online-learning

name	online-learning
description	Mutable parameter families with bounds, shadow→canary→promote pipeline, and statistical gates for parameter promotion.

Online Learning

The system mutates a small set of bounded parameters weekly based on closed-trade stats and post-mortems. Hard risk caps are NEVER mutated by this loop.

Mutable parameter families

Family	Examples	Bounds
sizing_aggression	r1_max_pos_pct, r3_dollar_floor	conservative ≤ value ≤ aggressive
stop_distances	default_stop_atr_mult, option_premium_stop_pct	0.5× ≤ value ≤ 2.0× current
entry_filters	min_setup_quality, min_iv_percentile_to_skip	exposed ranges per family
regime_thresholds	crisis-overlay vol percentile	0.75 ≤ value ≤ 0.95
candidate_count	max_candidates_per_scout	1 ≤ value ≤ 12

Pipeline

1. Spec — Learning Critic proposes a parameter change with rationale, expected impact, and min_canary_trades.
2. Synthesis — System validates bounds, anti-thrashing (no change if same param touched in last 4 weeks).
3. Implementation — New value committed as a param_versions row with mode=CANARY.
4. Validation — N trades run with the canary version. Statistical gates checked:
   • Wilson 95% CI lower bound on win rate ≥ control's mean
   • Profit factor ≥ control's profit factor × 0.95
   • Max drawdown ≤ control's max drawdown × 1.1
5. Analysis — If gates pass, promote to mode=ACTIVE. If fail, revert.

Shadow vs canary

• Shadow — proposed parameter runs in parallel with control on the SAME trades; only logged, not actuated. Used to gather signal cheaply.
• Canary — parameter actually drives a slice (e.g. 25%) of trades. Statistically gated.

Hard rules

• N_canary_min ≥ 20 trades for any parameter family before considering promotion.
• No more than 1 active canary per family at a time.
• Risk hard-caps NEVER mutated; only "soft" parameters above.
• If a canary fails its drawdown gate within the first 5 trades, immediate auto-revert.