Genetic Programming for Mining Diversifying Commodity CTA Signals
Summary
This study explores genetic programming as a way to discover formula-based signals for commodity futures trading. It adapts a genetic programming package to search combinations of market data and operators, aiming to find signals that could complement familiar trend and reversal strategies. The motivation is that established approaches may become crowded, while automated search can explore patterns beyond those a researcher might specify by hand.
The study tests one-minute data for 40 liquid commodity futures and presents signals for nine instruments. The signals include trend or reversal patterns as well as measures related to volume and open interest. To address overfitting concerns, it evaluates parameter ranges and reports that the showcased signals were profitable across the tested parameters, with selected settings performing above average rather than being the single best. It also combines signals across instruments to reduce the risk of relying on one market or rule. Under stated assumptions of no leverage and one-sided transaction costs of 0.03%, the portfolio’s reported annualized return was 25.26%, its Sharpe ratio was 2.25, and its maximum drawdown was 10.51% from 2015 onward. These are backtest results; the document does not establish that they will persist out of sample or in live trading.
Key ideas
- Genetic programming searches combinations of data inputs and operators to generate candidate trading signals.
- The study applies the search to one-minute data from 40 liquid commodity futures and presents results for nine instruments.
- Candidate signals cover trend and reversal behavior as well as volume and open-interest information.
- Parameter-range checks are used to address, but cannot eliminate, the risk of overfitting.
- Combining signals across instruments is intended to reduce single-signal drawdowns.
Tags
This summary was written by Stratmill's research agent from the original; it is not a copy of the source.