discussed here. Loss of principal is possible. You are encouraged to seek financial advice from a licensed professional prior to making transaction decisions.

Further, you should not assume that the future performance of any specific investment or investment strategy will be profitable or equal to corresponding past performance levels. Past performance does not guarantee future results.

This round, we have expanded our collection of drivers from 3 to 21, so we have a better set of results to look at. It’s worth casting a wide net with crypto, as there has not been much research on the space yet. Our large set of drivers balances looking at too few drivers and looking at too many.

We will do a quick recap of the framework we are using then discuss recent performance of market drivers in crypto.

Quick Recap - Market Drivers

The market driver framework we use involves backtesting rules-based portfolios that concentrate investment in a certain attribute of a cryptocurrency, like volatility or momentum. Our portfolio strategy is to buy the “best” coins according to the attribute and short the “worst” coins.

Take the measurement of price acceleration, represented by an approximated second derivative of daily prices. This is one way to quantify a coin’s price momentum. Our hypothetical portfolio buys the coins with the most price acceleration and shorts the coins with the least price acceleration. We are concentrating our conviction in a specific investment thesis - accelerating price means higher momentum means higher returns for our portfolio.

We assume that we repeat this every week, rebalancing into the best/worst coins every Friday. We’re looking at quintiles in this analysis. That means our strategy will go long 10 coins and short 10 coins each week.

The results we will discuss take 21 attributes, backtest the associated strategies, and report the performance of the hypothetical portfolios with this strategy. In a way, this is a set of measurements reporting the success of specific investment theses over a given timeframe.

Driver Performance Results

Let’s start with the 4-week performance as of March 29, 2024:

Share

First, let’s talk about the colors of the bars. This is a useful way to visually group similar drivers. We currently have four categories:

• Momentum (Red)

• Volatility (Yellow)

• Technicals (Green)

• Volume Activity (Light Blue)

This is neither an exhaustive nor definitive list of categories. “Momentum” and “Technicals” can be very fuzzy categories, for example. Further, there is another category to be added to our set here, which we will talk about in the near future.

Let’s look at the top performer: “Volatility 365D”. What this is reporting is how well someone investing in volatile coins performed over 4 weeks. Every Friday, this portfolio goes long the 10 coins with the highest volatility over the last year, and short the 10 coins with the lowest volatility.

This simple strategy, along with four other simple strategies built around volatility measures, performed incredibly well over the 4 weeks.

It’s fascinating how some drivers will have periods of excellent performance and periods of underperformance. This is clearly not the time to look at common measures of momentum, for instance.

It’s also interesting to note that while straight volatility performed well, the “Vol of Vol” portfolio did not. Meaning that concentrating on the volatility of volatility would not have been a good strategy. The best drivers were consistent volatility, not inconsistent volatility.

Let’s look at a more detailed performance summary:

Even more to unpack here. We see the color-coded categories again, which helps us continue to group drivers. The performance columns report the numerical cumulative performance. The 4W columns is the same data from the earlier bar chart.

(Quick sidebar - isn’t that a beautiful table? That’s completely custom work with pandas! I’m proud of this one.)

The “Strength” columns are mostly familiar to us at this point. “Avg Corr” is the average correlation of a driver value with the forward returns. It’s one way to measure the predictive value of a driver. Values around 0 are common. A general rule of thumb is that values above 0.02 and below -0.02 are significant.

“Sharpe” is the strategy’s annualized risk-adjusted performance, and the “Win Rate” here is the percentage of weeks with positive performance.

Looking at this table, we can see many conflicting pieces of information. The “Volatility 365D” strategy has high performance but negative correlation with forward returns. This indicates that the performance over the past year could have been more luck.

Conflicting measurements are common in market driver research, so we take notice when the signals line up. “Volatility 30D” achieved extremely high performance, has a high win rate, and a high correlation with forward returns. That is noteworthy. It doesn’t mean we should all go start buying volatility. Drivers can be inconsistent, but it is worth watching these in the future.

Finally, let’s look at cross-driver correlation:

(Another beautiful table, right? We’re on a roll with this post!)

The other fascinating part of driver analysis is the diversification potential. Volatility is understandably correlated with volatility, but not with all strategies. And some momentum strategies, for example, have very low correlation with all categories of drivers. This tells us that volatile coins tended to not have momentum in the past year.

A big takeaway is that volatile coins has a very, very good year. Volatility wasn’t the only driver of the crypto markets, but it was a major one.

We will spend some more time looking at market drivers in our next posts, including drivers in the equity markets. Always more to find here.

Until next time, keep on the cutting edge, everyone.

Disclaimers

The content on this page is for educational and informational purposes only. Any views and opinions expressed belong only to the writer and do not represent views and opinions of people, institutions, or organizations that the writer may or may not be associated with.

No material in this page should be construed as buy/sell recommendations, investment advice, determinations of suitability, or solicitations. Securities investment and trading involve risks, and not all risks are disclosed or discussed here. Loss of principal is possible. You are encouraged to seek financial advice from a licensed professional prior to making transaction decisions.

