User Guides

Every feature of skforecast explained with runnable examples, grouped by topic.

• Data

  The format skforecast expects for series and exogenous variables, and the datasets used in the examples.

  • Input data
  • Datasets

• Choosing a model

  Machine learning, statistical, foundation and deep learning models compared with the same backtesting.

  • Machine learning vs statistical vs foundation models: which one to use

• Single series Forecasters

  One model for one series: recursive and direct strategies, baselines and classification.

  • Recursive multi-step forecasting
  • Direct multi-step forecasting
  • Forecasting baseline
  • Autoregressive classification

• Global Forecasters (multiple series)

  One model trained on many series at once, either independent or multivariate.

  • Global forecasting: local vs global models and which forecaster to use
  • Independent multi-time series forecasting
  • Series with different lengths and different exogenous variables
  • Dependent multivariate series forecasting

• Foundation Models

  Pretrained models that forecast zero-shot, without training on your data.

  • Forecasting with foundation models
  • Heterogeneous series with different lengths, exogenous variables and missing values

• Statistical Models

  ARIMA, ETS and ARAR with the same API as the rest of the forecasters.

  • Forecasting with statistical models
  • ARIMA, SARIMAX, AutoARIMA
  • ETS, AutoETS
  • ARAR

• Deep Learning

  Recurrent neural networks (RNN, LSTM) for multi-step forecasting.

  • Deep learning Recurrent Neural Networks

• Feature Engineering

  Build the predictors of the model: lags, window features, exogenous variables and transformations.

  • Autocorrelation analysis and lag selection
  • Window and custom features
  • Exogenous variables
  • Calendar features
  • Categorical features
  • Data transformation
  • Differentiation
  • Handling missing values
  • Sktime pipelines
  • Feature selection

• Model Evaluation and Tuning

  Measure the error as the model will be used in production and search for the best configuration.

  • Metrics
  • Backtesting forecaster
  • Hyperparameter tuning and lags selection

• Probabilistic Forecasting

  Prediction intervals and quantiles, and how to evaluate them.

  • Probabilistic forecasting
  • Bootstrapped residuals
  • Conformal predictions
  • Quantile forecasting
  • Probabilistic global models
  • Conformal calibration
  • Metrics in probabilistic forecasting
  • Continuous Ranked Probability Score (CRPS)

• Model Explainability

  Understand what drives each forecast and inspect the matrices the model is trained on.

  • Model Explainability
  • Extract training and prediction matrices

• Model Deployment

  Save, load, run and monitor a forecaster in production.

  • Save and load forecaster
  • Forecaster in production
  • Drift Detection
  • Skforecast in GPU

• Plotting

  Built-in plots for residuals, prediction intervals and autocorrelation.

  • Plotting

• Additional Resources

  Migration from earlier versions and specific techniques: weights, stacking and gradient boosting.

  • Skforecast Migration guide
  • Weighted time series forecasting
  • Stacking multiple models
  • Forecasting with XGBoost and LightGBM

• Migrating from other libraries

  How the concepts of other forecasting libraries translate to skforecast, with side-by-side backtests.

  • Migrating from Prophet to skforecast

We hope you find these guides helpful. If you have any questions or need further assistance, please don't hesitate to reach out to the skforecast community. Remember to visit the FAQ and forecasting tips page for answers to frequently asked questions and forecasting tips.