WideChaos / Forecast uncertaintySEPARATE SELECTION & CALIBRATION

NOMINAL TARGETS / OBSERVED COVERAGE

Inspect the interval.

Use a fixed historical point model with residual-rank intervals. Compare the nominal target with what later months actually covered.

先选模型,再用另一时段的误差定区间,最后检查未来时段的实际覆盖。这里的90%是名义目标,不能当作每条新输入或未来月份的保证。

The historical windows overlap and change with season. Exchangeability is not established here. Separate calibration labels keep them out of checkpoint selection; they do not establish a future-coverage guarantee.

90%Nominal target · α = 1/10
91.45%Observed test coverage · 3,894/4,258
172.98Mean interval width · rentals
364Primary misses · all retained

Monthly primary coverage ranged from 86.11% in October to 96.84% in December. The full test includes 4,258 eligible targets; 118 with incomplete history are omitted and listed.

Try the fixed study model

Enter exactly 24 consecutive nominal hours, oldest first. Counts are integer-valued numbers from 0 to 100,000. No timezones, gap filling or current-city validity is inferred.

Loading frozen point model and calibration scores…

Inspect the rank arithmetic

For n calibration scores and error fraction a/d, use the one-based rank ceil((n + 1) × (d − a) / d). If the rank exceeds n, the interval is unbounded. This calculator is separate from the fixed study model.

Open the score-rank calculator

Defaults: n = 4, α = 1/10, rank = 5, so the result will be unbounded.

For nonnegative outcomes. Up to 10,000 finite scores and point values from 0 to 1e9; score JSON up to 1 MiB. Scores are user supplied and unauthenticated. This arithmetic does not fit a model or establish exchangeability. The download includes the full score list.

Complete time split and observed coverage

All 40 training epochs, observed monthly coverage against nominal levels, mean widths, and complete four-role eligible and omitted target counts

Open full-resolution scientific figure

All three levels and all months are shown. These are descriptive results from one point fit and fixed earlier calibration scores; there is no adaptive threshold update or guaranteed monthly coverage.