finnts 0.7.0

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 foundation model. Uses the same Chronos API as chronos2 but does not support external regressors. Passes model_type = "chronos-bolt-base" to the API.
  • Added chronos-bolt-tiny foundation model. Lightweight Chronos model variant using the same API as chronos2 and chronos-bolt-base. Does not support external regressors. Passes model_type = "chronos-bolt-tiny" to the API.

• 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
  • Declared nixtlar as an optional dependency. TimeGPT requires R 4.1 or newer, while core FinnTS workflows remain 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.

• Updated how outliers are handled in prep_data(). Outliers are removed from the training data, but still kept in the testing splits during time series cross validation.

• 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.
  • The daily engine validates nonseasonal, weekly-difference, 364/365-day-difference, and Fourier-with-ARIMA-errors strategies on an internal holdout, then refits the simplest competitive strategy. Finn’s outer back-tests continue to score 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.
  • Candidate failures fall through to another validation-successful strategy or a deterministic drift fallback. No model timeout or process-termination behavior was added.
  • Agent model summaries identify the actual engine, selected strategy, transformed and effective ARIMA orders, Fourier/seasonal settings, validation WMAPE, candidate scores, and fallback status.
  • Added forecast as a direct Imports dependency. It was already transitively required by modeltime; using its mature ARIMA implementation avoids reimplementing numerical estimation. This adds no new runtime service, credential, or network surface and retains the package’s existing open-source dependency chain.

• Updated optional variable-importance support for vip 0.5.0. vip remains in Suggests and is resolved from its maintainer’s r-universe repository. ranger is now declared directly in Suggests, and Boruta feature selection uses Boruta’s ranger adapter to preserve behavior after Boruta 10.0 changed its default importance provider. FinnTS continues to install and run without these optional packages: feature selection now fails early with installation guidance, while model summaries retain all sections except variable importance.

Bug Fixes

• Prevented random MARS tuning failures when prune_method = "cv" was selected without the required folds. Automatic grids now use the five non-CV pruning methods, while explicit multistep CV pruning supplies a bounded fold count.
• Fixed subscript out of bounds failures for time series combos containing non-ASCII characters. File name hashes are now stable regardless of how the text was read in (e.g. read.csv vs vroom), so input data, EDA, and forecast outputs resolve to the same file.
• Fixed partial-fold models incorrectly winning Best_Model selection. Models that fail on some back-test folds are now excluded from best-model ranking while other complete models continue normally.
• Fixed null_converter() crash in agent workflow when input is NA.
• Added retry with exponential backoff (up to 3 retries) for Chronos and TimesFM API calls on transient failures (HTTP 429, 5xx, 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 issue when future values of external regressors exist in some series but not all, leading to missing data issues when training a global model.
• 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 now preserves one prediction per original assessment row and fails explicitly on missing, duplicated, padded, truncated, recycled, or non-finite output.
• Removed XGBoost multistep prediction padding and truncation that previously masked row-alignment defects.
• Custom multistep lag_periods now propagate consistently through feature engineering, feature selection, model training, and forecast updates. Lag lists that do not cover the forecast horizon automatically include the horizon as a final boundary.

Breaking Changes

• experiment_name within set_run_info() has been changed to project_name to comply with new AI agent capabilities.
• Migrated from the qs package to qs2 for fast object serialization. The qs2 package is actively maintained and CRAN-ready with improved compression. Files previously saved with qs format cannot be read by qs2; any cached .qs files from prior runs will need to be regenerated.

finnts 0.6.0

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 0.5.0

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, regardless if selected as best model. This allows for improved scaling with large data sets.
• 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 0.4.0

Improvements

• Tidymodels speed up
• Added external regressor support for ARIMA by introducing a new model option of arimax, which uses engineered features in addition to any external regressors supplied.
• 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 0.3.0

Improvements

• Spark data frame support. Initial input data can now be a spark data frame, enabling millions of time series to be ran across a spark compute cluster.
• Updated train/validation/test process for multivariate ML models.
• In addition to existing forecast_time_series(), added new sub components of the finnts forecast process that can be called separately or in a production pipeline. Allows for more control of the forecast process
  • prep_data()
  • prep_models()
  • train_models()
  • ensemble_models()
  • final_models()
• Automated read and write capabilities. Intermediate and final Finn outputs are now automatically written to disk (see options below). This creates better MLOps capabilities, easier scale on spark, and better fault tolerance by not needing to start the whole forecast process over from scratch if an error occurred.
  • Temporary location on local machine, which will then get deleted after R session is closed.
  • Path on local machine or a mounted Azure Data Lake Storage path in spark to save the intermediate and final Finn run results.
  • Azure Blob Storage to store non-spark runs on a data lake. SharePoint/OneDrive storage to store non-spark runs within M365.
• New MLOps features that allow you to retrieve the final trained models through get_trained_models(), get specific run information thorough get_run_info(), and even retrieve the initial feature engineered data through get_prepped_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 processing through spark now needs a mounted Azure Data Lake Storage path supplied through set_run_info(). Please refer to the vignettes for more details.

finnts 0.2.4

Dependency Fixes

• Fixed dependency issue with timetk.

finnts 0.2.3

Dependency Fixes

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

finnts 0.2.2

Bug Fixes

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

finnts 0.2.1

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 0.2.0

Improvements

• Added spark support to run Finn in parallel on Azure Databricks or Azure Synapse.
• Added error handling when creating simple model averages. Should allow forecast to keep running even if there are memory issues when averaging individual forecast models, which helps on large data sets.
• Expand Azure Batch task timeout from one day to one week. Prevents errors when running large forecasts that take over a day to run in Azure Batch.

Deprecated

• Deprecated azure_batch parallel compute option within forecast_time_series function since the Azure Batch R packages are deprecated. Please use the new integration with spark on Azure.

finnts 0.1.1

Default Function Behavior

• Change default behavior to only run R1 feature engineering recipe when the argument run_global_models is set to TRUE or NULL and recipes_to_run is set to NULL in the forecast_time_series function. Running R2 recipe with global models on large data sets often results in RAM issues when running 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 bug in hierarchical forecasting. Missing values in the hierarchy are converted from NA to zero, which fixes how data is aggregated at various levels of hierarchy.

finnts 0.1.0

• Initial CRAN Release