#51 — Post-shock counter-trend retrace live

2026-05-25 (Session 074) backtest + live deploy on account 009; taken offline 2026-06-17, container stopped 2026-06-21

Verdict: positive · ✅ validated & deployed live (edge later decayed live; offline)

Large S5 velocity shocks (z>2.5) are followed by a tradeable counter-trend retrace after a ~44-bar (220s) peak window: 72 OOS-positive configs, WF 12/12, all mc_p=0.0000, +56 p/d at the zero-inflation fill model.

Hypothesis — what it tests

Whether large S5 velocity shocks (z-score > 2.5) systematically overshoot and retrace, so that fading the move from its peak after a ~220-second observation window is a bounded-risk, spread-surviving edge.

Description

The hypothesis: after a sharp price shock (30-second velocity z-score above threshold), price retraces counter-trend from the shock peak. The backtest detects a shock, waits up to 44 S5 bars (~220s) for the peak, then enters opposite the shock direction with TP=20p and SL=30p. A three-phase pipeline (OOS sweep of 225 configs → walk-forward over 3 sub-chunks → 1000-shuffle Monte Carlo) produced 72 OOS-positive configs, with the top 40 all passing WF 12/12 and all 20 WF-passers at mc_p=0.0000; the honest sd=3 (zero fill inflation) benchmark gave +56 p/d. The 44-bar window independently matched the MSP wavelet study's D3→D5 cross-scale correlation peak at 110–220s. Deployed straight to live money (fx-retrace-live, 4 JPY pairs, eventually account 009) with a causal rolling-deque z-scorer and an IDLE→MONITORING→IN_POSITION state machine. Live it accumulated roughly +110p over ~250 trades (~+0.44 p/trade) but decayed to ~+3–4 p/d versus the +56–70 OOS estimate as volatility compressed, and was taken offline 2026-06-17; a validated M5-ATR≥5 entry-gate rescue (+79.4 p/d OOS) was kept paper-only.

Key result

72 OOS+ configs / 225; WF 12/12 on top 40; all 20 WF-passers mc_p=0.0000; +56 p/d OOS at sd=3 (zero inflation), +337 p/d at sd=1 (~2-pip inflated); live ~250 trades +110.3p (+0.44 p/trade) then decay to ~+3-4 p/d; ATR≥5 gate rescue +79.4 p/d OOS, WF 11/12, mc_p=0 (paper-only)

Indicators

S5 velocity z-score (30s velocity, MAD-normalized)shock-peak detection windowIS-P90 spread gate

Algorithms

grid search sweepwalk-forward validation (3 sub-chunks)Monte Carlo sign-shuffle (1000 shuffles)causal rolling-deque state machine

Code

research/experiments/shock_straddle/backtest_shock_straddle.py research/experiments/shock_straddle/backtest_post_shock_retrace.py research/experiments/shock_straddle/backtest_retrace_atr.py services/strategy_retrace_live/main.py services/strategy_retrace_paper/main.py services/strategy_retrace_atr_paper/main.py

Interactive version (search, filters, figures): the experiment explorer.

← #50 Price momentum M15+M5 live #52 Markov D1 regime filter →

This page is one experiment from the audited record behind the book The Edge That Wasn't — get the book · about the project.