Further, you should not assume that the future performance of any specific investment or investment strategy will be profitable or equal to corresponding past performance levels. Past performance does not guarantee future results.

Overall, March was an exciting month. There’s a lot to talk about. We experimented with new concepts, discovered useful strategies, and laid even more foundations for future work. Some surprises came up along the way, too.

Onward and upward!

Trend Following with Neural Networks

We started March with training our CNN-LSTM model and getting a real backtest with FX pairs.

What happened was that our simulated strategy worked surprisingly well. The model expressed confidence throughout the out-of-sample period in the FX pair with the highest performance over the period. In other words, a trading strategy would have remained invested in the top performing asset over 1 year.

Combined with the original research’s success, this gives a lot of confidence in this model. We haven’t returned to the CNN-LSTM yet because, honestly, it exceeded our expectations already. The next steps we might take are to try long-short strategies, experiment with different feature sets, or simply take the model into live trading.

We also want to experiment with other model architectures and targets, too. Some fascinating research is being published that will be fruitful for our work.

R2T Backtests on the RSI and RSX

Our research with the RSX truly arose organically. It came from pure curiosity about the RSI alternative. Like the first post mentioned, it’s rare to find a commercial alternative to an established indicator that is worth looking into. But, the RSX worked. That research really was delightful - not in the Keebler elf Christmas kind of way, but in the “what a fun find!” kind of way.

And in a wild coincidence, Quantified Strategies posted a solid RSI trading strategy the day before our first post was published. We just had to drop everything and test the RSX again. Again, with more delightful surprises, the RSX version worked again.

We can’t help but be curious about Jurik Research’s other work. Their moving average wasn’t a stand out in our MA crossover backtests, but admittedly, the MA crossover itself isn’t a compelling strategy. It might be worth revisiting the JMA at some point.

AI-Generated Trading Strategies

Putting GPT-4’s trading knowledge to the test was a fun experiment. We asked it for a reversal strategy, and it gave us one that we could test on the S&P 500. We tried millions of parameter combinations and found scattered success. The performance didn’t stun us, but it wasn’t bad, either. The experiment was the main point (savor the journey over the destination, you know).

Trying out more AI generations is worth the time. LLMs can’t perform backtests, but they can draw on massive sets of trading knowledge. There might be gold in there somewhere.

Finally, Market Drivers

It’s hard to overstate how important factor research is for professional investors. Knowing what’s differentiating outperforming securities helps investors and traders make much more informed decisions.

And yet, this kind of research isn’t easily available for wider audiences and for markets outside of equity markets. Partially, that’s because factor research is far more established for equities. Equity market drivers are more well understood and documented. And that’s also partially due to the lack of computing resources and data available to non-institutional investors.

But that doesn’t mean we can’t 1) experiment with factors and 2) share that work. Indeed, we intend to expand our market driver research and, possibly, make it a more regular part of our publications.

Thanks for joining us, and look forward to more in April. We’re excited to share the research we’re working on. Feel free to reach out with any ideas or feedback. Your input always helps!

Until next time, keep on the cutting edge, everyone.

If only it were this easy to identify cause and effect in capital markets.

The drivers (or factors) behind market returns are the subject of libraries of research. It is a fascinating area of study, and in this post, we will discuss a useful type of research that uses rules-based portfolios to analyze those drivers. By keeping an eye on these drivers, traders can make smarter choices and diversify their portfolios easily.

Forming Factor Portfolios

First, we have to choose our investment universe. Often, researchers will choose stock indexes, so for fun, we will go against the grain and look at the top 50 cryptocurrencies by market cap. We exclude stablecoins and coins with less than two years of trading history.

Second, we choose a factor to analyze. This analysis begins with the trailing 30-day volatility, measured by the standard deviation of daily returns.

Third, we use a straightforward ranking methodology to form portfolios. Our algorithm will measure the 30-day volatility of each of the 50 cryptocurrencies, rank them, and separate them into quintiles. We then go long the top quintile (buy the 10 CCs with the highest 30-day volatility) or really concentrate our factor exposure and go long the top quintile and short the bottom quintile.

Say we re-measure 30 day volatility each week and buy the 10 coins in each quintile with equal weighting. Here’s how the portfolios would have performed:

Share

For half of the 1 year period, the quintile 1 portfolio outperformed the quintile 2 portfolio, the quintile 2 portfolio outperformed the quintile 3 portfolio, and so on. This is what makes market driver research so fascinating - even a simple ranking methodology can create superior portfolios.

And what’s more, every single portfolio outperformed Bitcoin over the period. Even the lowest volatility portfolio achieved superior performance.

We often see this stratification on a risk-adjusted basis, too:

So does this mean that volatility explains price action in crypto markets? Not quite, but it does mean that volatility is a factor that might drive performance. Our job is to observe the strength of that driver and measure it against other possible market drivers.

Measuring Market Drivers

Let’s introduce two more factors into our analysis:

• Ratio to 52W High - The ratio of the last close to the previous 52 week high. This assumes mean-reverting behavior in prices.

