parsing out noise in predictions ?

lexx · Post by **lexx** » Sat Mar 07, 2015 3:17 am

I have the following case :

for N assets I make M forecasts using different methods.

What would be the most efficient way to combine those forecasts into 1?

I can normalize each of M forecasts so that they are comparable to each other and then simply average them.

But is there a better way (other than averaging) to extract common component out of them ?

Would much appreciate your opinions.

Maggette · Post by **Maggette** » Sat Mar 07, 2015 4:35 pm

Hi. Sure there are more sophisticated ways of combining forecasts. These depend on further information about the performance of your models/predictors and their relationship. Google "boosting".

First you could and should check how many of your M predictors actually add information.

I would start comparing the prediction errors of your predictors/models. Are they correlated? Are they from the same distribution (KS test)?

goldorak · Post by **goldorak** » Sat Mar 07, 2015 7:12 pm

An important point further down the road is to ask yourself if, once you have found a way to combine the forecasts, you can consider this as a definitive result and that it will not evolve with time.

svisstack · Post by **svisstack** » Sat Mar 07, 2015 10:59 pm

@Maggette: can you elaborate more about correlated signals? There are some techniques to liquidate repeted information to not bias towards correlation structure that will appear in most signals (for example when simplest method used - averaging).

When this was not clear enough then I will explain more:

imagine 3 signals with good error ratio where 2 signals have strong correlation, when averaging we should account this 2 signals approx as 1 not 2.

il_vitorio · Post by **il_vitorio** » Sun Mar 08, 2015 7:21 pm

Hello Lexx, let my add my nickle

I would suggest to do a minimax method, that is like taking the intercept in the confidence intervals of your forecast that should be the region of total confidence of your intervals. (knowing that your forecast do have a confidence intervals).

lexx · Post by **lexx** » Tue Mar 10, 2015 2:42 am

Thanks a lot for replies, guys! Quite helpful, will definitely try boosting.

@il_vitorio - do you have a reference for the method you mentioned ? - book, article ?

Anyway, thanks everyone!

il_vitorio · Post by **il_vitorio** » Tue Mar 10, 2015 3:26 pm

Yes, e-mail me and I will hand it to you!

Maggette · Post by **Maggette** » Tue Mar 10, 2015 3:57 pm

"Thanks a lot for replies, guys! Quite helpful, will definitely try boosting."

But you should realy consider the advice given by goldorak.