FAQ and forecasting tips

Answers to frequent questions and practical tips to get the most out of skforecast, grouped by topic. If your question is not answered here, open an issue on GitHub so that other users can also benefit from the answer.

• Forecasting Tips

  Practical answers to common modelling problems, such as negative predictions, missing values, trend or delayed data.

  • Avoid negative predictions when forecasting
  • Forecasting time series with missing values
  • Forecasting with delayed historical data
  • Forecasting time series with trend: differentiation
  • Forecasting time series with trend: tree based models

• Model Evaluation

  Evaluate forecasts and prediction intervals without overly optimistic results.

  • Backtesting vs One-step-ahead
  • Continuous Ranked Probability Score (CRPS)
  • Calibration of probabilistic forecasting intervals
  • Avoid data leakage in pre-trained forecasting models

• Feature Engineering

  Encode cyclical calendar features and change the frequency of a series.

  • Cyclical features in time series
  • Time series aggregation

• Performance Optimization

  Measure and speed up training, backtesting and hyperparameter search.

  • Benchmarking skforecast
  • Parallelization in skforecast
  • Profiling skforecast