larrylarp3-btc-avg9-1h
Not financial advice. Demo / paper trading only until a positive after-cost record exists.
Look-ahead disclosure: the feature dpo is the ORIGINAL MT5 centred Detrended Price Oscillator (pandas-ta dpo(length=21), centred), which uses 11 future bars. Holdout metrics of models that use it are inflated versus live, where dpo is not yet known on the newest 11 bars (live it is edge-filled with the last computed value; every fill is written to logs/flags.log). dpo_causal / dpo_causal_z are causal companions. Ablations without dpo show no edge after costs (negative average net return per trade on every timeframe).
- Binary P(up in 4 bars), group
avg9, timeframe 1h, trained 2026-10-22T05:57:38; features (9): price, atr, ema_diff, rsi, garch_vol, macd_diff, vwap, price_vwap_diff, bb_position - Uses strict centred dpo: no (fixed MT5 subset). Feature source: original MT5 script via
mt5_utf8(MANIFEST sha25631c328326dd9ec2ddbb818985d1dbdf9469dc6f2fca6dd11305a8360a4eb55d3). - Walk-forward OOS: accuracy 0.5181 (majority 0.5098), AUC 0.5253; threshold 0.6: futures taker 9040 trades avg net -0.120%, spot taker 9040 trades avg net -0.820%.
Execution gate (hard rule, 2026-10-22): trade only if 0.77 <= p_hit < 0.9888 and EV after costs > 0; p_hit >= 0.9888 = HOLD + flag (overconfident). Defensive mode never changes the gate.