Changelog

Source: NEWS.md

finnts 0.7.0.9009 (DEVELOPMENT VERSION)

Improvements

• Improved model selection balances accuracy and plausibility, preserves supported growth and seasonality, and rejects invalid predictions.
• Improved averaging, forecast updates, and run recovery.

Bug Fixes

• Forecast updates verify saved models and forecasts, refit damaged results, and recover interrupted logging without rewriting valid outputs.
• Shared global updates preserve valid local winners; preparation, artifact formats, and worker payloads are unchanged.
• Restart checks do not prevent concurrent attempts from overwriting files.
• Fixed future actuals becoming 0 with clean_outliers = TRUE; they remain NA, while backtests retain original actuals.
• Nonnegative hierarchical reconciliation caps inverse-error weight ratios at 1e15, reducing slowdowns from near-perfect fits without adding timeouts.
• Extreme-weight forecasts may change; valid weights and negative-allowed reconciliation remain unchanged. Invalid residual variances produce actionable errors.
• Local and ADLS-mounted workflows read known artifacts by exact path and reuse necessary listings, reducing repeated storage discovery.
• Condensed forecast getters preserve batch precedence; storage failures propagate, while optional missing files and valid empty CSVs remain supported.
• Legacy reads, remote downloads, and Spark data-frame routing remain unchanged.
• update_forecast() excludes removed series before update routing, avoiding missing-artifact errors; new series retain default local forecasts.
• Global Agent iterations compare only hierarchy approaches supported by their input scope and consolidated EDA.
• Pre-expanded hierarchy ID inputs use bottoms_up internally, followed by one outer reconciliation.
• Bottom-level inputs may compare bottoms_up against their detected standard or grouped hierarchy after bottom-level reconciliation.
• iterate_forecast() skips completed global optimization and resumes unfinished local series; incomplete metadata or higher iteration targets retain retries.
• Forecast updates choose the newest completed predecessor with nonempty required final outputs, skipping canceled or incomplete versions.
• Required outputs include run metadata, forecasts, model summaries, and EDA; hierarchical runs also require hierarchy summaries.
• Predecessor metadata validates reuse fields only; other final outputs are not checked for schema, combo, or model identity.
• Storage and read failures remain hard errors during predecessor selection.
• Agent duplicate detection treats NULL defaults, equivalent explicit values, and reordered multipart settings as identical, avoiding redundant iterations.
• Trimmed character combo boundaries before validation and hashing, aligning artifacts across workflows; blank values and normalization collisions fail early.
• Agent EDA summaries report absent outlier dates and omit non-finite regressor-lag correlations, keeping infinite values out of LLM context.
• Missing intermediate forecasts now explain incomplete or inconsistent runs and recommend restarting or regenerating the missing step.
• Validated custom seasonal_period values reach stlm-arima, stlm-ets, and tbats; default NULL remains logged as NA.
• Agent seasonal-period proposals require values above one; exhausted correction retries preserve and finalize the best forecast.
• Yearly defaults use two- and three-year periods; earlier-version replay presents default-backed settings as literal NULL.
• Invalid legacy seasonal periods warn once with their original value, then use cadence defaults during replay and forecast updates.
• Agent reasoning validates all proposed settings, using typed correction retries and graceful exhaustion.
• Retries reuse history; completed iterations refresh it once. Current-version runs control iteration limits, accuracy signals, duplicates, and change budgets.
• Earlier versions remain replay context; external-regressor exploration has no separate cap beyond the overall iteration limit.
• Canonical lag, rolling-window, and seasonal-period settings prevent reordered, default, or previously tested configurations consuming additional changes.
• Finalization and verification avoid repeated listings; best-run listing failures propagate as storage errors.

finnts 0.7.02026-08-27

Improvements

• New AI Agent Capabilities

  • iterate_forecast() can use LLM’s to find the optimal combination of data and inputs to create the most accurate forecast.
  • update_forecast() can take previously trained models from iterate_forecast() to create forecasts on new data fast.
    
  • ask_agent() can be used to ask questions about the forecast, data, or models to get insights.
  • Helper functions set_project_info() and set_agent_info to assist in iterating and updating forecasts.
  • New functions to retrieve information from agent runs:
    • Added get_agent_forecast() to retrieve the final forecast output from an agent run.
    • Added get_best_agent_run() to retrieve the run metadata information from an agent run.
    • Added get_summarized_models() to retrieve model summary information from an agent run.
    • Added get_eda_data() to retrieve the exploratory data analysis results from an agent run.

