---
title: StochLOB: Heavy-Tailed Score Matching for Robust Limit Order Book Trend Prediction
url: https://www.ml-quant.com/papers/arxiv/2610.21081/
site: ML-Quant (https://www.ml-quant.com)
updated: 2026-10-16
license: Summaries CC BY 4.0; links go to the original sources
index: https://www.ml-quant.com/llms.txt
identifier: arXiv:2610.21081
source_url: https://arxiv.org/abs/2610.21081
featured: 2026-10-16
citations: unknown
topic: ML & AI Methods
---


# StochLOB: Heavy-Tailed Score Matching for Robust Limit Order Book Trend Prediction

StochLOB replaces Gaussian noise with jump-Lévy and stable-noise corruptions to capture bursty limit-order-book flow, improving tail-window stress diagnostics on held-out LOB data.

- Source: https://arxiv.org/abs/2610.21081
- Identifier: arXiv:2610.21081
- Released: 2026-10-16
- First featured: Quant Letter No. 135 (2026-10-16): https://www.ml-quant.com/issues/2026-10-16/
- Citations (Semantic Scholar): not tracked
- Published in: not yet
- Topic: ML & AI Methods
- Authors: Arshia Abolghasemi, Seyed Reza Tavakoli, Parsa Naderi, Nazanin Mirzaei, Behrad Jahani, Saba Yousefzadeh, Tahereh Javaheri, Mohammad Hosein Rohban, Gholamali Aminian, Ali Habibnia

## Abstract (arXiv, CC0)

Score-based diffusion methods usually perturb data with Brownian Gaussian noise, corresponding to a light-tailed forward stochastic differential equation (SDE). For market time series this SDE can be a poor abstraction: limit order book (LOB) order flow is heavy-tailed, bursty, and occasionally jump-like. We study STOCHLOB, a training-only score-matching framework that changes the forward diffusion corruption law while leaving the deployed predictor deterministic. Instead of Brownian-only Gaussian corruption, the noising process uses two non-Gaussian SDE-inspired alternatives: jump-Lévy noise through a compound-Poisson variance mixture, and clipped $α$-stable noise through a positive-stable Gaussian scale mixture. These corruptions inject finite jumps and heavy-tailed residuals into training without requiring reverse-time diffusion sampling at inference. The resulting score target has a useful interpretation: its radial factor $h_t(r)$ is the posterior expected corruption precision, so large residuals are weighted according to the latent corruption scale that most plausibly generated them. On chronologically held-out Coinbase LOB data, the proposed variants remain competitive on uncorrupted BTC and altcoin test horizons and improve several realized tail-window stress diagnostics. An exploratory Feishu A-share transfer study further tests whether the probability outputs remain usable under transaction costs and trading-rule constraints. Code and configuration files are available at https://github.com/ArshiAbolghasemi/Penny.

## Related

- [Reinforcement Learning for Jump-Diffusions](https://www.ml-quant.com/papers/arxiv/2405.16449/): The study shows that continuous-time reinforcement learning can be applied to both pure diffusion and jump-diffusion processes, useful in portfolio selection with stock prices modeled as a jump-diffusion.
- [Reinforcement Learning and Rational Expectations Equilibrium in Limit Order Markets](https://www.ml-quant.com/papers/ssrn/4682574/): A paper suggests that simple payoff-based reinforcement learning can help achieve rational expectations equilibrium in limit order markets, with speculators mainly providing liquidity.
- [Historical calibration of SVJD models with deep learning](https://www.ml-quant.com/papers/ssrn/4650097/): The paper suggests using deep neural networks to calibrate parameters of Stochastic Volatility Jump Diffusion models, proving to be more accurate, robust, and faster than other methods.
- [What Do Foundation Models for Market Microstructure Learn? A Controlled Study with Synthetic Limit Order Books](https://www.ml-quant.com/papers/arxiv/2610.20104/): Autoregressive transformers trained on limit-order-book message streams learn amortized Bayesian inference over market regimes, matching optimal prediction and enabling accurate parameter calibration.
- [Exact Phase-Type Point Processes for Certified Measurement and Prediction of Event Streams](https://www.ml-quant.com/papers/arxiv/2610.14948/): A multiscale phase-type Hawkes model measures how fast prices absorb macroeconomic news versus self-exciting echoes, showing direct absorption in about one second and extended responses up to 137 seconds.
- [Jump Components in Jump Diffusion Models](https://www.ml-quant.com/papers/ssrn/4732803/): A proposed nonparametric test can determine if a jump diffusion process contains a jump component or is a diffusion, with the test statistic showing standard normal distribution if there are no jumps.