LoginSign Up
← Back to all sparks
F

FoRecoML

INFRA · APIS
Velocity0.0

Forecast Reconciliation with Machine Learning

github.com/danigiro/forecoml ↗

The machine-learning arm of a forecast reconciliation toolkit, four months old and already sharing its sibling's plumbing.

forecastingmachine-learninghierarchical-reconciliationtime-seriesr-package
◆Current state
FoRecoML brings machine-learning approaches to forecast reconciliation across cross-sectional, temporal, and cross-temporal frameworks through csrml(), terml(), and ctrml(). It reached CRAN in April 2026 and has since spent both releases integrating with FoReco rather than expanding its own method set: results are now FoReco's foreco objects, and print() and summary() report framework, approach, problem dimensions, features, training sample size, combination matrix, and trained models.
◆Where it's heading
This package is being built as a satellite, not a competitor. Adopting FoReco's exported new_foreco_class() constructor within days of that class appearing means FoRecoML results drop straight into the same print, summary, plot, and components methods as analytically reconciled ones — which is what makes machine-learning and classical reconciliation directly comparable in a single workflow. The 1.1.1 argument-validation work landed in the same minute as the equivalent change in FoReco, so the two are being maintained as one release train.
◆Prediction
With the integration work done, the next release is more likely to add or expose machine-learning approaches than to keep reshaping output; the structured summary already enumerates features and trained models, which suggests inspection tooling is where attention has been.

◆Recent moves

  1. 1mo ago

    Structured print and summary for fitted reconciliation models

    print() and summary() on rml_fit objects now lay out the framework, approach, dimensions, features, training sample size, combination matrix, and trained models instead of a bare dump. The stricter argument validation shipped here mirrors the same change in FoReco, released a minute apart — these two packages move together.

    View source ↗
  2. 1mo ago

    Adopts FoReco's foreco class for all reconciliation output

    The adopting half of a two-package move: FoReco defined and exported the foreco class, and three days later every function here returns one, built through the same new_foreco_class() constructor. The directional decision was made upstream; what happens here is the payoff, with FoReco's print, summary, plot, and components methods now working on machine-learning results.

    View source ↗
  3. 3mo ago

    Machine-learning forecast reconciliation arrives on CRAN

    ⚡ SPARK

    The package's first release, opening a machine-learning route to a problem that has been solved analytically. The notes are two lines, but the scope they state is the whole point: all three reconciliation frameworks, learned rather than derived.

    View source ↗