Strategy Discussion: 698

It appears to be a tree with thousands of nodes and hundreds of constant values, perfectly fitted to the historical data. Are you trying to verify whether a model that was perfectly fitted to the past will also generalize to the future?

Hi Jason, could you please publish the content of self.tree as text?

Really interesting strategy, thanks for sharing it publicly. Three questions from digging through the tree:

1. The description mentions weighting "proportional to a metric such as inverse volatility", but I can't find any in the tree — of 765 weight nodes, 756 are equal-weight and 9 are constant-weight. Max drawdown is listed in your description above as a headline indicator too but appears only 6 times. Is this the same version the results came from, or did I miss something?

2. How was the tree built and fitted to the historical data? 2,666 nodes, 377 if-branches, and thresholds like 0.621995, with RSI cutoffs spread across 0.74 / 0.77 / 0.79 / 0.81 / 0.84 in different branches. Six decimal places suggests a tool rather than hand-tuning — which one?

3. It often holds both sides of the same trade. On 2022-10-04 and 2022-10-27 the whole book is 50% BIL + 25% TQQQ + 25% SQQQ — effectively 100% cash while paying two expense ratios. Same on 2021-09-14 (15% SOXL + 15% SOXS) and 2021-10-19/20/21. Across the backtest SOXL+SOXS coexist on 14% of days. Intentional vol-dampening, or a side effect of the four sleeves being equal-weighted against each other?

Good luck with the out-of-sample test! 🍀

Hi Jason/Jowalz

Great questions.

1/ Some inverse-vol weightings were replaced with equal weight in this version. 

I'm not sure if I understand the drawdown question. The Max DD in the description is generated by QC by running a 5 years look back. The drawdown you see in your back-test is looking back at 1-1-2020, which include COVID crash. So 2 different versions.

2/ The tree aren't meant to fit historical data. It's an combination of simple RSI/SMA/Cumm Return weighting etc… as mentioned in the title. Eugene simplified the tree, then translated into human readable version. You can check out his version here: https://www.quantconnect.com/terminal/cache/embedded_backtest_220e5a4a7c472b9ecb63a5ead610e6ae.html

The decimal you saw are python rounding fraction to float. No fine-tuning. 

3/ Side effect of equal weighting of different strategies. 2 strategies effectively cancelling each other out when there is disagreement. In production version, improvement can be made to net the inverse pairs.

Let see if OOS hold true.

Thanks for taking the time on all of this — genuinely appreciate you sharing both the strategy and Eugene's version. And thanks for clearing up the inverse-vol point.

Tree Construction & Scale The main thing I'd love to understand is how the big tree came together, mostly because I'd like to learn the method rather than just run the result.

Eugene's version: 58 tickers

Published version: 118 tickers across 2,666 nodes

That's a big gap for a "simplification," so I'm curious which direction the derivation ran — did Eugene or you reduce the large tree by hand, or did the large one grow out of something closer to Eugenes? Either way, I'd be interested in what the workflow for the tree looks like, as building something at that scale by hand seems like a lot of work.

Constants & Parameter Generation One small thing that made me wonder about an automated generator in the tree chain is the specific constants:
e.g. Exact constants: 0.621995 and 0.049226
e.g. RSI cutoffs across branches: 0.74 / 0.77 / 0.79 / 0.81 / 0.84

These examples from the tree look like clean, deliberate values rather than standard float rounding artifacts (like Python's 0.1 + 0.2 = 0.30000000000000004). Not a criticism at all — just genuinely curious where they came from or how they were derived.

Will be watching the OOS. 🍀

Thanks again for engaging with all this!

jowalz I guess you'll need some Trading-strategy graphs—convert strategy logic into decision trees or signal-flow diagrams. For example, look for a custom Roslyn → graph converter. Or just use QuantConnect itself. QuantConnect provides excellent performance visualization, automatic order-fill plotting, custom charts, Alpha Framework modules, and a newer strategy-prompt interface. Its charting system visualizes what a strategy produced, not the complete logical structure of the source code. QuantConnect charting documentation

Henry H How did you come up with big tree or rules ? was AI used to generate this based on historical data ? otherwise it seems too much work to come up with something of this nature