Can deep learning be used to quantify transactions?

Can deep learning be used to quantify transactions?

• Yes, but don't play predictions (except high-frequency trading)

  I've seen a lot of articles, publications, or brokers write about deep learning as input based on historical indicators, using networks like LSTM to predict future stock and futures returns and match them into trading strategies. This method I've basically tried, either through the classification method or the regression method to predict, and the results are bad.

  This is not to say that new technology is unreliable in predicting the price of assets such as stocks, but first let's think about why you can predict the future with just a few inputs. This hypothesis of predicting the future based on historical data is strong, and under a strong hypothesis, using a black box to run out of a narrow winning rate is somewhat unimpressive.

  The key is that there is a stable data-dimensional correspondence between the image and the name, which is more complex, but the relationship is stable. And the financial sequence is different, and the logic of historical data predicting the future is itself unstable, which will only make the results more confusing with such a complex tool. But in fact, deep learning has a particularly suitable application in secondary market quantitative trading, specifically what I am not convenient to say, the characteristic of this application is definitely stable correspondence.

Translated from Quantified Transactions by Zenino