Multi-Horizon Limit Order Book Forecasting with Seq2Seq Models and IPUs

Summary

The document summarizes research on forecasting multiple future steps from limit order book data. Rather than predicting only one future point, the proposed approach uses sequence-to-sequence encoder-decoder networks with attention to generate a path of forecasts. The study builds on an existing deep network architecture and evaluates the method on the FI-2010 public dataset and a year of London Stock Exchange order book data. The reported findings describe competitive results for short horizons and stronger performance for longer horizons than the multi-level comparison approaches discussed.

The work also addresses the computational cost of recurrent networks by testing Graphcore intelligent processing units against GPUs. It reports that IPUs substantially reduced training time in the tested setup. These findings are specific to the data, models, and hardware configuration described; the document provides no detailed numerical results in the excerpt. Forecast paths may inform trading or risk decisions, but predictive accuracy alone does not establish execution-aware profitability, and the authors note low signal-to-noise as a central difficulty in financial time series.

Key ideas

• The method predicts a sequence of future limit order book outcomes instead of a single horizon.
• It uses sequence-to-sequence encoder-decoder networks with attention to generate forecast paths.
• Tests use the FI-2010 dataset and a year of London Stock Exchange order book data.
• The study reports competitive short-horizon results and better longer-horizon results than its multi-level comparisons.
• In the tested setup, Graphcore IPUs accelerated training relative to GPUs.
• Forecast accuracy does not by itself establish profitability after trading costs and execution constraints.

Tags

This summary was written by Stratmill's research agent from the original; it is not a copy of the source.