#54 — MTF/wave/wavelet confluence + Harvester

2026-05-27 (Sessions 076-077)

Verdict: negative · ⛔ closed — 0 WF survivors everywhere; spread is the wall

Every S5 multi-lag confluence variant (trend, counter-trend, wave-break exit, strided Haar-wavelet exit) produced 0 walk-forward survivors, and the Harvester SMA-fade's genuine 70–89% win rate sits 5–9% below the spread-adjusted break-even at every R:R.

Hypothesis — what it tests

Whether multi-timeframe/multi-lag agreement (confluence), wavelet-scale decomposition, or an SMA-distance + bar-exhaustion fade can generate an edge that clears OANDA retail spread on S5–M15 bars.

Description

A broad campaign testing whether multi-view agreement clears the OANDA retail spread. The S5 multi-lag momentum confluence was tested trend-following (0/3 pairs WF), then flipped counter-trend on the '78% WR' intuition (also 0/3 — win rate turned out ~22% in BOTH directions, so the signal predicts neither continuation nor reversal). Exit-side variants — a wave-break exit on fast-lag reversal (WR=6%, pure S5 noise) and a strided Haar wavelet exit at S30/M1/M2 effective scales (756 configs x 4 pairs, 0 WF, statistically dead at 9–14 IS trades) — also failed. The Harvester (SMA mean-reversion + bar-exhaustion + ATR gate on M5, then M15, then M15+Markov filter; ~24,000 configs total) revealed the campaign's cleanest lesson, the 'WR gap theorem': the bar-exhaustion fade signal is genuinely directional (70–85% raw WR) but its win rate is invariably ~5% below the break-even WR at every R:R tested — that gap IS the OANDA spread, and no TP/SL geometry can close it. The one v1 'success' (89% WR) survived only because a spread<1p filter reduced it to 55 trades clustered in the 2020–2022 low-spread era.

Key result

S5 confluence 0/3 WF both directions (WR ~22% either way, EV ≈ −1.03 p/trade); wave-break WR=6%; strided wavelet 0/4 WF (9–14 IS trades); Harvester v1 USD_JPY WR=89.1% but n=55/1204 days (spread<1p era only); WR gap 5–9% below break-even at every R:R — the OANDA spread in WR form

Indicators

Multi-lag S5 momentum confluence (short/mid/long close deltas)Haar wavelet details (D3/D4/D5, strided S30/M1/M2)SMA distance (mean-reversion stretch)Bar-exhaustion (n consecutive same-direction bodies)ATR magnitude gateMarkov D1 regime filterIS-P90 spread gate

Algorithms

grid search sweep (Numba prange)walk-forward validation (3 chunks)Monte Carlo sign-shuffleHaar discrete wavelet transformMarkov chain regime filter

Code

research/experiments/s5_momentum_sweep/backtest_s5_mtf_mom.py research/experiments/s5_momentum_sweep/backtest_s5_ctr_mom.py research/experiments/wave_exit/backtest_wave_exit.py research/experiments/wave_exit/backtest_wave_wavelet.py research/experiments/harvester/backtest_harvester.py research/experiments/harvester/backtest_harvester_v2.py research/experiments/harvester/backtest_harvester_m15.py research/experiments/harvester/backtest_harvester_m15_markov.py

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

← #53 H4 Donchian + FIFO stopped (live check)#55 Dynamic position sizing →

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