#71 — Regime-conditioned entry Phase A (triple-barrier, against-drift)

Hypothesis — what it tests

Whether a 7-feature causal regime vector (drift/efficiency-ratio/volatility/envelope-slope/leg-structure/variance-ratio) computed on a closed 10-minute S5 window carries directional information that survives triple-barrier random-entry cost accounting, using a coin-flip control arm and dual information+money gates rather than the naive 66.7% barrier baseline.

Description

A 7-feature intra-bar regime vector (drift, efficiency ratio, realized vol, peak/trough envelope slopes, leg expansion, variance ratios at k=2/4/8, leg count) computed on a closed 60-slot S5 window drives a triple-barrier (TP/SL/timeout) random-entry test across against-drift, with-drift, and coin-flip arms, at TP in {1.5,1.8,3.2,4.0}p and horizon in {2,4} M5 bars. The primary arm was pre-registration-amended from with-drift to against-drift given 7 prior negative with-drift studies. A mid-run amendment (Amendment 2) found that at the 2-bar vertical barrier ~80% of trades timeout, and among decided trades win rate inflates to 82-85% on pure random walks (verified by simulation/shuffle controls) — invalidating the 66.7% gambler's-ruin baseline as the null and replacing it with the coin-flip arm plus a net-ECN-expectancy money criterion. IS results (first 70%, 12 pairs, confirmatory cell against/hiER/TP3.2p/h2): the information criterion passed (against beat coin gross on 12/12 pairs, with-drift worst on 12/12, exactly the amended prediction), but the money criterion failed (0/12 pairs net-positive at ECN 1.0x cost; best -0.57 p/trade; gross tilt 5-10x below the 0.7p ECN floor, ~20x below OANDA cost). The user elected to keep OOS sealed rather than spend it on a foregone failure, preserving one future confirmatory shot at a redesigned exit (wider TP / passive ECN entry).

Key result

Indicators

Algorithms

Code