Skip to content

Time Series

methods

https://news.ycombinator.com/item?id=37877443

  • On extremely high dimensional data: MLP that just uses lagged values as features

  • On mid-dimensional data: LightGBM/Xgboost is by far the best and generally performs at or better than any deep learning model

  • On low-dimensional data: (V)ARIMA/ETS/Factor models are still king

model history

https://medium.com/@ycwong.joe/a-brief-history-of-time-series-models-38455c2cd78d

summary

https://chartexpo.com/blog/time-series-forecasting

How to feed forecast into next step's input feature

https://www.kaggle.com/code/ryanholbrook/forecasting-with-machine-learning

  • multioutput model

  • each model for one step

  • one model, update feature based on prevous forecast

  • each model for one step, update feature based on prevous forecast