• Momentum 1W4W - The ratio of the 1 week return to the 4 week return. A higher ratio above 1 suggests an accelerating price increase, meaning that the price could have positive momentum.

The traditional performance measure of a driver/factor is the long-short performance - long the top quintile, short the bottom quintile, and rebalance after every period (weekly in our case). This measures the performance of a portfolio concentrated to capture that driver’s effects.

Here is the long/short performance of the three factors over 1 year:

Our volatility and mean-reversion factors show very similar performance, while the momentum factor was mostly flat over the period.

This illustrates another use for this methodology: diversification. While the volatility and mean reversion portfolios are highly correlated, the momentum portfolio is less correlated to both. And it turns out that all three portfolios have low (less than 0.3) correlation to Bitcoin, as well. This means that traders might be able to mitigate risk through this methodology, as well.

There is one more useful metric of driver strength - the correlation of the factor value with the subsequent security returns. (This is also called the information coefficient.) This is another way to measure how strong the relationship of the factor to performance might be. In practice, a correlation above 0.02 or below -0.02 is often considered significant.

Summarizing the measurements of the three factors here:

“Avg Corr” means that we measure the correlation of factor values to forward 1 week returns each period, then average the measurement for all periods.

A few things stand out here. Momentum outperformed on an absolute and risk-adjusted basis, but the low correlation with returns suggests that the relationship might be spurious. That portfolio could have just been lucky, in other words.

Additionally, our original volatility factor has the correlation with the highest absolute value, suggesting that it has the strongest relationship to returns among the three factors. This doesn’t mean that we should all go invest in volatile coins. Rather, it means that this is a factor that we could watch through time and, if the strength of the relationship holds, look at more deeply.

I hope this helps give a better understanding of why we like talking about market drivers around here. Expect to see more material on this soon.

(And it’s a good time for me to sign off, dinner time is clearly a factor driving my dog’s behavior right now!)

Until next time, keep on the cutting edge, everyone.

Disclaimers

The content on this page is for educational and informational purposes only. Any views and opinions expressed belong only to the writer and do not represent views and opinions of people, institutions, or organizations that the writer may or may not be associated with.

No material in this page should be construed as buy/sell recommendations, investment advice, determinations of suitability, or solicitations. Securities investment and trading involve risks, and not all risks are disclosed or discussed here. Loss of principal is possible. You are encouraged to seek financial advice from a licensed professional prior to making transaction decisions.

Further, you should not assume that the future performance of any specific investment or investment strategy will be profitable or equal to corresponding past performance levels. Past performance does not guarantee future results.

In our first post on this strategy, we introduced the strategy in broad strokes, discussed the trading rules, and performed an initial backtest on SPY.

In this second post, we will look at a more extensive set of backtests on the stocks in the S&P 500.

You all know how we like to do things around here. We literally did over 64 million simulations to test this strategy. If anyone can be said to have really put an AI’s claim to the test, it’s us.

We will also take a beat to discuss parameter stability and feasible parameter spaces. For a strategy like this, these considerations matter a lot.

Subscribe now

Parameter Space

Believe it or not, this strategy has 12 parameters:

• Fast MA Window

• Slow MA Window

• Stochastic K

• Stochastic Slow K

• Stochastic Slow D

• Stochastic Lower Threshold

• Stochastic Upper Threshold

• Ichimoku Cloud Tenkou

• Ichimoku Cloud Kijun

• Ichimoku Cloud Senkou

• Stop Loss ATR Multiplier

• Take Profit ATR Multiplier

The parameter space for this strategy can grow really, really quickly. Even just 3 values per parameter creates a parameter space of over 1 million possibilities, which strains both computational resources and usefulness for such a large parameter space. We limited our test to ~150,000 parameter configurations to give a solid foundation for our results. The parameters all stayed around the “standard” values for the indicators - 9 for the Ichimoku tenkou, etc.

One consideration in backtesting is parameter stability. This can be defined a number of ways, but the idea is simply to help us determine if we are curve fitting or not. If the strategy’s results change dramatically for a change in one parameter, this is evidence that the “optimal” parameter configuration is an anomalous result.

Result Overview

For each stock in the S&P 500 with history available from 2008, we ran simulations for each of the ~150,000 parameter configurations, which gave us over 64 million simulations to look at.

If you take the best metric for each stock, the results look like this:

Share

A couple of things stand out - this strategy tends to give favorable win rates. Historically, it has been better than a coin flip for the best parameter configuration on almost every stock tested. That’s impressive.

What’s less impressive are the risk-adjusted performances of the strategy on each stock. Even with the “risk management” prompt, GPT4’s suggested use of ATR-based SL and TP thresholds did not give us incredible Sharpe ratios. They aren’t terrible, though.

Looking at the outlier expectancy values - this strategy happened to catch a few lucky pops on stocks with very few signals. These results filter out simulations with low numbers of trades, but on some of the stocks here, you can expect to see outlier results. That’s why we like to look at medians and distributions!

