Temporal Kolmogorov-Arnold Networks (T-KAN) for High-Frequency Limit Order Book Forecasting: Efficiency, Interpretability, and Alpha-Decay

Battle Count: 107

Highly relevant to quantitative trading, particularly in the HFT and market-making domains. The paper directly addresses alpha decay, transaction-cost-adjusted profitability, and low-latency implementation (FPGA). The T-KAN architecture offers both improved predictive performance and interpretability, which are critical for trading strategy development. The 132.48% terminal return vs. -82.76% for DeepLOB under 1.0 bps costs is economically significant. However, the simplified backtesting methodology and single-dataset evaluation temper the practical applicability. The interpretability via learned B-spline dead-zones is valuable for strategy validation and regulatory compliance.

Moderate to high complexity. Requires implementing a hybrid LSTM-KAN architecture with learnable B-spline activation functions (Cox-de Boor recursion), inverse frequency weighting, L1 sparsity regularization, and custom classification heads. The B-spline parameterization and knot grid management add complexity beyond standard deep learning. FPGA/HLS implementation would require additional hardware engineering expertise. However, the code is publicly available, reducing the barrier to reproduction. The architecture uses 532,675 parameters (or 104,451 for the classification head), which is manageable for GPU training but the spline evaluation adds computational overhead per forward pass.

Code is publicly available on GitHub. Uses the standard FI-2010 benchmark dataset. Detailed hyperparameters are provided (feature dimensionality, look-back window, prediction horizons, LSTM units, L1 penalty, transaction costs). Strict chronological train/test split is described. However, some implementation details of the KAN classification head and B-spline grid configuration could be more explicit. The paper is a student research journal article, so peer-review rigor may be limited.