TS models are constrained by the realities of the world they exist in. While you can chase benchmarks in lots of ML problems, forecasting is something that's us

TS models are constrained by the realities of the world they exist in. While you can chase benchmarks in lots of ML problems, forecasting is something that's used by basically every large business with huge consequences to getting it wrong or right. Therefore, people stick with relatively performant & interpretable methods such as ARIMA and friends. Additionally, most TS problems are relatively data constrained (your company/product has only existed for so long) so methods that are sample efficient (which most "modern" ML methods are not) are much more useful. Also, time series/forecasting is a ghetto ;)