Edge Log · Wave 05
Trade only when the evidence is strong?
No validated edgeStrict policy stayed in cash
The idea.
Before each quarter, three policies decided how to trade the same gold rule: always use it, pick the version with the higher past average, or trade only when a simultaneous lower bound on past results sat above zero. Then we judged their results together with earlier tests, as one registered family of 812.
FIXED always uses the BOTH rule. Each quarter, ROLL_MEAN ranks eligible COST and BOTH versions by mean daily R at double costs. Training uses 36 calendar months ending before the last 20 observed New York dates available before the quarter. Eligibility requires complete look-back coverage, at least 60 resolved training positions, a finite positive stressed mean, and complete mapping to the shared observed-date calendar. ROLL_LCB also requires a valid simultaneous 95% lower bound above zero, calculated across both parent rules. Each policy stays in cash if no version passes its requirements. Cash means no new trades; a trade already open still finishes.
The test.
Gold CFD for the policies; gold, Nasdaq-100 and S&P 500 CFDs in the test family · Dukascopy bid prices, 15-minute candles, January 2018 to September 2026; policies run quarter by quarter from Q1 2021 to Q3 2026; joint test on 1,167 New York dates observed in all three markets, January 2023 to September 2026 · Assumed round-trip costs: gold 0.40, Nasdaq-100 2.0, S&P 500 0.60 points (the last two only in the test family); stress test at 2×.
The result.
Each square is one quarter, Q1 2021 to Q3 2026, colored by the summed base-cost R of the positions opened in it: green above zero, red below. Outlined: cash; no new trades originated in that quarter, so their summed R contribution was zero. The first quarter's look-back was incomplete, so both rolling policies began in cash.
Each row represents one or more labelled tests with identical reported statistics and raw p below 0.05, Jan 2023 to Sep 2026. Open circle: raw p. Dot: p adjusted for the family of 812 tests (single-step max-T, 3,999 circular-block replicates of 20 dates). Eleven tests had raw p below 0.05; after adjustment none did, and no simultaneous lower bound was above zero. Results shared by several labels are shown once. Four labelled tests shared the smallest raw p, 0.0185; each had adjusted p 0.93075. They are Wave 2's H1 fade with the trend, 08:00–12:00, both filters; Wave 3's BOTH; Wave 4's BOTH; and Wave 5's FIXED_BOTH policy, all gold at base cost. A p-value is not the chance that a rule works.
What each selector changed.
Mean R per trade: +0.179 → +0.148
Jan 2023 to Sep 2026, base cost. ROLL_MEAN kept 153 of FIXED's trades, dropped 22 (−0.663 R) and added 86 of its own (+3.420 R; −2.124 R at double costs). Total: +31.269 → +35.352 R. Both simultaneous lower bounds of the difference were below zero.
Mean R per trade: +0.148 → none (no trades)
Jan 2023 to Sep 2026, base cost. ROLL_LCB took none of ROLL_MEAN's 239 trades: it skipped 125 losers (−116.571 R) and 114 winners (+151.922 R). It ended 35.352 R behind, 21.976 R behind at double costs. Cash is not a gain.
What it taught us.
In this retrospective run, the strict policy stayed in cash in all 23 quarters and executed no trades. Cash isn't profit: next to the simpler selector, it skipped 125 losing trades but also 114 winners worth more. The smallest raw p in the family, 0.0185, became 0.93075 after correcting for all 812 tests.
Next wave drops with our next Reddit post.
Evidence and limits
- Retrospective: every date and the gold rule itself had been studied before. Registering this test before running it does not make the history untouched.
- The family of 812 tests covers 400 labelled strategies from Waves 2–4 at two costs plus 12 policy streams. It leaves out Wave 1, unlogged variants and earlier human choices, so it is not a correction for our whole search.
- Some labels share identical results: the 400 labelled strategies form 390 distinct series.
- The max-T adjustment is approximate (a fixed scale, circular blocks of 20 dates, 3,999 replicates), and its coverage in finite samples may differ from the nominal level. It resamples the realized results and does not refit the selectors.
- The joint test uses 1,167 New York dates observed in all three markets, partial Sundays included; it is not a count of full trading days.
- Costs are assumptions, not measured broker quotes; no financing. R is the planned signal-to-stop risk, not an account return.
- ROLL_LCB stayed in cash throughout this run and executed no trades. A cash selection blocks new admissions; existing positions and pending entries retain their original lifecycle. Zero drawdown from zero trades is not a safe or winning strategy.
- When it decided, the confidence gate looked only at the two parent rules; the 812-test correction came afterwards and did not cause the cash result.
Provenance
- Research journal record
RVD-W5-RESULT-001 - Research journal record
RVD-W5-PROTOCOL-001 wave5/results/policy-final.json· sha2568c7e92ec2ef7…wave5/results/policy-study.json· sha25616573d5ff7ea…wave5/results/joint-audit.json· sha256ce66ce63f2b8…wave5/reports/claim-ledger.json· sha25631a3dc635fac…wave5/results/training-decisions.json· sha256e7c80f76c474…wave5/PREREG.json· sha256b196627463ca…wave5/PROTOCOL.md· sha256464f3c436e71…- Claim review record
RVD-W5-WEB-EDGE-LOG-NOTE-001