Skip to contents

Changelog

Source: NEWS.md

chronofeat 0.6.0

Major Changes

  • pkgdown Documentation Site: Complete documentation overhaul with 6 comprehensive articles
    • Getting Started guide
    • Building Custom Models (primary focus)
    • Feature Engineering Reference
    • Data Preprocessing
    • Cross-Validation
    • Advanced Workflows (including ParBayesianOptimization)
  • Complete Forecast Engine Refactoring: Eliminated ~155 lines of duplication through systematic helper extraction and unified forecast loop
    • .build_future_grid(): Unified future date generation for grouped/ungrouped data
    • .prepare_feature_row(): Centralized target/calendar/xreg feature assembly
    • .apply_schema(): Centralized schema harmonization and predictor selection
    • Unified forecast loop: Replaced separate grouped/ungrouped branches with single loop treating ungrouped as pseudo-group

Bug Fixes

  • Window Validation: C++ forecasting path now returns NA for incomplete rolling/trend windows instead of silently shortening them, aligning with R-side behavior and preventing train/test feature distribution mismatch
  • Cross-Validation: Added panel alignment validation that detects and reports misaligned date ranges across groups
  • Grouped Features: Fixed grouped rolling/MA/lag features that were bleeding across group boundaries by ensuring operations respect group_by()
  • Trend Index: Corrected trend feature indexing during recursive forecasting - now uses length(y) + 1 for next step instead of length(y)
  • Factor Coercion: Now errors when attempting rolling statistics on factor columns instead of silently converting to meaningless numeric codes

chronofeat 0.5.0

Initial release with formula-based time series forecasting engine.

Features

  • Formula interface for feature specification: value ~ p(12) + q(7) + month() + dow()
  • Automatic feature engineering:
    • Target lags: p(k) creates k lags
    • Moving averages: q(w1, w2, ...) creates MAs with specified windows
    • Calendar features: dow(), month(), woy(), eom(), dom()
    • Rolling statistics: rollsum(), rollsd(), rollmin(), rollmax(), rollslope()
    • Trend features: trend(1, 2, ...) for polynomial trends
    • Exogenous variable lags/MAs: lag(var, k), ma(var, w)
  • Model-agnostic interface supporting any R model with fit/predict
  • C++ accelerated recursive forecasting via Rcpp
  • Multi-group/panel data support
  • Time series cross-validation with cv_forecast()
  • TimeSeries object with frequency detection and preprocessing