Member-only story
How causal impact models improve time series forecasting accuracy
Knowing why something changed is even better than knowing that it changed.
Time series forecasting has traditionally focused on identifying patterns and extrapolating them into the future. Techniques like ARIMA, Prophet, and LSTM-based neural networks excel at modeling seasonality, trends, and autocorrelation. However, these methods generally assume that future behavior is a function of past observations alone. They don’t answer a critical question: what caused the change?
This is where causal inference comes in. It shifts the conversation from correlation to causation. Instead of asking, “What will happen next?” causal models ask, “What caused this to happen?” This distinction becomes crucial when external interventions — like marketing campaigns, product launches, or policy changes — might alter the trajectory of a time series.
Causal impact models, especially those built on Bayesian structural time series (BSTS), provide a statistical framework to measure the effect of an intervention on a time series. These models not only enhance interpretability but also improve forecasting by adjusting for real-world events that affect trends. In this article, we’ll explore how causal impact analysis works, how to implement it in Python, and…
