Time Series Forecasting

Predict what a measured quantity does next — demand, load, price, temperature. What makes it its own problem rather than regression with a clock is that the training and test sets are separated by time rather than by sampling, so every mistake about what stays the same is a mistake the model cannot see until it is deployed.

Evaluation is where most of the errors live. Shuffling the split leaks the future into training, a model tuned on one period silently absorbs its regime, and the naive baseline — predict the last value — beats a surprising fraction of published deep models. The 2022 result that a single linear layer matched several transformer architectures on the standard benchmarks was less about transformers than about how weak those benchmarks were.

Fit ŷ sub _t+1:t+h  = f(y sub _1:t ) with the split ordered rather than random, so validation must roll forward rather than sample. Error grows with the horizon h for any model, because uncertainty compounds — which is why a comparison at one horizon says little about another, and why reporting a single number for a multi-step forecast hides where it failed.