But About Those Parameters

Let’s return to the topic of parameter stability. Again, defining “stable” parameters is more an art than a science, but our goal is to see if our results are anomalous. If we change the fast MA window 5 bars, say, and the simulation’s performance tanks, that’s not a sign of a robust strategy.

With such a large parameter space, my idea was to take a simple approach. I wanted to look at the simulated Sharpe ratios for 3 sets of simulations:

1. The optimal parameter configuration for each stock (the values plotted in the histograms above)

2. The parameter configuration with the highest median Sharpe ratio across all stocks

3. The parameter configuration comprised of the values most like to be used in the optimal configuration for each stock.

The third set deserves a little more explanation, because I have not seen it used elsewhere. I calculated the probability that each parameter value would be in the optimal configuration for each stock.

For example, we tested 2 different values of the ATR multiple for a stop loss. Of the 432 optimal configurations, 87% of those used a multiplier of 4.0.

One main message here is that as soon as you step out of the “optimal” configuration for each stock, you potentially lose 0.3-0.4 on your Sharpe ratio. This gives some pause for the generalizability of this strategy. Our tests here did not use cross validation, which we might do in the future (the primary obstacle is computing power - 64 million simulations would become 1.2 billion simulations).

Another main message is that this strategy depends on the stock it’s used on. In that vein, here are the 10 stocks with the highest optimal Sharpe ratio from this strategy:

Share

Thanks for checking out our research into this reversal strategy, everyone! The results are promising. We will definitely try more AI-generated strategies in the future. If AI will truly take all of our jobs in the future, we need to know what it can do!

Until next time, keep on the cutting edge, everyone.

Resources

The AotE Toolkit

Disclaimers

The content on this page is for educational and informational purposes only. Any views and opinions expressed belong only to the writer and do not represent views and opinions of people, institutions, or organizations that the writer may or may not be associated with.

No material in this page should be construed as buy/sell recommendations, investment advice, determinations of suitability, or solicitations. Securities investment and trading involve risks, and not all risks are disclosed or discussed here. Loss of principal is possible. You are encouraged to seek financial advice from a licensed professional prior to making transaction decisions.

Further, you should not assume that the future performance of any specific investment or investment strategy will be profitable or equal to corresponding past performance levels. Past performance does not guarantee future results.

I’m still skeptical about that, but I have a lot of respect for the capabilities of current generation large-language models like OpenAI’s GPT-4 and Google’s Gemini. They have their flaws but can achieve incredible results, especially with programming. One recent study found that many programmers are using LLMs for about 50% of their jobs, and I can believe it. LLM code is far from perfect but gives a great starting point to iterate.

I am very, very curious about the quality of trading algorithms LLMs can produce. In theory, LLMs have access to an enormous corpus of trading strategy material. So, I am enjoying the exercise of translating AI-generated trading algorithms to Python and backtesting them.

My first test is on a reversal strategy from GPT-4. I prompted GPT-4 to create a reversal strategy with good risk management, and it generated something we can look at in depth.

I want to start this short series with introductory material. At the end of this post, we will look at a simple backtest on SPY. Our next posts will look at full R2T results.

Share

Introduction to Reversal Strategies

Reversal trading strategies are predicated on the idea that markets move in phases: trends, corrections, and sideways movements. These strategies aim to identify points where a price trend is likely to reverse direction, enabling traders to enter early in a new trend. The core of reversal trading lies in recognizing patterns and signals that indicate the end of the current trend and the beginning of a new one.

The foundation of reversal strategies is rooted in market psychology and supply and demand dynamics. When an asset is overbought, it suggests that buying pressure may soon wane, leading to a potential reversal to the downside as sellers take control. Conversely, when an asset is oversold, selling pressure may be nearing its end, possibly giving way to a reversal to the upside as buyers step in.

The Ichimoku Cloud and the Stochastic Oscillator

We’re using a couple of new indicators here, so I wanted to offer brief explanations. The Ichimoku Cloud in particular is a star of visual storytelling in technical analysis, and I truly enjoy using it. The Ichimoku Cloud provides information on support/resistance, trend direction, momentum, and potential buy/sell signals, all within one chart. Developed by Goichi Hosoda in the late 1930s, the indicator is built on five main components:

1. Tenkan-sen (Conversion Line): Represents short-term momentum, calculated as the average of the highest high and the lowest low over the last 9 periods.

2. Kijun-sen (Base Line): Indicates medium-term momentum, derived from the average of the highest high and the lowest low over the last 26 periods.

3. Senkou Span A (Leading Span A): One of the two lines forming the "cloud", calculated as the average of the Tenkan-sen and Kijun-sen, plotted 26 periods ahead.

4. Senkou Span B (Leading Span B): The second line of the cloud, determined by the average of the highest high and the lowest low over the last 52 periods, plotted 26 periods into the future.

5. Chikou Span (Lagging Span): Shows the closing price plotted 26 periods into the past.

The "cloud" is formed between Senkou Span A and Senkou Span B and is the hallmark of the Ichimoku system. Its color (green for bullish, red for bearish) and position relative to the price action provide insight into the market sentiment. A price above the cloud indicates an uptrend, while a price below suggests a downtrend. The cloud's thickness can also signal the strength of the support or resistance it provides.

