cryptocurrency-price-prediction
cryptocurrency-price-prediction copied to clipboard
abhinavsagar
The block comparing prediction to test data is wrong
The following line is plain wrong:
preds = test[target_col].values[:-window_len] * (preds + 1)
You're basically taking the result you want to see (test[target_col].values[:-window_len]) and multiply it by (1 + noise) where te amplitude of the noise was in my case always smaller than 0.3.
The other plot is given as:
targets = test[target_col][window_len:]
which is the same as preds, but shifted by five. (1 + preds) is near to 1, so basically the stuff you see is the stuff you wanted to see in the first place.
The correct way is to:
preds = test[target_col].values[0] * (preds + 1)
which does show no correlation at all - the predicted value diverges from the test data.
I changed this value and now it looks not that accurate. @tastypenguinbacon is this the only modification needed? The results look way more reasonable that way.
This code was featured on some major crypto news outlet clickbate article which explains why the "bug" was never fixed. In this case the bug is intentional to generate internet "clout".