• New Chronos2 Model Integration

  • Added Chronos2, in addition to the existing model, to support zero-shot forecasting
  • It supports both historical and future external regressors
  • It can be used as a global model
  • Added a Chronos controller, which will support other Chronos variant API POST requests as well
  • Added two new package dependencies: jsonlite and httr
  • Integrated Chronos2 into the finn agent workflow
  • Added chronos-bolt-base: uses the Chronos2 API without external regressors, with model_type = "chronos-bolt-base".
  • Added lightweight chronos-bolt-tiny: shares the Chronos API without external regressors, with model_type = "chronos-bolt-tiny".

• New TimesFM Integration

  • Added TimesFM as a new foundation model for zero-shot time series forecasting
  • TimesFM is a local-only model (not global) that does not support external regressors
  • Uses its own API endpoint, configured via TIMESFM_API_URL and TIMESFM_API_TOKEN environment variables
  • Supports daily, weekly, monthly, quarterly, and yearly frequency data

• New TimeGPT Integration

  • Added TimeGPT in addition to existing model to support zero shot forecasting
  • Added support for both historical and future external regressors
  • Added timegpt-long-horizon model
  • Added finetuning for depth and layers
  • Enabled TimeGPT as a global model
  • Added support for padding time series that don’t meet minimum data requirements
  • Integrated TimeGPT into the finn agent workflow
  • Optional nixtlar enables TimeGPT on R 4.1+; core FinnTS remains available on R 4.0.

• Updated Train Model function

  • Added debug arg to trace error while training over various models
  • Fixed differencing restoration for combo data in global models.

• prep_data() removes outliers from training data while retaining them in time-series cross-validation testing splits.

• Adaptive daily ARIMA to reduce runtime

  • Users continue to select "arima"; daily workflows now use the bounded arima_fast engine while non-daily workflows retain classic auto_arima behavior.
  • Daily ARIMA validates nonseasonal, weekly-difference, 364/365-day-difference, and Fourier-with-ARIMA-errors strategies on an internal holdout, then refits the simplest competitive strategy.
  • Outer backtests retain original targets when outlier cleaning is enabled.
  • Daily candidate searches use nonseasonal frequency-one ARIMA fits and never construct the expensive period-365 seasonal state-space model.
  • Failed daily ARIMA candidates use another validated strategy or deterministic drift, without timeouts or process termination.
  • Agent summaries report ARIMA engine, strategy, transformed/effective orders, Fourier/seasonal settings, validation WMAPE, candidate scores, and fallback status.
  • Declared forecast in Imports, reusing the mature ARIMA implementation already required transitively by modeltime.
  • This adds no runtime service, credentials, or network surface and preserves the existing open-source dependency chain.

• Updated optional variable importance to vip 0.5.0 from its maintainer’s r-universe; vip remains in Suggests.

  • Declared ranger in Suggests; Boruta feature selection explicitly uses its ranger adapter after Boruta 10.0 changed defaults.
  • FinnTS works without feature-selection packages; requesting unavailable feature selection fails with installation guidance.
  • Without vip, model summaries retain everything except variable importance.

Bug Fixes

• Avoided MARS tuning failures by excluding CV pruning from automatic grids; explicit multistep CV pruning uses bounded folds.
• Fixed non-ASCII combo hashing across readers, preventing mismatched input, EDA, forecast paths and subscript out of bounds errors.
• Excluded models with incomplete backtest folds from best-model ranking while retaining complete candidates.
• Fixed null_converter() crash in agent workflow when input is NA.
• Added exponential-backoff retries, up to three, for transient Chronos and TimesFM failures: HTTP 429, 5xx, and connection errors.
• Fixed hierarchical forecast reconciliation failure caused by floating-point Target discrepancies across models.
• Improved error messages during hierarchical reconciliation to include the underlying error for easier debugging.
• Fixed aggregation error when running hierarchical forecasts with standard hierarchy approach.
• Fixed hierarchical issues when a combo variable contains a single unique value.
• Fixed issue when reconciling standard hierarchical forecasts.
• Fixed weighted mape calculation when target variable has negative values.
• Support for latest xgboost 3x version.
• Fixed model summary for global models by considering average models too.
• Fixed global-model failures when future external-regressor values exist for only some series.
• Fixed issue around NA handling with external regressors.
• Fixed issue when reconciling hierarchical forecasts that are very close to zero.
• Fixed issue when checking if best models have been selected before.
• Fixed multistep Cubist, GLMnet, MARS, polynomial SVM, and radial SVM failures caused by non-unique fiscal date-index joins expanding assessment rows.
• Multistep prediction preserves one row per assessment and rejects missing, duplicate, padded, truncated, recycled, or non-finite outputs.
• Removed XGBoost multistep prediction padding and truncation that previously masked row-alignment defects.
• Custom multistep lag_periods propagate through feature engineering, selection, training, and updates; uncovered horizons are appended as final lag boundaries.

