Multi-Horizon Forecasting for Limit Order Books: Novel Deep Learning Approaches and Hardware Acceleration using Intelligent Processing Units (2105.10430v2)

Abstract: We design multi-horizon forecasting models for limit order book (LOB) data by using deep learning techniques. Unlike standard structures where a single prediction is made, we adopt encoder-decoder models with sequence-to-sequence and Attention mechanisms to generate a forecasting path. Our methods achieve comparable performance to state-of-art algorithms at short prediction horizons. Importantly, they outperform when generating predictions over long horizons by leveraging the multi-horizon setup. Given that encoder-decoder models rely on recurrent neural layers, they generally suffer from slow training processes. To remedy this, we experiment with utilising novel hardware, so-called Intelligent Processing Units (IPUs) produced by Graphcore. IPUs are specifically designed for machine intelligence workload with the aim to speed up the computation process. We show that in our setup this leads to significantly faster training times when compared to training models with GPUs.

• The paper introduces a multi-horizon forecasting model using encoder-decoder architectures with Seq2Seq and attention to generate sequential limit order book predictions.
• It demonstrates enhanced long-range prediction accuracy on datasets like FI-2010 and LSE, outperforming traditional single-step models in real-time trading.
• The use of Graphcore IPUs significantly accelerates training, illustrating practical benefits for high-frequency trading applications.

Multi-Horizon Forecasting for Limit Order Books: An Analysis

The paper presents a thorough investigation into the design and application of multi-horizon forecasting models specifically tailored for limit order book (LOB) data using advanced deep learning techniques. This domain of high-frequency trading and financial technology has long posed formidable challenges due to the inherent stochastic nature and low signal-to-noise ratio of financial time series. The authors leverage complex neural network architectures, namely encoder-decoder models with Sequence-to-Sequence (Seq2Seq) and Attention mechanisms, to address these challenges effectively.

Methodological Advancements

Central to this research is the shift from single-point predictions to multi-horizon forecasts using neural architectures inspired by natural language processing advancements. The encoder-decoder framework is adapted to extract temporal features efficiently and subsequently make informed long-term predictions. This method contrasts with traditional models that typically limit themselves to single-step or short-range forecasting due to the noisiness and dynamic nature of financial data.

1. Encoder-Decoder Models: By employing Seq2Seq and Attention mechanisms, the models can create a full trajectory of the predicted outputs rather than a solitary forecasted point. This approach allows for a more nuanced understanding of the market dynamics by predicting a sequence of future market states.
2. Model Performance: The authors validate their methods against well-established benchmarks using the FI-2010 dataset and data from the London Stock Exchange (LSE). Their approach demonstrates competitive performance, especially at longer prediction horizons. This is significant as it suggests the robustness of the autoregressive framework built into their forecasting system.
3. Hardware Acceleration: The implementation on Graphcore’s Intelligent Processing Units (IPUs) showcases a significant improvement in training speeds. This aspect is particularly salient for financial applications where speed is critical. The results indicate a remarkable increase in training efficiency, highlighting IPUs as a promising hardware solution for deep learning models in financial contexts.

Numerical Results and Implications

The empirical evaluations reveal that the proposed models achieve substantial improvements over existing methods, particularly for predictions at extended horizons. For instance, the models managed to outperform contemporary techniques for multiple steps ahead, which can be extremely beneficial in real-time trading strategies where understanding the future trajectory of prices is crucial.

Furthermore, the utilization of attention mechanisms not only improves predictive performance but also provides interpretability insights. By observing attention weights, practitioners can discern which features of the LOB are most influential at different time stamps. This might offer traders and financial analysts more depth in understanding how market conditions are influenced by varying factors.

Theoretical and Practical Considerations

The theoretical contribution of adapting Seq2Seq and Attention architectures from NLP to financial contexts demonstrates the versatility and power of these models. It sets a precedent for cross-disciplinary applications of machine learning techniques. Moreover, the paper highlights the potential of deep learning and hardware acceleration in overcoming the computational barriers frequently encountered with complex financial datasets.

From a practical standpoint, the research holds implications for algorithmic trading and electronic market-making where latency and prediction accuracy can lead to a tangible financial advantage. Future developments might explore the integration of these forecasting models with trading algorithms, potentially employing reinforcement learning as suggested by the authors.

Future Work

The authors suggest examining IPUs in more extensive configurations and their applicability in other domains of finance beyond LOB models. Another intriguing avenue is reinforcing the encoder-decoder model with methods from reinforcement learning, which could enhance decision-making in market conditions that are rapidly changing.

In conclusion, the paper makes a substantive contribution to the fields of quantitative finance and deep learning by presenting a thoughtfully crafted approach to LOB forecasting, augmented through novel technological implementation strategies. This work not only expands the toolkit available for financial data analysis but also positions multi-horizon forecasting models as a pivotal asset in understanding and anticipating market dynamics.