Paper: arXiv 2601.02310

Authors: Ahmad Makinde

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.

Complexity vs Empirical Score

  • Math Complexity: 8.5/10
  • Empirical Rigor: 9.0/10
  • Quadrant: Holy Grail — high math complexity, high empirical rigor

Why this score: The paper introduces advanced mathematical concepts, including the Kolmogorov-Arnold Representation Theorem, B-spline theory, and Cox-de Boor recursion, alongside substantial LaTeX-formatted equations. It demonstrates high empirical rigor by using a specific benchmark dataset (FI-2010), providing backtest-ready results (F1-scores, returns, drawdowns) with transaction costs, and offering a public GitHub repository for code.

Research Flowchart

  flowchart TD
  Start["Research Goal<br>Address Alpha Decay & Interpretability<br>in High-Frequency LOB Forecasting"] --> Data
  subgraph Data
      A["FI-2010 Dataset<br>High-Frequency Limit Order Book Data"]
  end
  
  Data --> Model
  
  subgraph Model
      B["Temporal KAN (T-KAN)<br>Replace fixed LSTM weights with<br>learnable B-spline activation functions"]
  end
  
  Model --> Proc
  
  subgraph Proc ["Computational Processes"]
      direction LR
      C1["Feature Extraction<br>Market Signal Shapes"]
      C2["FPGA Optimization via HLS<br>Low-Latency Implementation"]
  end
  
  Proc --> Outcome
  
  subgraph Outcome ["Key Findings & Outcomes"]
      F1["F1-Score: +19.1% relative improvement<br>(k = 100 horizon)"]
      Ret["Strategy Return: 132.48%<br>vs. DeepLOB -82.76% (1.0 bps cost)"]
      Int["Interpretability: Visible 'Dead-Zones'<br>in spline activations"]
  end

Evidence Card

What the authors claim, on what data, and who has checked it. Read from the paper's own text by a language model: every quote is verbatim, nothing was re-run. How to read this card

Proposes

The paper proposes Temporal Kolmogorov-Arnold Networks (T-KAN) that replace fixed linear LSTM weights with learnable B-spline activation functions for LOB forecasting.

Source sentence
we introduce Temporal Kolmogorov-Arnold Networks (T-KAN) to replace the fixed, linear weights of standard LSTMs with learnable B-spline activation functions.

Assumes 5

  • The universal approximation approach with fixed activations is parameter‑inefficient for capturing localized oscillations in microstructure data. model
    Source sentence
    We assume that this ’universal approximation’ approach is parameter-inefficient for capturing localized oscillations found in microstructure data.
  • Mean and standard deviation are computed over a rolling window to keep the auction environment stationary. data
    Source sentence
    where µ and σ are the mean and standard deviation calculated over a rolling window to maintain a stationary auction environment.
  • Class imbalance is addressed by inverse frequency weighting in the loss function. model
    Source sentence
    To address the significant class imbalance in the FI-2010 Auction dataset, we utilize Inverse Frequency Weighting
  • A flat 1.0 basis‑point transaction cost is applied to the mid‑price trading simulation. execution
    Source sentence
    A mid-price trading simulation was conducted using a 1.0 bps transaction cost to evaluate the economic significance of the model’s predictions.
Show 1 more
  • The FI‑2010 dataset provides a representative benchmark for high‑frequency LOB forecasting. data
    Source sentence
    The FI-2010 benchmark dataset [4] was used to evaluate the performance of Temporal Kolmogorov-Arnold Networks against the DeepLOB baseline.

Data

Backtest sample: intraday

  • FI-2010 · Ntakaris et al. order-book public
    Source sentence
    using data from the FI-2010 dataset

Recognised sources: FI-2010

Costs

Modeled Costs are part of the reported headline results.

A flat 1.0 bps transaction cost is applied in the backtest.

Source sentence
A mid-price trading simulation was conducted using a 1.0 bps transaction cost to evaluate the economic significance of the model’s predictions.

Finds 4

Reported performance (Temporal Kolmogorov-Arnold Networks (T-KAN)): total return 132.5%

  • T‑KAN improves F1‑score by 19.1% relative to DeepLOB at a 100‑tick horizon (0.3995 vs 0.3354). performance
    Source sentence
    T-KAN achieved an F1-Score of 0.3995, representing a relative improvement of 19.1% over the baseline of 0.3354.
  • In a transaction‑cost‑adjusted backtest, T‑KAN yields a terminal return of 132.48% versus DeepLOB’s –82.76%. performance
    Source sentence
    the T-KAN model resulted in a terminal return of 132.48%
  • Learned B‑spline activations create a dead‑zone near zero‑mean inputs, filtering micro‑structural noise. other
    Source sentence
    This learned function effectively creates a "dead-zone" near zero-mean inputs, filtering out micro-structural noise while non-linearly amplifying high-conviction signals.
  • T‑KAN uses more parameters (104,451) than DeepLOB (58,211) but achieves higher profitability density per parameter. other
    Source sentence
    While T-KAN uses a far higher parameter footprint (104,451 vs. 58,211), the ’profitability density’ per parameter is significantly higher.

Checked by others

No other paper in the archive has been found to support or contradict these findings (compared with 45 related papers). Coverage is partial, so this is neither replication nor refutation.

Extracted 2026-11-14 · same card as JSON · A language model chose these items: check the source sentence before relying on one.