Timeline for Time Series forecasting feature creation/engineering
Current License: CC BY-SA 4.0
5 events
| when toggle format | what | by | license | comment | |
|---|---|---|---|---|---|
| Apr 12 at 12:45 | comment | added | Comte | Detrending, helps trees to extrapolate. And if one really wants they can use linear trees (results may vary). | |
| Apr 12 at 11:51 | comment | added | Comte | @noNameTed GBDTs and random forest are excellent algorithms for modelling time series data. They do well on competitions, and in production for time series forecasting, especially when you have a lot of exogenous variables. They don't need to be in time, you just need to make time into features. | |
| Feb 22 at 15:05 | comment | added | noNameTed | I see your point. However, I'm still not aware of any tree based models (except for model trees) that are capable of extrapolating outside the range of response variables seen in the training data. A traditional forecasting problem like "given a history of car sales by month, predict car sales for each of the next 12 months" would usually desire that behavior, especially when there is a positive trend in sales. ARIMA and exponential smoothing models can both handle that case, but I'd expect the tree models to predict the maximum value seen in the training set for the next 12 months. | |
| Feb 21 at 21:32 | comment | added | m13op22 | Tree-based models can be used for forecasting if you reformat the data by windowing, just as you would for neural networks, and can be very powerful for forecasting. | |
| Feb 21 at 16:31 | history | answered | noNameTed | CC BY-SA 4.0 |