Edge Log · Wave 02
Can a look-back selector pick hours and filters?
No validated edgeNo candidate qualified
The idea.
Before each quarter, a selector looked back 36 months and chose one New York time window and one filter for the quarter ahead, or no trade at all. We asked whether choosing this way survives costs and our qualification checks.
The Wave 1 range patterns, limited to one of four New York signal windows (00:00–04:00, 04:00–08:00, 08:00–12:00, 12:00–16:00) and one of four filters: none, cost (assumed cost small next to the planned risk), range (previous range wide next to recent volatility) or both. Before each quarter, each selector ranked the recipes that passed its training checks by a score built from their double-cost results over the previous 36 months, and froze its pick; if none passed, it stayed in cash. BASE chose among 32 unfiltered recipes per market, EXPANDED among all 128. Entry, stop and target follow Wave 1; a position still open when its window ends exits at the next bar's open.
The test.
Gold, Nasdaq-100 and S&P 500 CFDs · Dukascopy bid prices, 15-minute candles, January 2018 to September 2026; selectors scored quarter by quarter from Q1 2021 to Q3 2026 · Assumed round-trip costs: gold 0.40, Nasdaq-100 2.0, S&P 500 0.60 points; stress test at 2×.
The result.
Each square is one quarter, Q1 2021 to Q3 2026. Filled: the selector traded; green when the positions it opened that quarter summed above zero at base cost, red when below. Outlined: cash. No recipe passed the training checks, so no new trades originated in that quarter; their summed R contribution was zero.
Dot: base-cost mean R per trade. Bar: descriptive 95% block CI. Ring: mean at double costs. Every lower bound is below zero. Gold · base had the fewest trades (13) and the widest interval, −0.749 to +1.126. The S&P 500 base selector stayed in cash in all 23 quarters and has no trades to plot.
Gold's expanded selector, with its limits.
Sum of base-cost R from the positions opened each year, not account return; 2026 ends in September. Chosen for display after seeing the results: it had the highest base-cost total of the six selectors. Its first two years summed −14.841 R.
Gold · expanded, Q1 2021–Q3 2026 · 264 trades
The same 264 trades, Q1 2021 to Q3 2026: mean R per trade falls from +0.062 to +0.005 when assumed costs double.
What it taught us.
None of the six selectors qualified. Gold's expanded selector had the highest base-cost total, its base-cost 95% block CI included zero, and at double costs it averaged +0.005 R per trade. The S&P 500 base selector never found a recipe to trade.
Evidence and limits
- Retrospective: every date here had been studied before, including in Wave 1, so this is not an untouched out-of-sample test.
- Costs are assumptions, not measured broker quotes. No swap or financing.
- R is the planned signal-to-stop risk, not an account return.
- The windows are New York clock windows with daylight saving time, not exact exchange sessions.
- The descriptive 95% block CIs use 2,000 moving-block replicates of 20 observed trading dates, including days without trades. They do not adjust for the 384 cells tested here or for earlier searches.
- Qualifying needed every check to pass, among them a 95% block CI lower bound above zero, a positive double-cost mean in 2025–Sep 2026 and no single year above half of the positive gross R. No selector passed them all.
- The S&P 500 base selector never traded, yet its paired difference against the prespecified fixed baseline (H1 follow, 08:00–12:00, no filter) had a 95% block CI above zero, because that baseline averaged −0.065 R per trade. Beating a losing baseline is not making money.
- All 384 cells and all six selectors are kept, failures included. Being picked by a selector is not the same as qualifying.
Provenance
- Research journal record
RVD-W2-RESULT-IMPORT - Research journal record
RVD-W2-PROTOCOL-IMPORT wave2/results/study.json· sha2561f4fa7f32a4b…wave2/PROTOCOL.md· sha2562a11bea0404d…- Claim review record
RVD-W2-WEB-EDGE-LOG-NOTE-001