pretty sure the results in the paper are because WT leaks future info

[JannyKul]

I tried implementing this too. 99% certain their implementation has the Wavelet transform leaking info. If you just do LSTM on daily close_price your odds are slightly better than a coin toss, adding WT they jump to around 65%. Running WT by windows so you don't leak info probability drops below 50%. SAE isn't going to recover anything from that.

I kinda elaborated a bit here

would love to hear your thoughts

@JannyKul
See the following thread/issue (relevant):
VivekPa/AIAlpha#13 (comment)

I am attempting to avoid this particular issue with the WT by applying it separately to a train-validate-test split as defined in the source article (Bao et al. 2017).

EDIT: I have read your medium article - I am skeptical of the results also, but that is also part of why I am attempting to replicate the results of the paper myself, first hand.

Example of scaling and applying the wavelet transform to the train-validate-test split independently:

https://github.com/timothyyu/wsae-lstm/issues/7#issuecomment-468628367

[mlpanda]

Hey guys, thanks for your comments. I totally agree that the paper most likely used the WT on the entire time-series. It should of course be done in a rolling fashion in which case even generating any profitable returns becomes very difficult. The approach might work on higher frequency data, but on those daily returns there is no chance that this approach works.

[JannyKul]

@mlpanda agree entirely