Probably a bit late to the discussion, but I disagree

The return is a function of the price. if we can correctly predict the price we can predict the return, that is not the problem here.

In investing we use returns not price because of non stationarity and auto-correlation of prices. like time differencing, using returns (integer differentiation) is a transformation that is “intended” to make the data series normal iid as this makes it easier for us to make inferences, traditionally with frequentist statistics. Neural Networks like RNN/LSTM/GRU do not explicitly require the condition of stationarity in fact it doesn’t require you to make any assumptions about the underlying distribution. In theory you do not need to make any of these transformations if you have enough data and the right architecture but in practice it is still very hard to predict non stationary time series. (Even when transformed, they are still often exhibit unfavourable characteristics like skew and kurtosis )

Researcher | Investor | Data Scientist | Curious Observer. Thoughts and insights from the confluence of investing and machine learning.

Researcher | Investor | Data Scientist | Curious Observer. Thoughts and insights from the confluence of investing and machine learning.