Fit intelligence

What the live fit-graph has learned for a product — region-level fit probability, the pooling path, audience split, and size recommendation with a confidence interval.

There's no backend endpoint yet for a per-product posterior (style selection, pooling, size recommendation) — in this version the fit-graph shows only aggregate metrics (FRR, signal count, body-area hotspots, tier quality) with live data.

Pooling path

The pooling path (global → brand → category → fit-block → style → size) requires a per-product posterior; this endpoint isn't live yet. No data.

Body-area hotspots

Misfit weight per body area across all returns — where fit issues concentrate the most.

No body-area hotspot data yet — it will populate once enough fit signal accumulates.

High weight = the body area where returns show the most fit problems.

Fit signal

Fit-driven returns / total lines

0fit return

fit share

No return lines yet — fit share has no data.

Fit return rate (FRR)

Store-wide · current

—

There's no time-series endpoint yet for the weekly FRR trend (sparkline) — only the current rate is shown.

Audience comparison · TR vs EU

no data

Same product, two body distributions

Ideal-fit probability by body area, side by side — posteriors diverge by audience, and the recommendation engine scales each market separately.

There's no backend endpoint yet for the audience (TR / EU) split — the fit-graph doesn't currently offer an audience-split posterior. No data.

Active viewaudience split · no data

Size recommendation strip

no data

Per-size recommendation strength + a 90% confidence interval require a product-size-level posterior; this endpoint isn't live yet. No data.

Cold-start queue

New products · signal collection

Styles that haven't matured yet — the estimate comes from a higher pool. The threshold is 40 signals; significant deviation starts at 5 returns.

The cold-start queue (new products' signal-collection rate + maturity estimate) needs a per-product endpoint; not live yet. No data.

Signal quality feeding the graph

Returns entering the pool per tier — only FIT-driven signals train the graph.

Total tier signal—

Signal quality

—

T2+/T2/T1 share · high confidence

Recommendation accuracy

—

all audiences · live

Size exchange success

—

exchange instead of refund

Flywheel: every T2/T2+ return sharpens this style's posterior; as accuracy rises, fit returns drop and signal quality improves.

How to read this

quick guide

Pooling path

A low-signal product gets its estimate from a higher level (category/fit-block). As signal accumulates, the posterior descends to style and size — prior influence decreases.

Body-area fit probability

Each bar shows the probability that the size fits that body area ideally. A low value + a direction label (tight/loose/short/long) marks where the engine scales the size.

TR vs EU

The same product's posterior can diverge across two audiences. The selector above chooses which audience is shown; the comparison card places both side by side.

Confidence interval

The thin strip on the size card is the 90% confidence interval. Narrow = the engine is confident, wide = little signal. The interval narrows as signal grows.

Cold-start queue

New products stay queued until they reach 40 signals. Significant deviation starts at 5 returns; the maturity column gives the estimated time at the current rate.

Signal quality

Only FIT-driven returns train the graph. T2+/T2 are the most valuable signal; EXCLUDE (non-fit) never enters. Tier weight summarizes learning power.