Breaking Changes

• experiment_name within set_run_info() has been changed to project_name to comply with new AI agent capabilities.
• Migrated from qs to maintained, CRAN-ready qs2 for fast serialization and improved compression.
  • Existing .qs artifacts are incompatible with qs2 and must be regenerated.

finnts 0.6.02025-09-04

Improvements

• Shortened global model list to just xgboost
• Faster xgboost model training for larger datasets
• Faster feature selection for global model training
• Added seasonal_period within prep_models() for more control over multiple seasonal periods in models like tbats

Bug Fixes

• Error in formatting of training data for global models
• Error when using multiple external regressors with future values
• Remove target_log_transformation within prep_data(), since box_cox has now replaced it for automated power transformations
• Error when running hierarchical forecasts with weekly data

finnts 0.5.02024-10-25

Improvements

• Added support for hierarchical forecasting with external regressors
• Allow global models for hierarchical forecasts
• Multistep horizon forecasts for R1 recipe, listed as multistep_horizon within prep_data()
• Always save the most accurate model average, even when unselected, improving scalability for larger datasets.
• Automatically condense large forecasts (+3k time series) into smaller amount of files to make it easier to read forecast outputs
• Improved weighted MAPE calculation across all time series
• Changed default for box_cox argument in prep_data() to FALSE
• Support for spark version 3.4 in Azure Synapse/Fabric

Bug Fixes

• Error in run_type column join in final forecast output
• Error in running feature selection

Breaking Changes

• Minimum R version now set to R 4.0 to comply with package dependency minimum version for tune

finnts 0.4.02023-12-01

Improvements

• Tidymodels speed up
• Added arimax for ARIMA forecasting with engineered features and supplied external regressors.
• Automated feature selection, refer to feature selection vignette for more details
• Error handling in hierarchical forecast reconciliation
• Box-cox and differencing transformations
• Added new function, list_models(), that lists available models in the package

Bug Fixes

• Best model selection
• Hierarchical forecast reconciliation

finnts 0.3.02023-08-09

Improvements

• Added Spark data-frame input support for forecasting millions of time series across a cluster.
• Updated train/validation/test process for multivariate ML models.
• Added independently callable forecasting components for finer workflow control and production pipelines.
  • prep_data()
  • prep_models()
  • train_models()
  • ensemble_models()
  • final_models()
• Automated intermediate and final artifact reads and writes improve MLOps, scalability, and restart recovery.
  • Temporary location on local machine, which will then get deleted after R session is closed.
• Local or mounted Azure Data Lake Storage paths persist intermediate and final results in Spark.
• Azure Blob Storage supports non-Spark data-lake runs; SharePoint/OneDrive stores results within M365.
• Added get_trained_models(), get_run_info(), and get_prepped_data() to retrieve fitted models, run metadata, and feature-engineered data.

Deprecated

• run_model_parallel has been replaced with inner_parallel within forecast_time_series()
• Data being returned as a list when running forecast_time_series(). Instead please use get_forecast_data() to retrieve Finn forecast outputs.

Breaking Changes

• No longer support for Azure Batch parallel processing, please use spark instead
• Parallel Spark processing now requires an Azure Data Lake Storage mount supplied through set_run_info(); see the vignettes.

finnts 0.2.42023-04-14

Dependency Fixes

• Fixed dependency issue with timetk.

finnts 0.2.32023-03-23

Dependency Fixes

• Removed package dependency modeltime.gluonts and its deep learning models because the package is no longer on CRAN.

finnts 0.2.22023-02-11

Bug Fixes

• Fixed hierarchical forecast reconciliation issues for certain forecasts that have high residuals.
• Compliant with latest dplyr v1.1.0

finnts 0.2.12022-11-15

Bug Fixes

• Fixed feature engineering issue around NaN/Inf values when computing log values of negative external regressor values.
• Fixed issue of ensuring random seed is set correctly in parallel processing.

finnts 0.2.02022-07-14

Improvements

• Added spark support to run Finn in parallel on Azure Databricks or Azure Synapse.
• Added error handling for model averages so forecasting can continue despite memory failures with large model sets.
• Extended Azure Batch task timeout from one day to one week for long-running forecasts.

Deprecated

• Deprecated azure_batch in forecast_time_series() following Azure Batch R package deprecation; use Spark instead.

finnts 0.1.12022-03-28

Default Function Behavior

• forecast_time_series() defaults to R1 when run_global_models is TRUE or NULL and recipes_to_run is NULL.
• This avoids R2 memory issues with large global-model datasets in Azure Batch.

Bug Fixes

• Fixed error when converting infinite values to NA values after model forecasts are created.
• Changed the cubist model to reference the new cubist model definition in parsnip package.
• Fixed hierarchical aggregation by replacing missing hierarchy values with zero.

finnts 0.1.02022-02-07

• Initial CRAN Release