Time Series: Why Baselines Win

The M-Competitions consistently show that simple statistical methods and tree-based machine learning outperform complex deep learning architectures for univariate forecasting.

Home / Guides / Time Series Forecasting: Why Baselines Win

The Danger of Overfitting

Time series data is notoriously noisy. Deep learning models, given their massive parameter counts, have a strong tendency to memorize noise rather than learning true underlying signals, leading to catastrophic generalization failure.

Required Baselines

Never deploy a neural network without first tracking metrics against:

  • Naive Drift: Predicting the last observed value.
  • Seasonal Naive: Predicting the value from the previous season (e.g., last Monday).
  • ARIMA / ETS: Classical statistical methods.

Feature Engineering over Architecture

Instead of building a complex LSTM, extract lag features, rolling means, and datetime components (day of week, month) and feed them into LightGBM. This approach is faster to train, easier to interpret, and usually more accurate.

Internal Resources