Model shadow. Naive = existing old30 model + Binance spot/futures L1 30m drawdown. Full = Naive + tested 1s absorption overlay. Shadow is read-only and never places orders. It needs 30 minutes of continuous Binance depth/trade history after start or reconnect. Full hit compares every prediction; signal hit compares each model's own frozen select threshold. Shadow cumulative PnL uses the same Binance 5m direction proxy, 5 shares per model signal, fixed 0.53 entry, and 7% taker fee; it is not wallet PnL.
Old vs Naive vs Full Shadow
Shadow Daily Detail, SGT
Shadow Weekly Detail, SGT
Recent Shadow Predictions
Definitions.
Definite fill rate = matched live orders / all live order attempts.
Possible fill rate = matched orders plus maker orders whose cancel result was not_canceled or order-not-found / all attempts.
Prediction win rate = signal side matched Binance spot 5m open-close direction.
Trade win rate = resolved order proxy PnL greater than zero.
Signal ask bucket groups rows by the ask seen when the executor decided. For ask above 0.53, the current strategy still used maker fallback at 0.53, not taker above 0.53.
Observed fee PnL uses the actual executor price and mode. Hypo taker fee PnL assumes every row in that bucket was taker-filled at the observed ask.
Experiment A = even round old max_ask 0.53. Experiment B = odd round new max_ask 0.55.
Shadow ask band records signals inside the configured ask band without sending an extra order. Its PnL is hypothetical taker-at-observed-ask PnL; real fill rate still needs live orders or depth replay.