The Edge That Wasn't
All 81 experiments › #63

#63 — Three measurement failures inverting verdicts (ledger, test proxy, risk arithmetic)

2026-06-17
Verdict: mixed · 🟡 mixed — meta-finding: the apparatus itself lied three ways
In one day, three independent measurement errors were caught, each inverting a decision: a P&L-recording bug made a +287p live pair look dead, a TP-proxy backtest made a dead pair look alive, and a JPY pip value 110× too large had driven a wrong risk cap.

Hypothesis — what it tests

Whether the project's own measurement layers — the live trade ledger, backtest exit proxies, and position-risk arithmetic — report the truth; each of the three was found capable of inverting a keep/drop or sizing decision on its own.

Description

A cluster of same-day discoveries showing that the measurement apparatus — not the market — can invert a verdict. (1) Ledger: the live SMA-Stack's PSAR-exit close handler discarded the broker fill price, so 17 GBP_USD trail-winners were logged as 'closed_unknown' / 0 pips; an OANDA backfill recovered +256.8p, revealing GBP_USD's true live record as +286.8p at 61% WR (the strategy's #2 pair) when the ledger had shown +30p/6% WR — a drop candidate. (2) Test proxy: a TF-separation sweep using a TP20 testbed exit crowned GBP_JPY H1/H4 as a winner (+2.04 OOS p/d), but re-running with the EXACT live PSAR exit over ~2.85 years gave OOS −3.41 p/d — the proxy exit faked the edge, and GBP_JPY proved dead at every speed and exit (it was dropped; GBP_USD was kept). (3) Risk arithmetic: risk documents had used a JPY pip value of ~$0.0069/unit/pip when the broker-verified truth is $0.0000626 — 110× too large — making the 200-pip emergency fence look like a 2×-NAV account-killer when it is really ~1.9% of NAV, which had motivated an unnecessary MAX_UNITS=30 cap. The audit table's script link points at research/experiments/gbpusd_regime; the actual fix/evidence scripts live in research/experiments/sma_exits.

Key result

Ledger backfill recovered +256.8p → GBP_USD true live +286.8p @ 61% WR (had shown +30p/6% WR); GBP_JPY TP20-testbed 'winner' +2.04 OOS p/d → −3.41 OOS p/d under the real PSAR exit (dead at every TF ratio and exit); JPY pip value error 110× ($0.0069 vs true $0.0000626/unit/pip) — 200p fence is ~1.9% NAV, not 2× NAV; GBP_USD S30/M1 PSAR OOS +2.67 p/d @ 92% WR (keep, don't widen)

Indicators

SMA stack (multi-TF alignment)PSAR trailing stopfixed TP (20p testbed proxy)

Algorithms

broker-ledger backfill/reconciliationwalk-forward backtestingTF-ratio grid sweepexact live-exit (PSAR) replay

Code

research/experiments/gbpusd_regime/diagnostic.pyresearch/experiments/sma_exits/gbpusd_oanda_backfill.pyresearch/experiments/sma_exits/gbpusd_psar.pyresearch/experiments/sma_exits/gbpjpy_h1h4_psar.pyresearch/experiments/sma_exits/gbpjpy_fast.pyresearch/experiments/sma_exits/tf_separation_sweep.py

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

← #62 Money management cannot manufacture edge (random vs trend entries)#64 FX stat-arb (Avellaneda-Lee eigen-residual) →
This page is one experiment from the audited record behind the book The Edge That Wasn't — get the book · about the project.