foxalgoMethodology
How to read every number on this site.
Descriptive statistics, not signals — touching a level is not earning money on it. Every number on this page ships with its sample size (n) and confidence interval. See methodology.
rate / base / delta
rate is the measured probability of an outcome under a stated condition (e.g. "price is above yesterday's high at 09:30"). base is the unconditional probability of the same outcome, same market, no condition. delta is rate − base: how much the condition actually moves the odds. A rate of 75% next to a base of 74% is not an edge — check delta, not rate alone.
n and confidence interval
n is the sample size the rate is measured on. Every rate ships with a 95% Wilson score confidence interval — a range that would contain the true probability 95% of the time under repeated sampling. Small n means a wide interval; treat those numbers as provisional even if the point estimate looks striking.
fit vs. held-out
Numbers are computed on two windows: a fit window (where the pattern was found) and a held-out window (data the discovery process never saw). This site publishes held-out numbers wherever available — they run cooler than fit numbers, and that's the honest number. A pattern that only exists in fit is called overfitting.
touching is not earning
A touch probability describes whether price crossed a level at some point in the window — not whether a trade placed there would have made money. Price can tag a target an hour after a stop-loss would have already been hit. These are descriptive statistics about price paths, not trade signals.
the skeptic process
Findings on the odds page carry a badge describing how they were checked:
| badge | meaning |
|---|---|
| SKEPTIC-CONFIRMED | Passed a dedicated adversarial review round beyond the standard gates. |
| WIDE (Bonferroni across sectors) | Survives a wide multi-sector search with Bonferroni-corrected significance across sectors. |
| CROSS-CLUSTER (correlated markets) | Consistent within a market cluster (peer-sector pattern), not yet independently re-verified. |
Ideas that fail this process end up on the graveyard page, with the specific reason they were killed — commonly lookahead bias (the signal secretly used future information) or a stratification confound (an apparent edge that disappears once you control for a market's own position).
crypto: a different clock
Crypto (BTC, ETH, SOL, AVAX, LINK, DOGE) is measured from Binance spot prices, 2019 onward, anchored to 00:00 UTC instead of a US market open — there's no exchange bell to anchor to since it trades 24/7. Numbers before 2023 were used to fit the curve; 2023 onward is the held-out, verified window. Weekends are quieter — the odds change, so we show separate weekend numbers.
eras and traps we watch for
- Lookahead — using information not yet knowable at decision time. The most common way a backtest lies. A naive placebo control does not reliably catch it.
- Compositional artifacts — combining two conditions and comparing to each alone without matching sample weights produces a fake "combo effect".
- Selection bias — measuring a rate only on a filtered subset (e.g. "when the signal fired") without stating what fraction of days it fires on.