#56 — Portfolio variation campaign (8 variants)

2026-05-28 (Session 079); variants 2-5 deployed to accounts 001-004 the same session

Verdict: mixed · 🟡 mixed — +250.4 p/d combined, but 130/130 passed MC (the tell)

A greedy correlation-capped portfolio of 8 momentum/exhaustion variants posted +250.4 p/d combined OOS at mc_p=0.0000 — and the fact that all 130 MC candidates passed was itself the warning that the gate had no power.

Hypothesis — what it tests

Whether a decorrelated portfolio of momentum-family variants (different timeframes, lags, TP levels, and a distinct exhaustion-continuation signal) smooths the combined equity curve and multiplies deployable p/d.

Description

The idea was to smooth the equity curve by combining decorrelated variations of the proven momentum family, spanning small-bite (TP=3-5p) to large-bite (TP=20-25p). Three sweeps — a TP frontier on the two deployed signals (288 configs), a bar-exhaustion continuation family (900 configs), and a full SMA x timeframe x lag x TP sweep (5,600 configs) — fed 130 candidates into 2000-shuffle Monte Carlo validation, and a greedy selector built a portfolio under a pairwise-correlation cap of 0.70. The 8 selected variants combined to +250.4 p/d OOS with portfolio mc_p=0.0000, and variants 2-5 were deployed live on repurposed accounts 001-004. The retrospective verdict is mixed: a 100% MC pass rate (130/130) was 'the tell' that the sign-shuffle gate has no power against this failure mode — the whole family was TP-only/no-SL, and the finite-margin re-validation (entry 57) showed the edge flips from +27.3 to −206 p/d the moment any stop is added, with all 6 simulated accounts wiped under realistic margin. The campaign's real product was the demonstration that the validation pipeline, not the signal, was the weak link.

Key result

Combined +250.4 p/d OOS (8 variants x 12 pairs), portfolio mc_p=0.0000, all 8 variants mc_p<0.025; 130/130 MC candidates passed (the tell); 2,536/5,600 variant configs all-12-pair WF-positive (45%); later finite-margin check: +27.3 → −206 p/d with a 30p stop, 6/6 simulated accounts wiped

Indicators

SMA16/SMA8 momentum tripletsRaw multi-lag price momentumBar-exhaustion continuation (n_consec same-direction bodies + SMA distance)TP ladder (3-50p)IS-P90 spread gate

Algorithms

grid search sweepMonte Carlo sign-shuffle (2000 shuffles x 130 candidates)greedy portfolio selection under correlation cap (<0.70)walk-forward validation

Code

research/experiments/portfolio_sweep/backtest_tp_frontier.py research/experiments/portfolio_sweep/backtest_exhaust_cont.py research/experiments/portfolio_sweep/backtest_variants.py research/experiments/portfolio_sweep/portfolio_rank.py research/experiments/portfolio_sweep/risk_revalidation.py research/experiments/portfolio_sweep/results/portfolio_final.csv services/strategy_pmom_h1m30_live/main.py services/strategy_sma16_m30m15_live/main.py services/strategy_exhaust_a_live/main.py services/strategy_exhaust_b_live/main.py

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

← #55 Dynamic position sizing #57 Finite-margin re-validation of the live momentum book →

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