Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

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

Ahmad Makinde

TL;DR

The paper tackles the problem of forecasting high-frequency LOB dynamics where alpha decay limits traditional models. It introduces Temporal Kolmogorov-Arnold Networks (T-KAN), which replace fixed LSTM gates with learnable spline activations to capture the shape of market signals, coupled with a KAN-based classification head. On the FI-2010 dataset and horizon k = 100, T-KAN delivers a 19.1% relative improvement in F1-score and a substantial backtest advantage (132.48% PnL vs -82.76% for DeepLOB under 1 bps costs), while remaining interpretable through dead-zone activations and hardware-friendly for FPGA-HLS. The work demonstrates improved alpha persistence, discusses economic viability, and points toward practical deployment in ultra-low-latency trading environments.

Abstract

High-Frequency trading (HFT) environments are characterised by large volumes of limit order book (LOB) data, which is notoriously noisy and non-linear. Alpha decay represents a significant challenge, with traditional models such as DeepLOB losing predictive power as the time horizon (k) increases. In this paper, using data from the FI-2010 dataset, we introduce Temporal Kolmogorov-Arnold Networks (T-KAN) to replace the fixed, linear weights of standard LSTMs with learnable B-spline activation functions. This allows the model to learn the 'shape' of market signals as opposed to just their magnitude. This resulted in a 19.1% relative improvement in the F1-score at the k = 100 horizon. The efficacy of T-KAN networks cannot be understated, producing a 132.48% return compared to the -82.76% DeepLOB drawdown under 1.0 bps transaction costs. In addition to this, the T-KAN model proves quite interpretable, with the 'dead-zones' being clearly visible in the splines. The T-KAN architecture is also uniquely optimized for low-latency FPGA implementation via High level Synthesis (HLS). The code for the experiments in this project can be found at https://github.com/AhmadMak/Temporal-Kolmogorov-Arnold-Networks-T-KAN-for-High-Frequency-Limit-Order-Book-Forecasting.

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

TL;DR

The paper tackles the problem of forecasting high-frequency LOB dynamics where alpha decay limits traditional models. It introduces Temporal Kolmogorov-Arnold Networks (T-KAN), which replace fixed LSTM gates with learnable spline activations to capture the shape of market signals, coupled with a KAN-based classification head. On the FI-2010 dataset and horizon k = 100, T-KAN delivers a 19.1% relative improvement in F1-score and a substantial backtest advantage (132.48% PnL vs -82.76% for DeepLOB under 1 bps costs), while remaining interpretable through dead-zone activations and hardware-friendly for FPGA-HLS. The work demonstrates improved alpha persistence, discusses economic viability, and points toward practical deployment in ultra-low-latency trading environments.

Abstract

High-Frequency trading (HFT) environments are characterised by large volumes of limit order book (LOB) data, which is notoriously noisy and non-linear. Alpha decay represents a significant challenge, with traditional models such as DeepLOB losing predictive power as the time horizon (k) increases. In this paper, using data from the FI-2010 dataset, we introduce Temporal Kolmogorov-Arnold Networks (T-KAN) to replace the fixed, linear weights of standard LSTMs with learnable B-spline activation functions. This allows the model to learn the 'shape' of market signals as opposed to just their magnitude. This resulted in a 19.1% relative improvement in the F1-score at the k = 100 horizon. The efficacy of T-KAN networks cannot be understated, producing a 132.48% return compared to the -82.76% DeepLOB drawdown under 1.0 bps transaction costs. In addition to this, the T-KAN model proves quite interpretable, with the 'dead-zones' being clearly visible in the splines. The T-KAN architecture is also uniquely optimized for low-latency FPGA implementation via High level Synthesis (HLS). The code for the experiments in this project can be found at https://github.com/AhmadMak/Temporal-Kolmogorov-Arnold-Networks-T-KAN-for-High-Frequency-Limit-Order-Book-Forecasting.
Paper Structure (21 sections, 14 equations, 5 figures, 1 table)

This paper contains 21 sections, 14 equations, 5 figures, 1 table.

Table of Contents

  1. Introduction
  2. Literature Review
  3. 2.1 Deep Learning in Market Microstructure
  4. 2.2 Kolmogorov-Arnold Networks (KAN) and Spline Theory
  5. Recurrence and the T-KAN Hybrid
  6. Methodology
  7. Data framework and Supervised Learning Setup
  8. The Sliding Window Unit
  9. Architectural Specification
  10. DeepLOB Baseline(CNN-LSTM)
  11. Proposed T-KAN Configuration
  12. Vector Graphic Diagrams
  13. Optimmisation and Inverse Frequency Weighting
  14. Results
  15. Comparative Performance Metrics
  16. ...and 6 more sections

Figures (5)

  • Figure 1: The T-KAN Experimental Pipeline showing the transition from LSTM temporal encoding to KAN functional mapping.
  • Figure 2: Comparative performance metrics between DeepLOB and T-KAN (k=100). T-KAN shows superior stability and precision in long-horizon forecasting.
  • Figure 3: Learned B-spline activation function of the T-KAN model. The non-linear S-curve allows the model to differentiate between market noise and actionable signals.
  • Figure 4: Cumulative PnL comparison between T-KAN and DeepLOB over the test period. The T-KAN model demonstrates significantly higher resilience to 1.0 bps transaction costs.
  • Figure 5: Alpha Decay Comparison: Information Coefficient (IC) vs. Forecast Horizon (k). T-KAN maintains higher predictive persistence over longer horizons compared to DeepLOB.