19 strategies vote. The engine fuses them every 5 minutes. 10 paper plans test parallel hypotheses on Polymarket. An adaptive layer measures outcomes, learns which signals work, and re-weights everything on a tight feedback loop.
Decisions flow downward; learning flows upward. Every 5 minutes a new decision is born; every 15 minutes the adaptive layer rewrites the rules.
Different cadences. Different noise profiles. Different feedback purposes.
The engine's 5-min directional prediction. Resolved against actual BTC price moves at 5m, 30m, and 1h horizons.
Paper trades on Polymarket binary markets. Resolves with fees + slippage included — the closest measure of real edge.
Parallel paper book on Binance with TP/SL — pure directional outcome on continuous price.
Each strategy returns a directional vote. The engine fuses them by category weight.
The engine doesn't stay still. Six self-tuning mechanisms continuously recalibrate from rolling outcomes.
Pearson correlation between each strategy's signal and the realised price move at its natural horizon.
If a strategy's IC goes deeply negative — instead of dampening, the engine flips its sign. An anti-predictor becomes a predictor.
Asymmetric Schmitt-trigger thresholds prevent flip-flop on regime noise. Sticky state persists across ticks.
Reality-check on engine confidence. 10 buckets compare predicted vs. actual win rate; multiplier scales conf at gate time.
Each plan can de-rate or boost signals by their TV alert TF (1m/5m/15m/30m/1h/1d). Auto-suggested by Bayesian shrinkage on rolling-50 PM win rate per TF.
Per-plan auto-pause when rolling cumulative PnL drops below threshold. 24-hour cool-off then auto-resume.
Each plan is a controlled experiment. Same signal stream, different gates and exits.
Kelly (default) — aggressive when edge × confidence are high. Risk Parity (opt-in) — constant-dollar risk; size ∝ 1/atrPct. Calm market = larger size; choppy = smaller.
From bootstrap warm-up to circuit-breaker pause and back.
Each btc5m_decisions doc keeps accumulating truth. By 1h, we have 3 measurement points — IC reads from whichever horizon matches the strategy.
Every signal passes through this funnel per plan. Each gate logs its reason in polymarket_decisions.
13 phases over a single session. Every layer earned its place.
Rolling-50 expectancy + sigmoid weights + EWMA blend
RSS → Claude Haiku 4.5 → BTC sentiment score / 30 min
Pearson r per strategy, weight multiplier from IC tier
predictionEngine scales conf by bucket multiplier
Constant-dollar risk · size ∝ 1/atrPct
Anti-predictor → predictor when IC ≤ −0.25
Sticky state, asymmetric thresholds
Per-plan auto-pause + 24h cool-off
Weekly p10/p25/p50/p75/p90 observability
useEngineFlip / useIcWeights / applyCalibration
1m–1d × 0.10–3.00 multiplier per plan
Bayesian shrinkage on rolling-50 PM WR per TF
5m + 30m + 1h — IC measures each strategy on its natural horizon
Quick reference for the terms that matter.
Pearson correlation between strategy signal magnitude and realised price move. 0.20 is excellent for noisy 5-min crypto.
Per-confidence-bucket scaling: actual_WR / predicted_WR. Engine claims 60%, right 9% → multiplier ≈ 0.15.
While plan trades < BOOTSTRAP_TRADES, strategy gates bypassed so plan can accumulate sample.
myProb − (price + FEE). Positive = expected-value trade.
Plan's blended probability estimate. Signal mode: 0.5×wr + 0.5×conf.
Asymmetric entry/exit thresholds that prevent flip-flop on noise.
size = TARGET / atrPct. Constant-dollar risk regardless of underlying volatility.
Adding prior pseudo-observations to small samples so a 1-trade 100% WR doesn't dominate.
new = 0.7×old + 0.3×computed. Smooths weights so they don't flap every cron tick.