InsufficientData
Fewer points than the model needs to be identified.
@statili · regression & distributions for TypeScript
Two packages: one fits the model, one explains the fit.
@statili/stats
Eight regression and smoothing models, in modern type-safe TypeScript. Every fit reports its standard error, p‑value and confidence interval — or a typed reason it could not fit at all. 557 tests, validated against closed-form solutions and published statistical tables.
@statili/forge
A deterministic fact generator. It turns a result into structured
claims — a kind, a narrowed claim, a
level, and the provenance to audit it. Never prose, so one
fact becomes an aria-label, a full text alternative, or a
row feeding an enrichment pipeline.
Most regression libraries return a plausible answer even when the data can’t support one. statili refuses.
Every result is a discriminated union — a fitted model, or a typed reason it couldn’t fit one. Standard errors, p‑values and confidence intervals ship with every estimate, not as an afterthought.
import { linear } from '@statili/stats'
const result = linear({ precision: 2 }, [[1, 12], [2, 19], [3, 29], [4, 37]])
if (!result.ok) return console.error(result.errorType, result.message)
result.slope // 8.5
result.pValueM // 0.0021
result.slopeInterval // [6.83, 10.17]
result.interval(6) // { fit: 54, lower: 47.88, upper: 60.12 }
Results are marked, not just returned:
finding caveat failureIt refuses to guess
Three points that share an x-coordinate. There is no line to fit. Here is what each library does with that.
slope: 0, r²: 0 — no warning
@statili/[email protected]
errorType: DegenerateInput
“all x-values are identical — no line is defined”
import regression from 'regression'
regression.linear([[3,10],[3,15],[3,22]])
// { equation: [0, 15.67], r2: 0,
// points: [[3,15.67],[3,15.67],[3,15.67]] }
import { linear } from '@statili/stats'
linear({}, [[3,10],[3,15],[3,22]])
// { ok: false, errorType: 'DegenerateInput',
// message: 'all x-values are identical —
// no line is defined' }
Every failure is typed
A discriminated union of five error types, not a thrown exception and not a
silent NaN.
DegenerateInput
The data has no variation along an axis the model requires.
InvalidInput
Non-finite values, mismatched lengths, or out-of-range options.
MathError
A computation that is undefined for these inputs, such as a singular matrix.
NumericalStability
A result the solver does not trust to the requested precision.
It states its uncertainty
Not just a slope. Standard error, p‑value, degrees of freedom, and a
band you can draw — on the left from @statili/stats, and on
the right the same numbers after @statili/forge has turned
them into facts. The sentences are generated from the fit, so they cannot
drift from it.
slope 8.5se 0.62p 0.0021df 2
rendered by @statili/forge
aria-label (short)
Signups rising with week — slope 7.85, p < 0.001.
full text alternative
Each additional unit of week is associated with an increase of 7.85 users in signups. The slope of 7.85 is statistically significant (p = 8.7e-9, 6 degrees of freedom).
insight
Unlikely to be due to chance (p = 0.0021).
caveat
n = 4 — treat the interval as indicative, not final.
Six packages
Use the whole stack, or take one layer at a time.
@statili/fp
Composition and currying utilities that underpin the data-last API.
@statili/math
Descriptive statistics, numeric helpers, linear algebra and activation functions.
@statili/stats
Regression models returning discriminated-union results.
@statili/forge
Insight generation layered on top of the regression models.
@statili/highcharts
Fitted series, uncertainty bands and accessible descriptions for Highcharts.
@statili/compat-regression
Drop-in replacement for the regression package, with its silent-failure bugs fixed.
557
tests, including validation against closed-form solutions and published statistical tables
2017
regression has not changed since
2020
highcharts-regression has not changed since, and has no test suite