0133 — Lodestar.Stats.TimeSeries is a package, and the serial-correlation lot moves into it

Status: accepted · Date: 2026-09-15 · Supersedes: 0114

Context

Decision 0114 measured #617's lot — the autocorrelation functions and the Ljung-Box test — against decision 0076's three criteria, found none distinct, and kept it in Lodestar.Stats. It also wrote down what would change that, and where it would be tested:

#671 re-measures rather than inherits. Three readings would have to change for a split to be earned, and #671 is where each is testable: a dependency this lot never needed — ADF's OLS is precisely that — a caller who wants the stationarity tests and no hypothesis test at all, or a release Lodestar.Stats is not on.

#671 is that lot: the augmented Dickey-Fuller test, KPSS and seasonal decomposition, at statsmodels 0.15.0 parity. Its spec (docs/superpowers/specs/2026-09-15_0671_stationarity-and-seasonal-decomposition.md) takes the placement question first, because the answer decides where every other line goes.

The measurement

0114's basis, stated before the numbers: lines per git diff --stat against origin/main; source-declared public members, accessors excluded, constructors counted only when explicitly written and public; enum fields counted apart.

The new lot: 1,040 lines of C# in seventeen files, thirteen public types and twenty-six public members, plus sixteen enum fields.

type	kind	public members
Stationarity	static class	2 — AugmentedDickeyFuller, Kpss
SeasonalDecomposition	static class	1 — Decompose
DickeyFullerOptions, KpssOptions, SeasonalDecompositionOptions	sealed records	3 each
DickeyFullerResult	sealed class	6
KpssResult	sealed class	5
SeasonalComponents	sealed class	3
TrendTerms, LagSelection, KpssLagRule, SeasonalModel, PValueBound	enums	16 fields

grep -c 'public ' src/Lodestar.Stats.TimeSeries/*.cs over the thirteen files sums to 39 lines; less the thirteen type declarations that is the twenty-six, and the arithmetic and the command agree.

With #617's five types the subject is eighteen types and forty-two members in 1,510 lines — a lot three times 0114's, which is the case the #617 spec's "What would make this spec wrong" named: "If the second lot turns out to be three times this one, the measurement taken here answers a question about a third of the eventual surface."

The dependency the lot needs, measured rather than assumed. A throwaway replay of 56 statsmodels cases through the published Lodestar.Stats.Regression 0.1.0 agreed to 5.5e-12 worst, and the frozen corpora now agree on 199 stationarity and 32 decomposition cases.

Against 0076's three criteria

Dependency profile — distinct. adfuller fits an ordinary least squares per candidate lag and reads each fit's t statistics and residual sum of squares. Lodestar.Stats.Regression's Householder least squares is that computation — as OrdinaryLeastSquares.Fit with the whole table, and as OrdinaryLeastSquares.Estimate without it, which is what a lag search reads. It restores Lodestar.Decomposition and, through it, Lodestar.Abstractions behind it. Lodestar.Stats depends on nothing, and Lodestar.Stats.Regression already depends on it, so the edge cannot run from Lodestar.Stats at all. Put in 0076's own terms — does shipping these types together force something on a caller who does not want it? — a Kruskal-Wallis caller would restore a QR, a Householder factorisation and a sparse-matrix package for a unit-root test they never call.

The two alternatives that keep the lot in Lodestar.Stats were weighed in the spec and refused there. A private least-squares solve would be a second OLS in one repository, which decision 0095 refused for the tails on the ground that two copies of one computation disagree eventually. An edge from Lodestar.Stats to Lodestar.Decomposition is the imposition above, one step removed.

Audience — still not distinct, and this record does not claim it is. 0114's reading holds: a reader checking residuals for normality is one call from checking them for independence. Neither does the audience argue against the split: the tests move, the reader's using does not.

Cadence — none yet. Nothing in the lot has been released, and Lodestar.Stats/v0.4.0, the last tag, carries none of #617's types.

One criterion is enough, and the dependency profile is the one 0076 itself was written about.

Decision

Lodestar.Stats.TimeSeries ships as its own core package, with two edges — Lodestar.Stats 0.4.0 and Lodestar.Stats.Regression 0.2.0 — and #617's five types move into it. The eighteenth package.

The whole namespace moves, not the half that needs the edge. 0114's third option, splitting the subject across two packages that share Lodestar.Stats.TimeSeries, stays refused for 0114's reason: a reader installing the package named after the subject would get half of it. The namespace was chosen in the #617 spec for exactly this contingency, so no type changes name and no caller's using changes — only the PackageReference.

The move is owed now, and costs nothing now. No Lodestar.Stats release carries these types. The next Lodestar.Stats tag would have been the point after which moving them needs a deprecation cycle; this record lands before it.

Options considered

• Keep everything in Lodestar.Stats with a private least-squares solve. Refused above on 0095's ground. The regression is not a line of arithmetic: OrdinaryLeastSquares.Fit solves through a Householder QR precisely because the normal equations square the condition number, and a private copy would either repeat that work or quietly skip it.
• Put the stationarity tests in Lodestar.Stats.Regression. Refused by 0096's sentence, which 0114 already quoted: "anything with a link function or a time index falls on that issue's side of the line." It would also make a correlogram cost a QR.
• A package for the new lot only, leaving SerialCorrelation in Lodestar.Stats. Refused with 0114's third option, above.

Consequences

• Lodestar.Stats loses SerialCorrelation, AutocorrelationOptions, AutocorrelationResult, LjungBoxOptions and LjungBoxResult before any release carried them; its CHANGELOG.md entry moves with them. docs/reference/stats/timeseries/serialcorrelation.md stays as a pointer, because decision 0122 links to it and a decision record is not edited.
• The fixed cost #566 measured is paid once: Version.props, the EXPECTED edges and a FLOORS dependent, the *.NetStandard.Tests mirror pinning four assemblies, every pack loop and release allow-list, wiki-map.json, the samples and the reference pages. dotnet test Lodestar.slnx reports thirty-six assemblies.
• A Lodestar.Stats.TimeSeries release follows Lodestar.Stats.Regression 0.2.0's, the floor its packed nuspec declares; Lodestar.Stats 0.4.0 is already on nuget.org.
• If a later lot needs a newer OrdinaryLeastSquares — an HAC covariance for Newey-West, say — the floor rises and Lodestar.Stats.Regression publishes first, the order CONTRIBUTING.md already gives for two packages edited together.