statcore

The statistical engine behind proj-mvp.vercel.app. Eleven methods, twenty modules, zero dependencies, runs in the browser. Every number on those sixty product pages is computed by this code at build time.

20 modules · 11 methods · 0 dependencies · 90 tests · MIT

Why you can trust the numbers

The changepoint core (cusum, bocd, bh) is a port of a tested Python implementation, and it is pinned to that implementation at 1e-6 by 25 parity tests running against fixtures generated from the original. If the port drifts, the suite goes red.

The eleven methods with no reference implementation to diff against are tested a different way — against properties the mathematics guarantees rather than against an oracle:

• conformal coverage holds at every alpha, on a deliberately skewed error distribution
• James–Stein shrinkage beats raw means on total squared error, on all 50 seeds
• the always-valid interval contains the truth along the whole monitoring path, not just at the end
• the GPD fit recovers a known shape parameter from exactly-sampled Pareto tails
• capture–recapture lands within 15% of a hidden population size

That distinction matters: a test that only checks the code agrees with itself proves nothing. These check it agrees with the theorem.

Install

npm install statcore

What is in it

| Module | Question it answers | Entry points | |---|---|---| | cusum, bocd, detect | Did this series change, and when? | cusum, bocd, runDetection | | bh | You tested a thousand things. Which findings survive? | applyBHCorrection, bhAdjust | | naive | What the threshold rule you use today would have said | naiveAlerts, compareToNaive | | power, bootstrap | Could you even have seen the effect? | requiredSampleSize, minimumDetectableEffect, pairedBootstrap | | sequential, ratesprt | Is it safe to look at this every morning? | mSPRT, rateSPRT | | conformal | How sure are you, with a guarantee? | fitConformalRegressor, calibrateAbstention | | causal | Did you cause it? | differenceInDifferences, syntheticControl | | survival | When did it happen, not how many | kaplanMeier, logRank, hazardByPeriod | | extremes | How bad can it get? | tailRisk | | hierarchical | Is that rank real? | shrinkRates, differsFromPopulation | | benford, capture | What is hidden, what did you miss? | benfordTest, captureRecapture | | judge | What did the LLM judge get wrong? | correctJudgedRate, compareJudged | | seasonal, stats | Weekly structure; the numeric utilities under everything | seededRandom, normPpf, quantile |

Three rules the code follows

1. CUSUM runs on day-over-day percentage changes, not levels. A growing series must not alert on its own growth.
2. Multiplicity is corrected across everything tested in a run. runDetection returns the suppressed findings as well as the survivors; the suppressed list is a feature.
3. Anything that peeks is always-valid. mSPRT and rateSPRT keep their error rate under continuous monitoring, and their intervals agree with their tests by construction.

Example

import { runDetection, mSPRT, calibrateAbstention } from 'statcore'

// Which of these series changed, after correcting for testing all of them?
const run = runDetection(series, { fdr: 0.05, minConfidence: 0.9 })
run.findings    // survived Benjamini-Hochberg
run.suppressed  // would have alerted without it

// Watch a canary after every batch without inflating the false-halt rate.
const seq = mSPRT(control, treatment, { alpha: 0.05, tau: 0.03, step: 250 })
seq.stoppedAt   // first n at which the always-valid test crossed, or null

// Turn any confidence score into a threshold with a guaranteed error rate.
const policy = calibrateAbstention(scores, correct, 0.02)
policy.threshold

Build

npm install
npm test     # 90 tests, including the 25 parity tests
npm run build

MIT.