How Signal Confidence Is Published and Versioned
What the confidence figure represents, why methodology versions are attached to every signal, and how published performance should be interpreted.
Every signal carries two pieces of metadata that are easy to overlook and important to understand: a confidence figure and a methodology version.
What confidence is
Confidence is the platform's own stated strength for a specific signal under a specific methodology version. It is not a probability of profit, and it is not comparable across methodology versions.
Concretely, two signals at the same stated confidence under the same version are claimed to be equally well-supported by that methodology. That is a meaningful statement about consistency. It is not a promise about outcomes.
Why methodology versions are attached
Every signal is stamped with the methodology version that produced it. This is the mechanism that makes performance claims auditable:
- Results can be grouped by version, so a change in method does not silently inherit the record of the previous one.
- A methodology change is visible as a version boundary in the history rather than an unexplained shift in behaviour.
- Signals cannot be retroactively edited to look better — the version and the original text are part of the published record.
Reading published performance
The performance surface publishes signals issued, hit rate, average confidence, and pair-level breakdowns for a period. Three cautions:
- Sample size governs everything. A small number of signals produces a hit rate with a wide confidence interval; treat short-window figures as noisy.
- Hit rate without magnitude is incomplete. A 40% hit rate with large winners can outperform a 60% hit rate with small ones — read hit rate together with average result.
- Regime matters. A period dominated by one regime tells you about performance in that regime, not in general.
Why the website cannot compute any of this
By architecture, this website neither computes confidence nor recalculates performance. It requests published values, validates them against a versioned contract, and renders them — including the staleness of the data. If the platform publishes something that violates the contract, the website refuses to display it rather than guessing. That refusal is a feature: it prevents a broken upstream from becoming a plausible-looking chart.
Key takeaways
- Confidence expresses methodology-relative strength, not win probability.
- Version stamps make performance auditable and prevent silent method drift.
- Judge published statistics by sample size, magnitude, and regime.