The undoings of deep learning in Finance
The promise around artificial intelligence centres around advances in the field of deep learning. Deep learning is a subset of machine learning (derived from statistical learning) which seeks to learn data representations instead of performing specific tasks. Unsurprisingly, this ability of deep learning algorithms to detect patterns has exciting possibilities for the field of finance where the primary objective of fund managers is to “beat the market”.
by Mrinal Mishra.
Traditionally, this is achieved by using past information (price information like stock prices is used for most forecasts apart from balance sheet information; of late, myriad variables like twitter reactions and satellite images are also being used) to forecast future prices using a variety of models while minimizing the error between the realized and forecasted prices. Deep learning models can minimize these errors as they are able to understand past patterns more precisely. This enables these models to achieve much superior forecasts with greater precision.
We proceed to illustrate a simple use of smart-indexing as demonstrated byHeaton, Polson and Witte (2016). Traditionally, index creation is done by selecting a small group of stocks which historically have given a performance very similar to that of the reference index. This process of selecting the stocks is mostly a trial and error method which involves the use of linear regression. These selected stocks (via the trial and error) are expected to be reasonable linear approximation of the underlying index. However, the deep learning model enables us to identify even non-linear relationships. This ability to deconstruct non-linear relationships makes in-sample approximation trivial apart from allowing us to train the model in a robust fashion to obtain out-of-sample accuracy in forecasts. Heaton et al. use autoencoders (as a technique) in their example in lieu of standard linear regression.
Given sufficient size and diversity in the input data, a deep learning model can often be trained to approximate the target data to almost complete accuracy.Classic models tend to focus on obtaining in-sample approximation quality. While obtaining accuracy in-sample is pertinent for causality (and academicians have tended to dwell on it considerably), it has not found flavor with practitioners. Due to its focus on out-of-sample performance as the optimization target, deep learning has been able to garner more interest in industry circles. However, the very strength of deep learning (which allows it to obtain impeccable non-linear relationships in data) can also be its undoing.The biggest criticism of deep learning models is “overfitting” i.e the tendency to have more than required parameters in the model. Overfitted models are able to mimic the training data exactly but perform poorly when exposed to unseen data. Hence, it is imperative that practitioners remain cognizant of these shortcomings before widely adopting deep learning techniques.