Backtesting and Machine Learning for a Self-Developed Futures Trading Strategy

This study investigates whether the poor backtesting performance of a discretionary futures trading strategy stems from important subconscious rules that were never formally documented. Testing ten years of Nasdaq-100 and S&P 500 futures data showed that adding an previously unwritten market-structure rule substantially improved performance, while machine learning found short-term price direction largely unpredictable but volatility expansion moderately predictable.

Having traded futures discretionary for about a year as most in this position I noticed that written rules would not reproduce in results. The usual explanation is that a trader's memory is fallible. I wanted to investigate another: the incompleteness of the written rules since the real most important rules are subconscious, applied without being noticed. I have tried to encode my strategy for Nasdaq-100 and S&P 500 futures exactly as I have written it, matched it against ten years of one-minute prices from 7.07 million bars. When traded it lost money: compared to break even probability 33.3%, 31.7% of trades were profitable with expectancy of -0.175R per trade over 205 trades. I have tried to switch parameters and none of the 24 of them resulted in profit. I looked at the only section of specification written without any measurable definition and what it depicted was the market structure or session-structure model: Accumulation-Manipulation-Distribution, that states whether to trade a day or not. Having added it without altering all other inputs resulted in trading 53.0 % of trades profitably with +0.480R of expectancy on 742 trades, positive on all five years of testing. In a separate analysis I have had a look at 300 101 samples through machine learning methods where it resulted in unpredictable short term price direction and predictable movement (expansion of volatility) with 57.6 % of hits against the base probability of 48.2%. The one rule I have never bothered to write down mattered most.

Explore More!

Research Paper

Source Code