We’ll be testing this strategy on SPY:

The second indicator is the Stochastic Oscillator. The Stochastic Oscillator is a momentum indicator developed by George C. Lane in the 1950s. It compares a closing price to its price range over a given period, with the premise that momentum precedes price. It oscillates between 0 and 100, with readings below 20 considered oversold and above 80 considered overbought.

The formula consists of two lines: the %K line, which measures current price relative to the high-low range, and the %D line, a moving average of the %K line, acting as a signal line.

This sounds like the RSI, right? The RSI measures the speed and change of price movements to evaluate overbought or oversold conditions. While both indicators aim to identify potential reversal points through overbought and oversold conditions, their calculation methods and interpretations differ. The Stochastics focus on the price's position relative to its high-low range over a period, making it sensitive to price swings and potentially leading to frequent signal changes. This makes it valuable in identifying short-term price reversals and cyclic moves within a trend.

GPT-4’s Reversal Algorithm

After being prompted for a reversal strategy with risk management, GPT-4 offered the following:

Entry Rules:

1. Bullish Reversal (Buy Signal):

   • The market shows a bullish trend (50-period EMA above 200-period EMA).

   • The Stochastic Oscillator is below 20, indicating an oversold condition.

   • The price is above the Ichimoku Cloud, confirming the bullish sentiment.

   • Enter a buy position when the Stochastic Oscillator crosses back above 20, signaling a potential reversal to the upside.

2. Bearish Reversal (Sell Signal):

   • The market shows a bearish trend (50-period EMA below 200-period EMA).

   • The Stochastic Oscillator is above 80, indicating an overbought condition.

   • The price is below the Ichimoku Cloud, confirming the bearish sentiment.

   • Enter a sell position when the Stochastic Oscillator crosses back below 80, signaling a potential reversal to the downside.

Exit Rules:

1. Stop-Loss: Set a stop-loss at 1.5x ATR from the entry point to protect against large losses.

2. Take-Profit: Set a take-profit level at a minimum of 3x ATR from the entry point to ensure a favorable risk-reward ratio.

3. Trend Reversal: Exit the trade if the price moves against the trend direction indicated by the EMA and Ichimoku Cloud alignment.

Results on SPY

Running this algorithm using Python and VectorBT PRO show initially disappointing results. We start with a long/short strategy that includes GPT’s suggested risk management, and our strategy has a win rate of 43% and an expectancy of 0.58. We miss almost all of SPY’s gains for some comparatively low results.

Here is the output from VBT’s portfolio simulator:

We obviously need to try some iterations! But, we have an interesting start, and in future posts, we will try some parameter configurations as well as variations on the suggested risk management techniques. Just like with programming, I suspect GPT gives us a good starting point, but not a good ending point.

Until next time, keep on the cutting edge, everyone.

Resources

The AotE Toolkit

Disclaimers

The content on this page is for educational and informational purposes only. Any views and opinions expressed belong only to the writer and do not represent views and opinions of people, institutions, or organizations that the writer may or may not be associated with.

No material in this page should be construed as buy/sell recommendations, investment advice, determinations of suitability, or solicitations. Securities investment and trading involve risks, and not all risks are disclosed or discussed here. Loss of principal is possible. You are encouraged to seek financial advice from a licensed professional prior to making transaction decisions.

Further, you should not assume that the future performance of any specific investment or investment strategy will be profitable or equal to corresponding past performance levels. Past performance does not guarantee future results.

We tested 8 variants of the classic RSI reversal strategy on SPY and compared the results with using the RSX in the same strategies. There was evidence that the RSX versions were both more consistent and better performing for most variants.

In a coincidence of good timing, our friends at Quantified Strategies posted a Triple RSI trading strategy a few days before my RSI vs. RSX post was published. So, naturally, I wanted to test out the RSX version of that strategy!

The Triple RSI Strategy

The trading rules are as follows:

• Buy when the following criteria are met:

  • 5 day RSI < 30

  • 5 day RSI has decreased for 3 bars in a row

  • 5 day RSI was below 60 as of 3 bars in the past

  • Closing price is higher than the 200 day moving average.

• Sell when the 5 day RSI crosses above 50

The folks at QS found some compelling results when backtesting on SPY, so lets see how the Triple RSI does on the stocks in the S&P 500 Index and compare with our own Triple RSX version!

Triple RSI vs. Triple RSX on S&P 500 Stocks

Quick bit of housekeeping before diving in. The backtests here differ from QS in a couple ways:

• Our backtests assume that buy and sell orders are made at the open price of the next bar, where QS uses the close of the signal bar

• Our backtests assume a small amount of slippage and fees

(See our backtesting methodology for more info)

Looking first at win rates, we see a similar pattern to our first tests on the RSX. The RSX has a narrower range of win rates, indicating a higher level of consistency. So it does not result in outliers as much as the RSI. What's new here is that the median win rate of the Triple RSX is a little higher than the Triple RSI.

