How many dates will there be in the OOS? It will be important to know when it comes to resources allocation and retraining frequency as we might hit the limit by the end of OOS period.
There is 36 date in the out of sample.
From 300 to 335.
That’s it? it is very little… this will be prone to so much luck imo. Why does it say that the out of sample is running until mid November? Will it be live data?
There may be some luck of course since the score can be volatile.
You can work on that by reducing the volatility of your spearman score to better control your risk.
Regarding the OOS phase. Yes it will be running until mid november. I did a detailed answer I think you saw already: Current score is from how many time periods - #2 by xgilbert
It will not be live data to answer your last question.
There are some trade off to do when splitting the data. BIgger OOS would mean less training points which is not ideal either. If you have any suggestion(s), we would be glad to here them of course !
Thanks for your reply!
It’s true that a trade-off is inevitable if the data is limited. I participated in the ubiquant market prediction on Kaggle, and I think the OOS there was fairly large, and indeed there was pretty big swings in the ranking at every updates.
So a large OOS allowed for rewarding the hopefully intrinsically better models.
On the other hand, if you are working with a limited/fixed data size… I guess the current split makes sense to me