The Problem With Deep Learning and Financial Applications

We are still witnessing relatively small number of success stories regarding the application of Deep Learning to common ( and novel ) tasks in financial industry. There are many success stories in general Deep Learning applications and then there are reports on how hedge funds are struggling to achieve meaningful results applying same techniques. Deep Learning also did not make any significant inroads into more mainstream financial corporations.

Why is this the case ?

Most recent advances were tied to applying Deep Learning techniques ( multi layer neural networks ) to niche applications — images and text recognition. Main characteristics of these data sets are that they are massive, static and unstructured.

Financial data, on the other hand, is dynamic, structured and not as massive. Simply applying ( impressive ) techniques we learnt with images and text just won’t do in financial world. Brute force massive computing power applied to very precise financial data won’t cut it. We have to come up with a more sophisticated approach- specialized algorithms, or look at algos that can work off of smaller amounts of data ( one shot learning ) to make advances.

The additional problem is the fact that often price change driver can be a piece of political news, which is something that even the smartest algorithm can predict. Human behavior is not something that can be modeled or predicted.