The unreasonable difficulty of time series forecasting

Opening excerpt

I’ve been thinking recently about what makes time series forecasting problems so difficult compared to other sequence learning tasks or IID Machine Learning problems.

First some motivating baselines illustrating the difficulty of the forecasting problem. I ran a series of models

These are interactive plots, click legend to hide a line, double click to isolate. Note that the historical sections have been truncated to 3x the forecasting window, for better visualisation.

There are many lines to look at, but the benchmark study’s main takeaway is that for series with strong, stable seasonality such as m4 hourly, the simple statistical baselines (Seasonal-Naive, MSTL) and the zero-shot foundation models win by large margins. However, the sophisticated models tend to be completely off for many time series tasks and are not much better than naive predictors because their predictions may trend in the completely wrong direction.

(Excerpted from the original; full article via the source link below.)