Dispatch from the Edge - March 2024
Hi everyone, and welcome to the new subscribers that have joined us! This is a monthly dispatch where we take stock of the previous month’s research, offer some commentary, and preview where we’re going.
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.