Artificial Intelligence in Finance

Part III. Statistical Inefficiencies

“There are patterns in the market,” Simons told a colleague. “I know we can find them.”¹

Gregory Zuckerman (2019)

The major goal of this part is to apply neural networks and reinforcement learning to discover statistical inefficiencies in financial markets (data). A statistical inefficiency, for the purposes of this book, is found when a predictor (a model or algorithm in general or a neural network in particular) predicts markets significantly better than a random predictor assigning equal probability to upwards and downwards movements. In an algorithmic trading context, to have such a predictor available is a prerequisite for the generation of alpha or above-market returns.

This part consists of three chapters that provide more background, details, and examples related to dense neural networks (DNNs), recurrent neural networks (RNNs), and reinforcement learning (RL):

• Chapter 7 covers DNNs in some more detail and applies them to the problem of predicting the direction of financial market movements. Historical data is used to generate lagged features data and to generate binary labels data. Such data sets are then used to train DNNs via supervised learning. The focus lies on identifying statistical inefficiencies in financial markets. In some of the examples, the DNN achieves an out-of-sample prediction accuracy of more than 60%.

• Chapter 8 is about RNNs, which are designed to accommodate the specific nature of sequential data, such as ...