Machine Learning Models Applied To High Frequency Financial Data

Chapter 1 is titled "A dynamic network model for high frequency order flows in financial markets." This chapter constructs a network model for high frequency trading volume data using a regularized vector autoregression moving average (VARMA) method. The network models how trading activity in one economic sector or asset group impacts trading in other asset groups. I explore the extent to which current trading volume is predictable from past trading volume history and how bursts of trading activity in one asset group is transmitted to other asset groups. I construct network connectedness measures which quantify the impact of these volume shocks. The results reveal that trading volume has a good deal of predictability: for the 51 asset groups considered, the model generates an average R2 of .16. About 45% of the network impacts as measured by impulse response functions are due to cross asset shocks. The results also reveal clear clustering of assets into groups that match economic intuition. Chapter 2 is titled "How useful are machine learning tools in predicting high frequency returns?" This chapter explores the predictability of high frequency returns for 63 exchange traded funds by applying recently popularized machine learning techniques. I consider factor models, LASSO, and random forests. From a baseline of predicting an asset’s return using its own return history, I build a sequence of predictive models of increasing complexity, and evaluate them using out of sample predictability. The results show that using the return history of all 63 assets improves out of sample predictability compared with the baseline model that uses only own return history. Incorporating market microstructure information in the form of volume, depth, and spread data into the model improves out of sample predictability further. However, introducing interactions, whether included in a linear model or created using a random forest, does not improve out of sample predictability. This predictability can be turned into substantial trading revenue, but transaction costs would mitigate these profits