Open Harmonics Get the app →

Open Harmonics · Experiments

Supported · Entries · 2026-10-11

What should 'random entry' mean? A price-blind reference control

What this means

The cleanest question is: if I had bought this stock on any day that year with the same stop and target, would I have done as well? On daily, yes: harmonic entries are no better than a coin-flip day (63% of random traders match them). On 4h, also yes. On 1h, harmonics do beat a coin-flip day by about 0.06R a trade, in both halves, and almost no random trader matches their total. The earlier +0.05 daily edge came from a control centred on dips.

Explanation checked against the data.

Daily: totals of 10,000 truly random traders (any day in the same year, same stock, blind to price) with the harmonic total marked.
Daily: totals of 10,000 truly random traders (any day in the same year, same stock, blind to price) with the harmonic total marked.
1-hour: totals of 10,000 truly random traders (any day in the same year, same stock, blind to price) with the harmonic total marked.
1-hour: totals of 10,000 truly random traders (any day in the same year, same stock, blind to price) with the harmonic total marked.
4-hour: totals of 10,000 truly random traders (any day in the same year, same stock, blind to price) with the harmonic total marked.
4-hour: totals of 10,000 truly random traders (any day in the same year, same stock, blind to price) with the harmonic total marked.
Hypothesis
A random entry is a fair benchmark only if it picks its day blind to price. Each earlier control leaked price information in a different way, so none measured the harmonic edge cleanly.
Method
Control (h): 20 random entries per harmonic trade, same ticker, drawn uniformly from every bar in the same calendar year as the harmonic entry (intraday: same year within the 2023-11 to 2026-09 data). No exclusions. Open entry, entry bar counts, same stop % and 2R target, full exit (build.py sim_many). Sensitivity (h2): same calendar quarter. Ticker-clustered bootstrap CIs, both halves, 10,000-random-trader Monte Carlo. What earlier controls leaked: (a/b) whole decade = market mismatch (randoms from other years and markets; worst on 4h, where trades bunch in 2025-26); (c/d) +-60 days = centred on the signal, so the window holds the decline into D and the bounce just after it; (e/f/g) span-excluded = hindsight (a stretch counts as 'a pattern' only because a decline into D later followed, so excluding spans removes the worst random days).
Result
Edge (95% CI) / halves / MC share of random traders >= harmonic total. (h): 1d -0.010 (-0.067,+0.045), +0.07 / -0.07, 63%; 1h +0.058 (+0.011,+0.107), +0.07 / +0.04, 0.6%; 4h -0.044 (-0.140,+0.049), -0.23 / +0.04, 83%. (h2): 1d +0.011 (-0.038,+0.054), 32%; 1h +0.059 (+0.016,+0.102), 0.4%; 4h +0.040 (-0.039,+0.115), 16%. For comparison, (d): 1d +0.047 (3.2%), 1h +0.050 (1.5%), 4h +0.012 (39%).
How much to trust it
Medium: Simple, price-blind and reproducible; same-year randoms still share the stock's survivorship drift with the harmonic trades, and a year is a coarse match for the market state
Did it change the live rule?
no live change; (h) is now the reference random control for all entry studies, with (h2) and (d) reported as sensitivities
Notes
Supersedes (d) as the reference used in entry-vs-random-control-fix and entry-vs-random-pattern-span. Quarter vs year matters most on 4h (-0.04 vs +0.04), where the data is short and trades cluster.

Scripts and data for this study will be linked here when the research code is published on GitHub.

← All experiments

Education and research only. Not financial advice. Backtests use today's index members (survivorship bias).