And because we are using he stocks in the S&P 500, we can break this down by sector, as well:

The Triple RSX achieved higher median win rates in 9 out of 11 sectors, which is very compelling. We see large differences in energy and financials, especially.

Let's look at the win rate vs expectancy:

We again see that the Triple RSX results are more clustered (with one interesting exception, back to that in a bit). However, the Triple RSX has a slightly lower correlation between win rate and expectation - 0.61 for Triple RSX vs 0.65 for the Triple RSI.

And since we have sectors split out, any pattern in the sectors here?

Not really.

But the outliers here look very interesting! So here are the top 10 Triple RSX stocks ranked by expectancy:

Share

Conclusion

We have additional evidence that the RSX is a superior choice than the RSI. The Triple RSX strategy results in more consistent win rates that tend to be higher than those achieved by the Triple RSI strategy. So, the RSX gives us a smoother indicator as well as a better strategy.

As a side bar, I promise that this is not some marketing for Jurik Research! I am honestly surprised that a custom version of the RSI performs so well. It gives me more respect for JR, and I’m now curious about their other custom indicators.

Until next time, keep on the cutting edge, everyone.

Resources

The R2T Backtesting Framework, Ver. 1

The AotE Toolkit

Disclaimers

The content on this page is for educational and informational purposes only. Any views and opinions expressed belong only to the writer and do not represent views and opinions of people, institutions, or organizations that the writer may or may not be associated with.

No material in this page should be construed as buy/sell recommendations, investment advice, determinations of suitability, or solicitations. Securities investment and trading involve risks, and not all risks are disclosed or discussed here. Loss of principal is possible. You are encouraged to seek financial advice from a licensed professional prior to making transaction decisions.

Further, you should not assume that the future performance of any specific investment or investment strategy will be profitable or equal to corresponding past performance levels. Past performance does not guarantee future results.

Why fix what isn’t broken?

The ubiquity of the RSI made me truly disbelieve Jurik Research’s Relative Strength Quality Index. It felt like I was seeing someone tell me that Coca-Cola Spiced is better than Coke Classic (it’s not - don’t even DM me).

So, of course, I wanted to test it. In this post, we’ll construct 16 different RSI and RSX strategies and use the R2T Backtest Methodology to evaluate them on SPY.

Subscribe now

Recap: RSI and RSX

The Relative Strength Index (RSI) is a momentum oscillator widely used in trading strategies to gauge the speed and change of price movements of an asset. The RSI is simple to understand because it ranges between 0 to 100. For interpretation, the RSI is typically used to identify overbought or oversold conditions in the market. A higher RSI indicates more frequent upward price moves, which might indicate more buying pressure. A really high RSI might indicate that an asset is “too hot” and that its price might correct or revert to the mean. (And the opposite applies for lower RSI values)

An asset is generally considered overbought when the RSI is above 70 and oversold when it is below 30. Meaning, an upward trend is seen as likely to reverse if the RSI is above 70, and a downward trend might be ready to reverse if the RSI is below 30.

The Relative Strength Quality Index (RSX) is a variation of the traditional RSI, designed to eliminate the choppiness associated with the RSI by applying smoothing techniques, making it more responsive to recent price changes without adding lag. The RSX maintains the same scale as the RSI, ranging from 0 to 100, and is used to identify overbought and oversold conditions in the same manner.

The primary difference between the RSI and RSX lies in their sensitivity and smoothness. While both aim to measure the momentum and identify overbought or oversold conditions, the RSX is designed to produce a smoother line that reacts more efficiently to recent price changes. This reduces the false signals that traders might encounter with the traditional RSI.

Despite its smoothing, the RSX manages to maintain a low level of lag, making it potentially more useful for traders looking for timely signals. The traditional RSI, while effective, can sometimes offer signals that are either too late or too sensitive to short-term price movements, leading to premature entries or exits.

The RSX's smoothed calculation makes it more suited for volatile markets, where the RSI's signals might be too erratic. Traders operating in these conditions might prefer the RSX for its ability to filter out noise and offer clearer signals.

We can see this visually in the below visualization of the RSI and RSX for SPY:

There are multiple points where the RSI bounces around thresholds, possibly giving multiple conflicting signals. The smoother RSX does not have such a problem. In addition, the lag in the RSX is not significant in this sample at all. That is a great feat for a smoothed indicator.

But does it give a better trading strategy? Let’s find out.

Trading Rules and Backtest Results

We start with a standard reversal strategy involving three parameters:

1. Window length

2. Lower Threshold

3. Upper Threshold

With the following trading rules:

• Go long when the RSI or RSX crosses the lower threshold from below

• Go short when the RSI or RSX crossed the uper threshhold from above

Our optimizations tested this parameter space:

• Window: [2, 4, 6, … 50]

• Lower threshold: [10, 12, 14, … 40]

• Upper threshold: [60, 62, 64, … 90]

Total parameter combinations: 6400

The RSX version of this strategy was far more consistent in terms of win rate but not as rewarding:

And the correlation between OOS win rate and OOS expectancy was also slightly lower:

If we look at IS/OOS accuracy for Sharpe ratios, there’s also a slight advantage for the original RSI:

Variations: Using Risk Management

It’s worth taking these somewhat basic strategies and looking at variations with stop losses, trailing stops, and trailing take profits. This tests the idea that these strategies might be asymmetrically flagging good or bad events. Meaning, one trade that gets unlucky could wipe out your capital. Good risk management will mitigate that asymmetry.

