/stockbee-20pct-study

Cohort Mining Rules

Purpose

Cohort summaries convert many 20% mover observations into evidence prompts. They should reduce hindsight bias, not create overfit rules.

Minimum Evidence Rules

• Require min_sample before promoting any rule candidate. Default: 10 matured records.
• Treat samples below 30 as preliminary unless the pattern is very stable and explainable.
• Split by market regime before increasing risk based on a cohort.
• Review representative charts from both winners and failures.
• Prefer simple groupings before adding more filters.

Default Cohort Dimensions

direction,catalyst.label,technical_context.pattern_label,technical_context.close_quality

Useful alternate groupings:

direction,technical_context.pattern_label,technical_context.extension_risk
catalyst.label,technical_context.close_quality,liquidity.volume_ratio_20d
direction,data_quality.flags,technical_context.pattern_label

Rule Candidate Thresholds

A continuation-favorable cohort can become candidate_for_review when:

• Sample size is at or above min_sample
• Direction-adjusted win rate is at least 58%
• Median direction-adjusted return is at least +2%

A weak cohort can become avoid_or_fade_study when:

• Sample size is at or above min_sample
• Direction-adjusted win rate is 42% or lower
• Median direction-adjusted return is -2% or worse

Overfitting Controls

• Do not add filters solely because they improved one backtest.
• Do not mine hundreds of tag combinations without recording the number of trials.
• Do not use future catalyst knowledge during historical classification.
• Do not count records with PENDING outcomes as matured.
• Do not mix current-universe-only and survivorship-complete backfills without a data-quality split.

Promotion Path

1. stockbee-20pct-study produces a rule candidate.
2. Human reviews charts and data-quality notes.
3. edge-hint-extractor or edge-candidate-agent converts it into an explicit research ticket.
4. backtest-expert validates the hypothesis with realistic execution assumptions.
5. monthly-performance-review decides whether to accept, reject, or keep monitoring.