A1002
Title: Bayesian bilinear neural network for predicting the mid-price dynamics in limit-order book markets
Authors: Martin Magris - Aarhus University (Denmark) [presenting]
Alexandros Iosifidis - Aarhus University (Denmark)
Mostafa Shabani - Aarhus University (Denmark)
Abstract: Recent advances in Variational Bayes (VB) for deep learning led to fast and efficient methods for scalable Bayesian inference applicable to complex modelling and predictive tasks, yet their use in econometric modelling is not widespread. In modern financial markets, traditional time-series forecasting models are often incapable of capturing the underlying complexity and interacting nature of the patterns driving price dynamics. On the other hand, data-driven Machine Learning (ML) methods have been proved effective. However traditional ML approaches are incapable of addressing parameters uncertainty and confidence interval for the predictions. Bayesian ML methods provide a natural remedy to bring together the predictive ability of ML methods and the probabilistic dimension typical of econometric research. Within the VB framework, we train a Bayesian bilinear network with temporal attention by adopting a state-of-the-art Bayesian optimizer to address the challenging prediction of mid-price movements in high-frequency limit-order book markets. Our results underline the feasibility of Bayesian methods in complex econometric modelling and their predictive and interpretative gains over several ML alternatives. We address the use of the Bayesian framework to analyse errors and uncertainties associated with forecasts and deliver insights for actionable trading decisions.
