Abstract
The success of deep learning-based limit order book forecasting models is highly dependent on the quality and the robustness of the input data representation. A significant body of the quantitative finance literature focuses on utilising different deep learning architectures without taking into consideration the key assumptions these models make with respect to the input data representation. In this paper, we highlight the issues associated with the commonly-used representations of limit order book data from both a theoretical and practical perspectives. We also show the fragility of the representations under adversarial perturbations and propose two simple modifications to the existing representations that match the theoretical assumptions of deep learning models. Finally, we show experimentally how our proposed representations lead to state-of-the-art performance in both accuracy and robustness utilising very simple neural network architectures.
Introduction
In quantitative finance, the Limit Order Book (LOB) serves as the fundamental record of all outstanding buy (bid) and sell (ask) orders for a given asset. Predicting short-term price movements from this high-frequency data is a central challenge for market participants. Most existing research utilizes "Level-II" data, which aggregates orders into the top $L$ price levels (e.g., the 10 best bids and asks).
The Problem: Non-Homogeneous Representations
The authors identify a critical flaw in the standard "compressed" Level-II representation: it is non-homogeneous. Standard deep learning architectures, such as Convolutional Neural Networks (CNNs), operate on the assumption that a specific input channel or index consistently represents the same data source.
In a level-based LOB, however, the meaning of "Level 1" is volatile. If a single small order is placed at a new "best" price, every subsequent level shifts down by one index. To a neural network, this looks like a massive transformation of the entire market state, even if the actual change in liquidity is negligible. This mismatch makes models fragile, prone to overfitting, and highly susceptible to "adversarial" noise—tiny orders placed at previously empty price ticks that trigger these level shifts.
Proposed Robust Representations
To address these issues, the authors propose moving from "level-based" indexing to "price-tick-based" indexing, which ensures spatial homogeneity:
- Moving-Window Representation: Instead of the top $L$ levels, a fixed-size grid is centered at the mid-price. Each index corresponds to a specific price tick (e.g., $0.01 increments). The value at each index is the volume at that price (positive for asks, negative for bids). This ensures a price level at a specific offset from the mid-price always maps to the same input index.
- Market Depth Representation: This representation uses the cumulative volume from the mid-price outwards. It captures the market's "depth"—its capacity to absorb large orders—in a smoothed format that is less sensitive to individual tick fluctuations.
Experimental Setup and Results
The study utilized the FI-2010 dataset, a benchmark containing 10 days of trading data for 5 stocks from the Helsinki Stock Exchange. The task was to classify future mid-price movements as Up, Down, or Stationary.
Key Findings
- Performance Gains: Simply switching to the proposed representations improved the accuracy of a basic Linear model by 7% and a Multi-Layer Perceptron (MLP) by 11%.
- Robustness to Adversarial Noise: The authors simulated a "valid market operation" by adding tiny orders (size 1) to empty price ticks. While these orders did not change market dynamics, they caused the accuracy of the state-of-the-art DeepLOB model to crash from 77.30% to 51.35%. In contrast, the Moving-Window and Market Depth representations maintained stable performance with zero significant decay.
- State-of-the-Art (SOTA): A simple Temporal Convolutional Network (TCN) using the Market Depth representation achieved a new SOTA accuracy of approximately 78.8%.
Conclusion
The research demonstrates that the "compressed" representations widely used in LOB forecasting violate fundamental assumptions of deep learning models. By adopting spatially homogeneous formats like Moving-Window and Market Depth, researchers can achieve significant improvements in both accuracy and robustness. This suggests that data representation is just as critical as architectural innovation in high-frequency trading applications. Future work will explore these representations in the context of Reinforcement Learning for trade execution and more advanced Transformer-based architectures.
Summary updated