TL;DR
TensorLOB is a novel deep learning architecture that forecasts financial market movements by treating Limit Order Book (LOB) data as a high-dimensional tensor rather than a flattened matrix. By preserving the natural multi-way structure of time, market side, price level, and order attributes, it achieves state-of-the-art performance on benchmarks while significantly reducing the number of parameters required for classification.
What Problem Does This Paper Address?
In high-frequency trading, the Limit Order Book (LOB) is the fundamental record of supply and demand. It is inherently multi-dimensional: it tracks orders over time, across two market sides (buy/sell), at multiple price levels, and records different attributes (price and volume) for each.
Most existing deep learning models, like DeepLOB, "flatten" these dimensions into a 2D matrix or a 1D vector before processing. This process collapses economically meaningful relationships—for example, the spatial relationship between adjacent price levels or the symmetry between bid and ask sides. The authors argue that this forces the model to relearn the basic geometry of the market from scratch, leading to a weaker "inductive bias" and sub-optimal forecasting.
Methods
The authors propose TensorLOB, a structure-preserving framework that operates entirely in the tensor domain. Its architecture consists of three main stages:
- Mode-Aware Input Normalization: Instead of standard scaling, it normalizes data from two views: a temporal view (across events) and a cross-sectional view (across price levels and sides).
- Multi-Head Kronecker-Factorized Feature Extraction: The model uses mode-wise tensor contractions. Instead of a standard dense layer that connects every input to every output, TensorLOB applies specific transformations to each "mode" (e.g., one transformation for time, another for price levels). Mathematically, this is equivalent to a Kronecker-factorized linear transformation, which enforces structure and reduces the parameter count.
- CP-Parameterized Tensor Classifier: To predict the final price movement (up, down, or stationary), the model uses a classifier based on Canonical Polyadic (CP) Decomposition. This allows the model to learn complex interactions between all four dimensions using a fraction of the parameters of a standard dense layer.
Datasets
The model was evaluated on the FI-2010 benchmark, a standard large-scale dataset for LOB forecasting.
- Data: It contains nearly 450,000 samples from five Finnish companies traded on the NASDAQ Nordic stock exchange.
- Task: Predict mid-price movements (Up, Down, Stationary) across four different time horizons: 10, 20, 50, and 100 future events.
- Comparison: It was tested against a wide array of baselines, including DeepLOB, Bilinear Networks (TABL), and traditional machine learning models like XGBoost and SVMs.
Key Insights and Ideas
- Structure as Inductive Bias: The core "aha" moment is that the geometry of the LOB is a feature, not a hurdle. By forcing the neural network to respect the 4D structure, the authors provide a "shortcut" for the model to understand market microstructure.
- Parameter Efficiency: The use of CP decomposition in the classification layer resulted in an 85.3% reduction in parameters compared to a standard dense classifier (2,115 vs 14,403 parameters). This suggests that tensor methods can achieve better results with much lighter models.
- Temporal Scalability: TensorLOB's performance advantage over flattened models actually grows as the prediction horizon increases (e.g., at the 100-event horizon), suggesting that preserving structure helps the model capture longer-term latent dynamics that are lost in translation when data is vectorized.
Key Results
TensorLOB achieved the highest mean accuracy, weighted recall, and weighted F1 score across all four evaluated horizons.
- At the 20-event horizon: TensorLOB's F1 score (72.2%) outperformed the next best method, BiN-C(TABL), by 3.3 percentage points.
- At the 100-event horizon: It maintained a significant lead with a 79.1% F1 score, while many baseline models (like DeepLOB-Seq and LSTM) saw their performance degrade significantly.
- Stability: The model showed very low standard deviation across different random seeds, indicating high training stability compared to attention-based or recurrent baselines.
Limitations and Future Work
- Computational Complexity: While parameter-efficient, performing multiple mode-wise tensor contractions can be computationally demanding on hardware not specifically optimized for tensor operations compared to standard matrix-multiplication-heavy architectures.
- Fixed Structure: The current model assumes a fixed number of price levels (10). Future work could explore how to handle LOBs with variable depths or incorporate more diverse data types (like order flow or metadata).
- Interpretability: Although the model preserves the "modes" of the data, further research is needed to decode exactly what the "heads" in the multi-head projection are learning about market behavior (e.g., are some heads focusing on liquidity while others focus on price volatility?).
Summary updated