How to solve Time Series problems with Machine Learning

by

There are different techniques or models to solve these type of problems. We can divide the type of models using this diagram:

Type of Time Series Models
Type of Time Series Models

Traditional Time Series forecasting techniques

  • Recursive: we can make predictions for tomorrow, the day after… to do the prediction of day #4 we will predict first #1, then #2, then #3 and then we will forecast #4.
  • Easy to expand: if we want predictions for the next 10 days, just keep recursively making predictions.
  • Thought to get right.
  • Cannot add time varying features.

Examples of techniques:

  • ARIMA
  • SARIMA (seasonal ARIMA)
  • Prophet
  • Neural Prophet
  • Vector Auto-regression (VAR)

Machine Learning models

  • Direct: make predictions directly independently of the horizon.
  • Thought to extend: training data increases linearly as we have more horizons to predict.
  • Easy to get right.
  • Can add time varying variables as features.

Example of techniques:

  • Neural network regressor
  • Catboost regressor
  • Any regressor

Leave a Comment