Performance Intelligence

Analytics use one canonical performance instead of counting every source row. When two sources describe the same exact performance, StrideSignal keeps both source assertions but counts the performance once. An ambiguous possible duplicate stays in the model until the user excludes it.

Exact dates can enter chronological charts, elapsed-time calculations, test/race pairing, and short-term models. Month- or season-only marks enter the robust all-history projection anchor and corresponding summaries, but StrideSignal never turns them into an invented July 1 observation.

The legal PB series applies the stored wind and timing rules. Wind-assisted or missing-condition marks remain visible as all-conditions results. A legal PB step requires an exact date; coarse-date bests stay in the season table.

This compares up to five recent exact-date meet summaries with the median of every earlier stored meet summary. It reports a deterministic 80% meet-bootstrap range when eligible. It is a competition-performance comparison, not a readiness, fatigue, recovery, health, or injury-risk score.

Consistency uses every stored exact-date mark across all available years. It calculates robust variation on log marks from the median absolute deviation, a deterministic 80% meet-cluster bootstrap range, and the classical sample coefficient of variation in technical details. Fewer than five results produce no repeatability estimate. Tight grouping means repeatability, not improvement.

StrideSignal counts canonical meets, competition days, and starts. This is recorded competition exposure. It is not training volume, total load, physiological stress, fatigue, or readiness. Training sessions remain separate.

Training tests are grouped only when test type, unit, timing, protocol version, device, lead-in, rest, start, venue, and surface match. Test/race relationships use exact dates inside 21 days. There is no same-season fallback. Fewer than 10 pairs produce no numerical coefficient; 10–19 pairs are exploratory; interpretive wording begins at 20 pairs. Association does not prove cause.

Long-term straight-line forecasts, next-PR marks, and PR probabilities are withdrawn. The 30- and 90-day estimate uses every positive non-excluded stored mark. Exact-date marks enter rolling-origin selection among robust all-history, recency-weighted, capped trend, and ensemble candidates. Coarse-date marks enter a capped robust anchor without receiving an invented day. Sparse and one-result estimates are explicitly preliminary. Every uncalibrated range is called model spread.

The old 0–100 reliability score is replaced by four dimensions: data integrity, comparability, sample adequacy, and model evidence. The overall tier follows the limiting dimension, and every page shows an analyzed-data receipt with included and excluded counts.

Review the analyzed data, source, conditions, timing, wind, classification segment, and limiting factor. Discuss important decisions with a qualified coach. StrideSignal does not diagnose health, readiness, fatigue, injury, or classification.

Calculation method last updated: 2026-08-01.