In addition, we also look at taking less capital risk by use only long orders instead of using long and short orders, as above.

In total, we test 8 different variations of the RSI and RSX strategies, totaling 16 variations:

• Long only

• Long only with 10% stop loss

• Long only with 10% trailing stop

• Long only with 10% trailing stop and 15% trailing take profit

• Long and short

• Long and short with 10% stop loss

• Long and short with 10% trailing stop

• Long and short with 10% trailing stop and 15% trailing take profit

A few interesting observations from this:

• Both strategies tend to achieve higher Sharpe ratios when long only no matter the use of risk management

• The RSX strategy has higher median OOS Sharpe ratios than the RSI strategy in 6 of the 8 variations tested

• The highest Sharpe ratios come from the long only strategy without risk management, with the long only RSX strategy achieving the highest median OOS Sharpe ratio among all variations tested

• The long/short strategy saw the highest discrepancy between RSI and RSX with the RSX achieving worse results, which might explain the shortfalls seen in win rates, expectancy, and IS/OOS Sharpe ratios

Conclusions

The RSX is undoubtedly a less volatile and less sensitive version of the RSI, which can reduce the noise produced by RSI threshold crossovers.

When looking at unbounded long/short strategies, however, the RSI tends to provide more favorable results, indicating that bets on the RSX’s upper threshold benefit from risk management. Indeed, in all of the variations of the long/short strategy tested with risk management tools, the RSX strategy outperformed on a risk-adjusted basis.

Overall, the evidence here supports consideration of the RSX in trading strategies. The superior risk-adjusted performance on most tests indicates that the indicator gives a more reliable signal, not just a smoother ride.

Until next time, keep on the cutting edge, everyone.

Resources

The R2T Backtesting Framework, Ver. 1

The AotE Toolkit

Disclaimers

The content on this page is for educational and informational purposes only. Any views and opinions expressed belong only to the writer and do not represent views and opinions of people, institutions, or organizations that the writer may or may not be associated with.

No material in this page should be construed as buy/sell recommendations, investment advice, determinations of suitability, or solicitations. Securities investment and trading involve risks, and not all risks are disclosed or discussed here. Loss of principal is possible. You are encouraged to seek financial advice from a licensed professional prior to making transaction decisions.

Further, you should not assume that the future performance of any specific investment or investment strategy will be profitable or equal to corresponding past performance levels. Past performance does not guarantee future results.

Neural networks are the backbone of the current generation of artificial intelligence, including LLM’s like GPT and Gemini, so we can confidently say that our humble CNN-LSTM is a kind of AI.

What we will do here is illustrate how this AI model can be deployed to form an actual trading strategy and perform a backtest using the FX major pairs. Our model actually did pretty well and shows some promise!

Subscribe now

Quick Recap

We are training a CNN-LSTM model to identify uptrends in the prices of a stock, FX pair, or other security. Because neural networks need labelled data in order to “learn” its complex relationships, we used a trend signal to label historical time series. The signal is simple - it uses hindsight to identify when a hypothetical trader could have bought or sold in order to avoid drawdowns beyond a given threshold.

This neural network, then, is trying to learn how to be this hypothetical perfect trader. It takes the contexts for the optimal buy and sell points and tries to predict if a new context means that there is an uptrend in the security.

What contexts did we use? For this round, we used technical indicators on the price series - SMA, EMA, RSI, MACD, and Bollinger Bands. These formed the feature set, in machine learning parlance, for each series. The idea is that we want to see if our AI could make perfect trades using those indicators for each FX pair.

Identifying a Trend

We trained a model for each FX pair using historical daily data for 2006-2023, with 2022-2023 used as the validation set for training. Then, 2023-Jan 2024 formed our test set or holdout set, which totaled 252 bars.

Let’s look at the the model we trained on CADCHF:

There’s a lot going on in this graph, so let’s break it down.

The left y-axis plots the CADCHF price for the duration of our test period. It was mostly ranging during this period and didn’t form an actual trend. When the price line is red, our actual signal (what the model is trying to learn) has a value of 0 for no trend. When the price line is blue, the actual signal is signaling that a trend is occurring. Our signal treats the tail end of the series as a trend because we have not hit a 10% drawdown after the previous drawdown.

The right y-axis plots the models raw output for each day. Because we are using a binary classifier, the outputs range form 0 to 1 (outputs of a sigmoid activation function). We can roughly interpret these values as the model’s predicted probability that we are in an uptrend.

With CADCHF, our model is never expressing much confidence that a trend is occurring. Its most certain predictions first occur in July 2023, when the price was dropping. If we were forming a trading strategy on CADCHF alone, we would need to decide a model output threshold that would trigger a buy signal. If we wanted our model to be “80% sure” of a trend, for example, we would ignore signals below 0.8 and would not hold CADCHF at any point during this period.

