Trend Decomposition

Split a series into the slow part, the repeating part, and what is left. The slow part is trend, the repeating part is seasonality, and the remainder is what a model actually has to work for. Doing this first is old, unglamorous, and still competitive — much of what a forecasting model learns is the part you could have subtracted.

A moving average estimates the trend, subtracting it leaves seasonality plus noise, and averaging over periods estimates the season. Modern architectures fold this in rather than replacing it: DLinear's whole contribution was decomposing and fitting each piece with one linear layer, which is why beating it is a real bar rather than a formality. Additive decomposition assumes the season does not scale with the level, which for anything growing is usually false.

y sub _t  = T sub _t  + S sub _t  + R sub _t  additively, or y sub _t  = T sub _t ·S sub _t ·R sub _t  multiplicatively, the second being the first applied to log y. The choice is a modelling assumption about whether seasonal swings grow with the level; taking logs first is the cheapest way to find out, since a multiplicative series becomes additive under it.