Auditing Loss Dependence in a Trader’s Trade History
Summary
The document describes a tool for checking whether a trader’s actions and recent outcomes are associated with subsequent losses. It trains a neural network on eight features from closed trades, then reports validation accuracy relative to a majority-class baseline, probability calibration, and permutation-based feature importance. A combined grade turns those diagnostics into written recommendations.
By default, the tool uses synthetic trades with an injected pattern of larger position sizes after losses; users can switch to closed-deal history for analysis of their own trades. The synthetic example demonstrates the workflow, but it does not establish predictive value on real trading records. The document gives no feature definitions, sample validation results, or evidence of out-of-sample performance beyond describing the diagnostic setup. Results from a trader’s own history depend on the selected features, data quality, and validation design, so the score should be treated as an exploratory audit rather than proof of causation or a guarantee of future losses.
Key ideas
- Trade-level analysis can reveal loss patterns that aggregate metrics such as win rate may conceal.
- The tool trains a neural network on eight features from closed deals to predict losses.
- Validation accuracy is compared with a majority-class baseline, and calibration compares predicted probabilities with observed loss rates.
- Permutation importance estimates feature relevance by measuring the effect of shuffling each input.
- Synthetic data demonstrates the workflow, while real-history results require careful interpretation.
Tags
This summary was written by Stratmill's research agent from the original; it is not a copy of the source.