Forecasting with Machine Learning

Learn how to forecast time series with machine learning models: from turning time series into tabular data, to single-step and multistep forecasting, multiseries forecasting, backtesting, and error metrics.

This repository contains the practical notebooks for the Forecasting with Machine Learning course. The examples use pandas 3, scikit-learn, Feature-engine, and skforecast.

Released: March 2024 Last updated: Nov 2026 Status: Actively maintained

What you will learn

• Frame time series forecasting as a regression problem.
• Tabularize time series with lag, window, and datetime features.
• Build benchmark models to know whether machine learning is adding value.
• Forecast multiple steps ahead with recursive and direct strategies.
• Forecast many time series at once with global models.
• Evaluate forecasting models with backtesting and choose the right error metric.

Course contents

1. Forecasting as regression

   • Time series forecasting
   • Tabularizing time series: lag, window, and datetime features
   • Endogenous and exogenous variables
   • Feature engineering with Feature-engine
   • Benchmark models: naive, seasonal naive, moving average
   • Single-step forecasting with machine learning

2. Multistep forecasting

   • Recursive forecasting
   • Recursive forecasting with window features and future-known features
   • Direct forecasting with skforecast and scikit-learn
   • Direct forecasting with future-known features

3. Multiseries forecasting

   • Global vs local forecasting
   • Independent multi-time series forecasting
   • Dependent multi-time series forecasting

4. Backtesting

   • Backtesting with or without refitting
   • Backtesting with expanding or rolling training windows
   • Backtesting with intermittent refitting
   • Backtesting with a gap

5. Error metrics

   • Choosing error metrics for forecasting
   • Implementing and using custom error metrics

Getting started

Clone the repository (Python 3.12+ required), then set up a dedicated environment to run the notebooks.

Create and activate a virtual environment

python -m venv fwml
source fwml/bin/activate        # macOS/Linux
fwml\Scripts\activate           # Windows

Install dependencies

pip install -r requirements.txt

Install Jupyter and register the kernel

pip install jupyter ipykernel
python -m ipykernel install --user --name=fwml --display-name "fwml"

The environment will now be available as a kernel named fwml in Jupyter Notebook. Start Jupyter and open the notebooks in numerical order:

jupyter notebook

Course

For lectures, explanations, and the complete learning path, visit the online course.