LoginSign Up
← Back to all sparks
F

finnts

INFRA · APIS
Velocity0.0

Microsoft Finance Time Series Forecasting Framework

github.com/microsoft/finnts ↗

Microsoft's automated forecasting framework, still mostly a one-maintainer effort.

forecastingtime-seriestidymodelsautomlhierarchical-modelsr-package
◆Current state
finnts automates time-series forecasting end to end — feature engineering, model selection, hierarchical reconciliation — on a tidymodels backbone. Recent releases have concentrated on global models (one model fitted across many series) and on hierarchical reconciliation, which has needed repeated correction at weekly granularity. Release notes are auto-generated pull-request lists, so the detail lives in the PRs rather than in the changelog.
◆Where it's heading
Cadence is roughly annual and the commit history is almost entirely one maintainer, with occasional outside contributions. The direction across the last four releases is consolidation of the forecasting internals — multi-horizon models, feature selection, reconciliation fixes — rather than new surface for users. The changelogs themselves are unedited PR dumps, which makes the arc harder to read than the work probably warrants.
◆Prediction
Hierarchical reconciliation has produced a bug fix in three of the last four releases, so the next one likely touches it again; nothing in the entries points to a specific new capability.

◆Recent moves

  1. 11mo ago

    Global model updates and a weekly reconciliation fix

    Four merged pull requests: a global-models update, general model fixes, and a bug fix in weekly hierarchical reconciliation. The release notes are a bare PR list, so the user-visible scope has to be inferred from titles — the reconciliation fix is the one with obvious downstream effect on forecasts.

    View source ↗
  2. 1y ago

    Multi-horizon models, hierarchy drivers, Synapse 3.4 migration

    The largest release in this window: multi-horizon forecast models plus the lag and global-model fixes they required, external drivers for hierarchical forecasts, best-model scaling, and a migration to Synapse 3.4. Multi-horizon work dominates, though the PR-list format leaves the modelling change undescribed.

    View source ↗
  3. 2y ago

    ARIMAX support and feature selection added

    Adds ARIMAX from an outside contributor — the package's first external code contribution — plus feature selection and a tidymodels process update. Another model for an existing task rather than a new class of forecasting, but the reconciliation guardrails added here (capping oversized forecasts, fixing residuals) address a recurring failure mode.

    View source ↗
  4. 3y ago

    dplyr 1.1.0 compatibility and a reconciliation fix

    Upstream compatibility work for dplyr 1.1.0, a typo correction, and another hierarchical reconciliation bug fix ahead of a CRAN submission. Maintenance rather than movement.

    View source ↗