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

Source: arXiv - 2601.02310v1

Overview

The paper introduces Temporal Kolmogorov‑Arnold Networks (T‑KAN), a new deep‑learning architecture for forecasting high‑frequency limit order book (LOB) dynamics. By swapping the static linear activations of conventional LSTMs for learnable B‑spline functions, T‑KAN captures the shape of market signals, dramatically reducing the “alpha decay” that plagues existing models such as DeepLOB.

Key Contributions

• Learnable spline activations: Replaces fixed LSTM weights with B‑spline functions, enabling the network to model non‑linear market patterns more flexibly.
• Significant performance boost: Achieves a 19.1 % relative improvement in F1‑score at a 100‑tick prediction horizon (k = 100) on the FI‑2010 dataset.
• Robust profitability: Generates a 132.48 % return versus a ‑82.76 % drawdown for DeepLOB when accounting for 1 bps transaction costs.
• Interpretability: The spline shapes expose “dead‑zones” (regions of low sensitivity), giving traders visual insight into which price movements the model deems irrelevant.
• FPGA‑ready design: Architecture is streamlined for low‑latency implementation via High‑Level Synthesis (HLS), making it suitable for hardware‑accelerated trading systems.
• Open‑source reproducibility: Full code and experiment scripts are released on GitHub.

Methodology

1. Data preprocessing – The authors use the FI‑2010 high‑frequency LOB dataset, extracting the top 10 price levels (bid/ask) and standard technical features (price differences, volumes, etc.).
2. Temporal KAN layer – Each LSTM cell is augmented with a Kolmogorov‑Arnold Network (KAN) activation: a piecewise B‑spline whose control points are learned during training. This allows the activation to adapt its curvature to the underlying market dynamics.
3. Network stack – A sequence of T‑KAN layers feeds into a fully‑connected head that outputs a multi‑class prediction (price up, down, or neutral) for the next k ticks.
4. Training regime – Standard cross‑entropy loss with Adam optimizer, early stopping, and class‑balanced mini‑batches to mitigate the heavy class imbalance typical of HFT data.
5. Hardware mapping – The spline evaluation is expressed as a series of simple arithmetic operations, which translates efficiently to FPGA pipelines using HLS directives (loop unrolling, pipelining).

Results & Findings

• Alpha decay mitigation: While DeepLOB’s predictive power collapses beyond k ≈ 30, T‑KAN maintains a steady F1‑score up to k = 100, indicating better long‑horizon signal retention.
• Interpretability demo: Visualizations of the learned splines show flat regions (dead‑zones) where the model ignores noisy micro‑fluctuations, and steep regions where it reacts strongly to decisive price moves.
• Hardware efficiency: The spline‑based activation reduces the number of multiply‑accumulate operations, cutting both power consumption and inference latency on an FPGA board.

Practical Implications

• Low‑latency trading bots: Developers can embed T‑KAN directly onto FPGA‑based market data adapters, achieving sub‑microsecond decision times—critical for latency‑sensitive strategies.
• Reduced model drift: The shape‑learning capability helps the model stay relevant longer, lowering the frequency of retraining cycles and operational overhead.
• Explainable AI for compliance: The visible dead‑zones give compliance teams a tangible way to audit why a model ignored certain price ticks, easing regulatory concerns around black‑box HFT models.
• Cost‑effective scaling: Because T‑KAN replaces heavy matrix multiplications with spline evaluations, it can run on modest hardware (e.g., edge devices or low‑cost FPGAs) without sacrificing accuracy, opening the door for boutique firms to compete with larger players.

Limitations & Future Work

• Dataset scope: Experiments are limited to the FI‑2010 dataset; performance on other exchanges, asset classes, or more recent market microstructure may vary.
• Hyper‑parameter sensitivity: The number and placement of spline knots require careful tuning; automated knot selection is not yet explored.
• Model complexity vs. interpretability trade‑off: While splines are more interpretable than deep ReLUs, the overall network depth can still obscure decision pathways for very deep stacks.
• Future directions: Extending T‑KAN to multi‑asset cross‑correlation forecasting, integrating reinforcement‑learning based execution policies, and developing automated knot‑optimization algorithms are suggested by the authors.

Authors

• Ahmad Makinde

Paper Information

• arXiv ID: 2601.02310v1
• Categories: cs.LG
• Published: January 5, 2026
• PDF: Download PDF