BNNmetrics

Bayesian Neural Networks for Econometrics

Abstract

The complexity and volume of financial data in modern financial markets have been exponentially growing during the last decades. Machine learning (ML) methods such as Deep learning (DL) have been widely utilized for several classification and prediction problems, given their intrinsic flexibility, appropriateness for large multidimensional problems, and ability to discover and adapt to non-linear patterns. However, the enormous number of parameters, their difficult interpretation and inability do deal with uncertainties represent DL’s main shortcomings. On the other hand, classic econometrics methods, of limited variables, great interpretability and with excellent probabilistic properties, have failed to prove appropriate for the analysis of modern high-frequency data. The application in financial econometrics of a DL sub-class of algorithms known as Bayesian neural networks (BNNs) is expected to revolutionize the process of modeling, analyzing, and understanding trading behavior in real markets. BNNs’ attractive properties have the potential of bridging the gap between classic econometrics and ML. This research will show measurable improvements over the current state of the art, both from the financial econometrics and the ML sides, in three problems defined on high-frequency financial data: volatility modeling, stock mid-price movement prediction, and interdependence analysis between stock prices.

Objectives

BNNmetrics is focusing on three research objectives of maximum relevance for researchers and practitioners. The planned research will rely on ultra-high-frequency, full-depth limit-order book market data, to develop and extend new and current time-series econometric models by using the powerful BNN framework. My research objectives are described in detail in the following.

Modeling volatility based on non-linear data-driven models

Complex non-linear patterns in time-series and generic non-normal distributions will beexplred within the bayesian neural network framework. Backtesting methods based on posterior and predictive distirubionts from BNNs' outputs will allow for precise assestment and validation of the results, unveiling the extent to which this methodology outperforms w.r.t. purely-econometrics and purely data-driven ML approaches.

Stock mid-price movement prediction

BNN-based strategies for variable selection along with the underlying econometric rationale will allow scalable and flexible models, but parsimonious and with increased interpretability, showing ML and econometrics complement each other in a multidisciplinary framework. I will address the role different sets of variables play on the forecasts’ uncertainty, and how to balance between model complexitty and forecasting performance.

Inter-stock dependency analysis

The adoption of data-driven methodologies with the expressive power of ANNs has an enormous potential towards a better understanding of inter-stock trading dependencies. My approach exploiting BNNs which show advantages compared to ANNs w.r.t. their ability to handle uncertainties and interpretability properties will allow me to better understand the dynamics between interacting stocks.

Implementation

The project is hosted by Aarhus University (Denmark), Department of Engineering, Section of Electrical and Computer Engineering.

BNNmetrics is supervised by the Machine Learning expert Assoc. Professor Alexandros Iosifidis, leader of the Machine Learning and Computational Intelligence (MaLeCI) group.

Project Partnership include The Quantitative Analytics Group of the Financial Markets Division of ING Bank, ING Group the Netherlands.

Publications

Submitted and under review

Magris, Martin, Mostafa Shabani, and Alexandros Iosifidis. "Bayesian Bilinear Neural Network for Predicting the Mid-price Dynamics in Limit-Order Book Markets." arXiv preprint arXiv:2203.03613 (2022).

Magris, Martin, Mostafa Shabani, and Alexandros Iosifidis. "Quasi Black-Box Variational Inference with Natural Gradients for Bayesian Learning." arXiv preprint arXiv:2205.11568 (2022).

Magris, Martin, Mostafa Shabani, and Alexandros Iosifidis. "Exact Gaussian Manifold Variational Bayes." arXiv preprint arXiv:??? (2022).

Mostafa, Shabani, Martin Magris, George Tzagkarakis, Juho Kanniainen, and Alexandros Iosifidis. "Predicting the State of Synchronization of Financial Time Series using Cross Recurrence Plot". arXiv preprint arXiv:??? (2022).

For more information see this page.

Find us at conferences and other venues

Statistics Seminars, University of Jyvaskyla, 8 Jan. 2021, presenting "Modelling and Forecasting High-frequency Volatility with Vine Copula " link

EcoSta 2021, Hong Kong, 24-26 June 2021,presenting "A vine-copula HAR forecasting model: Application to Dow Jones stocks" link

IAAE 2021, Rotterdam, The Netherlands, 22 - 25 June 2021 (Virtual), presenting "Approximate Bayes factors for unit root testing" link

BAYSM 2021, Virtual, June 10, poster "Approximate Bayes factors for unit root testing" link

EcoSta 2022, Kyoto, Japan, 4-6 June 2022, presenting "Bayesian Bilinear Neural Network for Predicting the Mid-price Dynamics in Limit-Order Book Markets" link

AU Digital Innovation Conference 2022, Aahus, Denmark, May 2, presenting "Bayesian inference for deep learning: gaining predictive edge on the market"

Eusipco 2022, Belgrade, Serbia, 29 Aug.-2 Sept., presenting Shabani, Mostafa, et al. "Multi-head Temporal Attention-Augmented Bilinear Network for Financial time series prediction." link

About the project

Horizon 2020 Call: H2020-MSCA-IF-2019 (Marie Skłodowska-Curie Individual Fellowships)

Topic: MSCA-IF-2019

Type of action: MSCA-IF-EF-ST (Standard European Fellowships)

Granted by: Research Executive Agency

Grant agreement number: 890690

Acronym: BNNmetrics

Funding period: October 1st, 2020 - September 31st, 2022 (24 months)

Page updated

Google Sites

Report abuse