Reproduce Strategy Statistics

[jorgebmann]

I run your script and were able to reproduce annualized returns (7.1% ), but the Sharpe Ratio was very low (0.3) with random forest 15-15 (EW) strategy. This is quite different to the Sharpe ratio of 0.88 reported in your paper. Where there any changes made?

[AndreasTheodoulou]

Hi Jorg,

Thanks a lot for reaching out. The script should be reproducing the output reported in the paper, it might be though that I have uploaded the wrong version of code by accident, as when I was working on this it wasn't through github and only uploaded after the end of the project. If you are keen to re-produce results I would try with the scripts attached below. Note that for at least one of the two you might have to change the model that is there from either XGB or NN to Random Forest and you might want to double check the data importing.

If I have time later, will try to figure myself why that is the case.

Thanks,
Andreas

MDLcoursework.txt
RF tuning2.txt

[AndreasTheodoulou]

Would be interested to hear back if you give it a shot.

[jorgebmann]

Hi Andreas, thank you for your fast reply. I will have a look at data importing and run these scripts and see if Sharpe ratio looks better. I also speeded up the random forest grid search a little with parallel computing, i.e.: library(parallel) G <- split(pars, 1:NROW(pars)) grd_temp <- mclapply(G, grid_par_hit, train_features = train_features, train_label = train_label, test_features = val_features, test_label = val_label, mc.cores = detectCores()) Results are identical and calculation is much faster. Cheers, Jörg Am Fr., 19. Juni 2020 um 16:04 Uhr schrieb Andreas Theodoulou < [email protected]>:

…

Hi Jorg, Thanks a lot for reaching out. The script should be reproducing the output reported in the paper, it might be though that I have uploaded the wrong version of code by accident, as when I was working on this it wasn't through github and only uploaded after the end of the project. If you are keen to re-produce results I would try with the scripts attached below. Note that for at least one of the two you might have to change the model that is there from either XGB or NN to Random Forest and you might want to double check the data importing. If I have time later, will try to figure myself why that is the case. Thanks, Andreas MDLcoursework.txt <https://github.com/AndreasTheodoulou/ML_Trading_Strategies/files/4804985/MDLcoursework.txt> RF tuning2.txt <https://github.com/AndreasTheodoulou/ML_Trading_Strategies/files/4804987/RF.tuning2.txt> — You are receiving this because you authored the thread. Reply to this email directly, view it on GitHub <#1 (comment)>, or unsubscribe <https://github.com/notifications/unsubscribe-auth/AJVVYJVVMCSMFGWZWQGRHP3RXNV7LANCNFSM4OCXHCTQ> .

[jorgebmann]

Ok, I run the scripts you provided. I change the model to Random Forest (copy and paste from first script), as you suggested. However, results are similar to the first one: Annualized return looks good (ca. 7%), but Sharpe Ratio is still at about 0.3. Data import seem to look good, thought I don't have a comparison. Did you have a chance to figure out what is the case?
Thanks,
Jörg