multi-horizon-forecasting-for-limit-order-books's Introduction

Multi-Horizon-Forecasting-for-Limit-Order-Books

This jupyter notebook is used to demonstrate our work, Multi-Horizon Forecasting for Limit Order Books: Novel Deep Learning Approaches and Hardware Acceleration using Intelligent Processing Units. We use FI-2010 dataset and present how model architecture is constructed here. The FI-2010 is publicly available and interested readers can check out their paper. The paper is available at https://arxiv.org/abs/2105.10430.

A blog and a video for our work are available at https://www.graphcore.ai/posts/graphcore-turbocharges-multi-horizon-financial-forecasting-for-oxford-man-institute.

multi-horizon-forecasting-for-limit-order-books's Issues

Original LSE data

Great paper! Any chance that the original LSE LOB data is available for open access?

additional features

I wonder how to incorporate additional features like volume, open interest to the model.

couple questions about the accuracy

Hello,
I really like your idea of the usage of 2D COV to LOB data and the attention layer. I have couple questions ; maybe you also think about them before. First, the accuracy doesn't improve much after several Epochs. Is that normal or suggesting an overfitting -- as the model has so many layers and the structure is complex. Second, can this model applied to regression problems? I changed the decoder input to shape (1,1), and decoder_dense output shape to 1. However, it often predict a constant y-pred value. Does that mean this structure is not suitable for regression problems? or need to be simplified. Thank you so much.

Recommend Projects