Deep Learning Techniques for predicting High-Frequency returns using Order Book data

A financial market is an ensemble of market agents willing to buy or sell a certain financial security, such as a stock, bond, or derivative. The natural infrastructure for a financial market is an exchange, a physical or virtual place that brings together buyers and sellers, facilitating the occurrence of transactions. Today most exchanges are electronic, allowing traders to access live order book information, i.e., the collection of all standing orders for a given security.

Market participants have different technological infrastructures, receiving data and submitting orders at different latencies. Over the past few decades, HFTs have engaged in a fierce race to zero latency, making vast economic efforts to reduce their latency by just a few microseconds. What we aim to explore in our research is one of the possible reasons why such a race happened in the first place.

Specifically, we aim to analyse the predictive value of order book data, i.e., to what extent can a trader with immediate access to the order book predict the market's future direction?

Empirical studies have shown that price formation dynamics, i.e., next mid-price moves, are predictable. In our research, we investigate whether such predictability persists at longer horizons. To do so, we employ deep learning architectures, leveraging their ability to learn complex data dependencies. So far, we have conducted an extensive empirical experiment on over one year of Nasdaq data to answer the following questions:

1. Do high-frequency returns display predictability? If so, how far ahead can we predict? 2. Which order book representations perform the best? 3. Can we use a single model across multiple horizons? 4. Can we use a single model across multiple stocks?

To answer these questions, we used model confidence sets, a structured statistical procedure particularly well suited for the problem.

There are some further questions we wish to explore. First, we would like to understand whether the structure of the order book can completely explain the predictability in returns or if recurring trading patterns play a relevant role. Moreover, we would like to understand whether such predictability is tradeable or if it might be useful for some market players, for example, helping market makers gauge the market's direction and adjust their quotes accordingly.

Graph supOU processes

While there are various discrete-time models for graph/network time series, our research project will focus on the continuous-time setting to allow for consistent modelling across time scales and account for irregular observations.

The project aims to extend previous work on Graph-Ornstein Uhlenbeck (GrOU) processes. There are various research avenues we wish to consider. First, we would like to allow for a more flexible autocorrelation structure, possibly displaying long memory.

In this context, we could consider merging and advancing the existing theory of GrOU processes with that of multivariate CARMA processes. Alternatively, we could consider defining a Graph supOU process and developing suitable inference techniques. Second, we wish to explore graphs with more complex topologies.

For example, we may want to allow multivariate observations on each node and/or consider networks with natural group structures.

In applications, it is often the case that the dimension of the graph's adjacency matrix is (much) larger than the number of time series observations. In this setting, an important question is whether we can estimate the adjacency matrix consistently. If so, one might want to detect sparsity in such networks. Possible real-world data sets with time-evolving graph structures stocks' realized volatilities and exchange rate pairs.

This project falls within the following EPSRC research areas: Artificial Intelligence Technologies, Digital Signal Processing, and Statistics and Applied Probability.