Apologies if something similar has been asked before here, I can't exactly find the right phrase to describe my question.

Suppose I have a full 100 years of daily temperature values (with Feb. 29s removed).

100 * 365 = 36,500 data points: call this time series Y

Let's say I also decide to generate 365 more time series (using the same dataset) one for each day of the year, and each consisting of 100 data points (i.e. one for each year). Lets call these series y1, ..., y365, where each has length 100.

Suppose, using Y, I build a seasonal SARIMA/GARCH model to generate short-term rolling forecasts, i.e. 1-10 days ahead.

My question is this:

Is there an existing technique/is it possible to improve the forecasts of the seasonal SARIMA/GARCH model trained using Y by fitting less complex SARIMA/GARCH models on each the relevant y series for the forecast period and combining the forecasts?


1 Answer 1


You do not bring new information, only treat it differently, so the final forecast will not be better.

If you think about some thing in line of AdaBoost, you will train 365 weak learners (the temperature of a given day) on 100 points, then find the better weighted average for the 365 weak learners. You will get something that is not very far of the average over 100 years for the 365 weak learners and for the weight the daily variation.

AdaBoost, and other boosting methods, are efficient because there is no obvious global method.


Your Answer

By clicking “Post Your Answer”, you agree to our terms of service and acknowledge you have read our privacy policy.

Not the answer you're looking for? Browse other questions tagged or ask your own question.