There is a popular trend of building machine learning teams in financial firms recently. The hope that the prosperity in image recognition community can be transplanted to financial industry by applying deep learning is too good to be true.

To understand why, we must first understand why deep learning has been so successful. In image recognition, the model was fixed as images. Different image have different information. By accumulating more and more image data, the information of the true model is accumulated and could be incorporated into the learning networks if the network is ‘large’ and ‘good’. ‘Large’ means the network need to have lots of nodes and layers which leads to the trend of making network deeper. ‘Good’ means the functions in the network need to be useful to characterize image features and also robust in training. This is why convolution function, residual function, batch normalization etc. are becoming mainstream.

However, in finance the type of true model is not known or constantly changing. Accumulating financial data would not help you find the type of model which generated the market prices. The signal to noise ratio approaches to 0. The seemingly rich financial data are essentially useless, or more accurately not useable since most of that is noise.

So what should we do? In fact, one model of financial market has started to gain much of attention. That is financial market is generated by human behavior under a set of legislation. We need to accumulate the data about the relation of human behavior and market price, and try to build ‘good’ and ‘large’ network to predict this type of relation.

Now, we still only have very limited understanding of human behavior. Shiller has some insights in his recent book on narrative economics. The power of this is already astonishing. Trump election and Brexit were results of using Facebook to influence human behavior, see Ted talk and report.

Reference

1 2