Using Behavior Cloning to Model Subjective Trading Decisions
Summary
The author describes why converting discretionary trading into fixed rules can be difficult: decisions may combine many implicit features and weights, and traders adapt as market conditions change. The post compares this challenge to recognizing a familiar person in a photograph, which can be easy for a human but hard to express as explicit conditions. The author says their own rule-based work has mostly captured narrowly defined market situations.
As a possible approach, the author proposes behavior cloning, a machine-learning method inspired by robotics and autonomous driving. The suggested setup is to provide formatted market data alongside position and trade records, train a model to imitate the trader’s actions, then assess it on validation data. This is a proposal and request for advice, not a demonstrated trading system: no dataset details, model design, validation results, or evidence of profitability are supplied. The post also notes that market data may be simpler in form than image-rich driving data, while acknowledging that building a working quantitative process remains unresolved.
Key ideas
- Discretionary trading decisions may depend on implicit features and changing judgments that are difficult to encode as fixed rules.
- The author proposes behavior cloning as a way to learn trading actions from examples.
- The suggested training inputs pair market data with position and trade records.
- Validation data could be used to assess how closely a model imitates the trader.
- The post reports no implementation or evidence that the proposed approach works in live trading.
Tags
This summary was written by Stratmill's research agent from the original; it is not a copy of the source.