So even though the model expressed its highest certainty at the wrong time, it never had a lot of conviction. I would interpret those as “IDK, maybe?”

Let’s look at a completely different case:

Both the actual prices and model predictions for CHFJPY form a completely different picture. Here, CHFJPY was actually in an uptrend the entire period. It never had a drawdown of over 10%.

And look at our model’s predictions. I would interpret that output as our robot trader holding a big neon sign saying “BUY THIS OR YOU’RE STUPID”.

Forming an AI Trading Strategy

We can use our model outputs for the 28 FX major pairs to simulate a trading strategy.

For this test, we wanted to long the FX pair with the highest conviction output from its model. We wanted to get into the most likely trend every 28 days (roughly a month). So each month, we ask our robot trader which FX pair it thinks is in an uptrend. Then, we buy that for 28 days and repeat.

What happened here was interesting. Our model had more conviction with CHFJPY than with any other pair. This strategy would have been long CHFJPY for the entire period, in other words.

And assuming buying and holding, that would have actually worked. CHFJPY outperformed the 27 other pairs for the period.

But was this the optimal pair for each rebalancing period? This graph illustrates:

The blue line in this chart shows a hypothetical portfolio over time. At each rebalancing period, we would have closed our previous position and opened another in the FX pair with the highest trend probability. The blue line shows what this strategy actually did.

An actual perfect strategy would have gone long the FX pair with the highest performance for the period. Because no other model had such high conviction, we did not buy any alternatives. CHFJPY was at or near the top of the pack in many periods, but not all. We left some outperformance on the table, in other words.

Still, these results are encouraging! We would not consider this a comprehensive backtest by any means, but it is a cause for some optimism about this modeling approach.

In future posts, we would like to explore other security sets, other feature sets, and other model architectures. There is lot more ground we can cover with AI.

Until next time, stay on the cutting edge, everyone.

Resources

Trend Following and Deep Learning, Part 1

References

Eggebrecht, Patrik & Lütkebohmert, Eva. (2023). A Deep Trend-Following Trading Strategy for Equity Markets. The Journal of Financial Data Science. 5. jfds.2023.1.120. 10.3905/jfds.2023.1.120.

Disclaimers

The content on this page is for educational and informational purposes only. Any views and opinions expressed belong only to the writer and do not represent views and opinions of people, institutions, or organizations that the writer may or may not be associated with.

No material in this page should be construed as buy/sell recommendations, investment advice, determinations of suitability, or solicitations. Securities investment and trading involve risks, and not all risks are disclosed or discussed here. Loss of principal is possible. You are encouraged to seek financial advice from a licensed professional prior to making transaction decisions.

Further, you should not assume that the future performance of any specific investment or investment strategy will be profitable or equal to corresponding past performance levels. Past performance does not guarantee future results.

“Give me six hours to chop down a tree, and I will spend the first four sharpening the axe.”
― Abraham Lincoln

Hi everyone! The past two months have been a whirlwind of activity here at the newsletter. We have been dutifully sharpening and sharpening and sharpening that axe, getting ready for the work to come. Preparatory work like this is just as important as the work itself.

And we are very, very excited about what’s to come.

We published four pieces explaining our approach to algorithm backtesting and strategy evaluation:

1. In-Sample and Out-of-Sample Testing: The backbone of our methodology, helping us find models that shine in the future, not just in the past

2. Cross Validation and Walk Forward Optimization: Exploring advanced techniques for robustness testing, their potential pitfalls, and outlining our probabilistic approach

3. Monte Carlo Simulation: Powerful reality checks via “what if?”-style simulations further cementing significance of strategies

4. Why do we need millions of backtests?: Our flag in the ground, highlighting metrics that we focus on to give you the best assessments of strategy effectiveness

These foundational posts culminated in version 1 of our "Really, Really Thorough Backtesting Methodology." This methodology is designed to be a guide for how we look at trading strategies. We want to be, well, very thorough and informative in backtesting. We see this methodology as an iterative guide. We intend to update it and engage with you all on how to best execute.

Armed with this methodology, we published results from two first research projects: the first analyzing 18 different moving averages across the 28 FX major pairs and the second conducting a sweeping backtest of the entire S&P 500, involving over 86 million simulations. This level of comprehensive research is rare outside the realms of academia and top-tier institutional firms, and we're just getting started. This is a powerful toolkit with a lot of potential. We know it, and there is a lot coming down the pipelines.

In addition to our focus on algorithm backtesting, we've also dipped our toes into the fascinating world of deep learning. While this exploration was momentarily paused to prioritize algorithmic trading strategy evaluation, the allure of deep learning in financial analysis is too potent to sideline for long. Rest assured, we will revisit this topic soon.

Looking ahead, Alpha on the Edge is brimming with potential harvests from our metaphorical seed planting. We are deeply grateful that you have decided to join us. Don’t hesitate to reach out and connect.

Until next time, keep on the cutting edge, everyone.