Case study / Trading behavior
Can a model learn my trading decisions?
I used my own futures trading history to study a narrow question: whether pre-entry market information and user state can reproduce the direction of an observed trade.
The model is evaluated conditional on recorded LONG/SHORT entries. It does not establish when to trade, when to pass, how much to risk, or whether a trade will be profitable.
- ML-eligible episodes
- 686
- Final chronological holdout
- 114
- Direction accuracy
- 71.05%
- Balanced accuracy
- 67.76%
Start with the decisions, not the win rate.
The reconstructed record contains 689 complete episodes. Its gross win rate was 66.47%, but the worst 1% of episodes accounted for 50.20% of all losing PnL. That concentration makes a single accuracy or win-rate statistic an incomplete account of trading quality.
I separated three questions: can the model imitate direction, can it identify better entries, and can it recognize severe losses? Success on the first question cannot stand in for success on the other two.
What I built.
I reconstructed complete trading episodes, aligned pre-entry market and user-state features, and applied public-data coverage checks. The final dataset contained 686 ML-eligible episodes.
The newest 114 BTCUSDC episodes formed a chronological holdout kept out of fitting and parameter selection. The reported HGB + user-state model achieved 71.05% direction accuracy, 67.76% balanced accuracy, and MCC 0.353 on this final conditional LONG/SHORT evaluation.
What the evidence did not establish.
Final entry-quality ROC-AUC was 0.5583. Tail-loss PR-AUC was 0.1364. These diagnostics do not establish a profitable-entry model or a reliable risk veto. The account’s returns are not attributed to this model.
The historical data also reflect occasions on which a trade was actually entered. A model learned from those entries alone does not describe all the times I chose not to trade.
The next connection: observe decisions as they happen.
A read-only observer was introduced to collect forward action and market-state evidence. That creates a separate prospective record rather than quietly treating the historical holdout as live validation.
The research direction connects naturally to market microstructure: even a plausible directional decision can have poor execution economics. Fill uncertainty, costs, and post-fill price movement need their own tests.
Evidence and provenance.
This case study reorganizes the behavioral results already published in the portfolio’s September 24, 2026 update. It introduces no new empirical measurement or outcome-based retuning. The underlying account record, research sample, and forward observations are different datasets.