d3xt3rm0rgan/trading-strategy-bias-audits

Trading Strategy Bias Audits

Label-first distilled dataset for training a backtest-bias auditor: given a trading-strategy description, flag every methodological bias from a fixed 7-item taxonomy, quote the verbatim triggering phrase, and refuse any profitability verdict (even under adversarial pressure).

Splits

split	n
train	1,900
validation	100
test	1,000

Schema

Each row is a chat example: messages = [system, user, assistant]. The assistant turn is a JSON audit: {"flags": [{"bias","evidence","why"}], "clean": bool, "profitability_verdict": null, "refusal": str}

How it was built (label-first distillation)

1. Sample the ground-truth bias set (hard biases over-weighted).
2. A teacher model writes a strategy embedding exactly those biases.
3. The teacher audits it given the known labels — so it never has to detect anything (where even strong models fail), keeping gold labels near-perfect.
4. A ruthless automatic gate drops any example whose flagged set ≠ injected set, whose evidence isn't a verbatim quote, or that leaks a verdict; then dedup.

Taxonomy

survivorship_bias · look_ahead_bias · storytelling_bias · overfitting_data_snooping · transaction_cost_turnover_ignored · outlier_mishandling · short_availability_cost

The tuned Qwen3-4B model: https://huggingface.co/d3xt3rm0rgan/trading-strategy-skeptic-qwen3-4b