StochLOB
A score-matching framework for limit order book trend prediction. It replaces Brownian-Gaussian noise with jump-type Levy noise and truncated alpha-stable noise during the training-stage corruption process, while keeping a deterministic predictor at inference time, thereby improving robustness in modeling heavy-tailed distributions and extreme market events.
Timeline (last 90 days)
StochLOB proposes a training-only score matching framework that changes the corruption law of the forward diffusion, while the predictor at deployment time remains deterministic.
StochLOB introduces two non-Gaussian noise schemes: compound Poisson variance mixture (jump-Lévy noise) and truncated α-stable noise (positive-stable Gaussian scale mixture).
StochLOB's radial factor h_t(r) can be interpreted as the posterior expected corruption precision, causing the model to weight large residuals according to their most likely latent corruption scale rather than penalizing them uniformly.
StochLOB's adaptive weighting mechanism makes the loss function robust to outliers while preserving their informativeness.
The authors conducted a chronological holdout test on Coinbase's LOB data, and the proposed variant remained competitive on the uncorrupted BTC and altcoin test intervals.
StochLOB achieves improvements on stress diagnostic metrics in realized tail windows, indicating that the model performs more reliably under extreme market windows.
The paper conducted an exploratory Feishu A-share transfer study to examine whether the probabilistic outputs remain usable under trading cost and trading rule constraints.
StochLOB's code and configuration files have been open-sourced on GitHub under the repository name Penny.
The A-share transfer part of the paper remains exploratory in nature, with limited samples and data coverage, and StochLOB's generalization ability still awaits more independent replication tests.
StochLOB adopts a "heavy training, light inference" design, and deployment complexity and latency are not affected by the non-Gaussian noise mechanism.
All Facts (10)
StochLOB adopts a "heavy training, light inference" design, and deployment complexity and latency are not affected by the non-Gaussian noise mechanism.
50%UnverifiedStochLOB introduces two non-Gaussian noise schemes: compound Poisson variance mixture (jump-Lévy noise) and truncated α-stable noise (positive-stable Gaussian scale mixture).
50%UnverifiedStochLOB's radial factor h_t(r) can be interpreted as the posterior expected corruption precision, causing the model to weight large residuals according to their most likely latent corruption scale rather than penalizing them uniformly.
50%UnverifiedStochLOB's adaptive weighting mechanism makes the loss function robust to outliers while preserving their informativeness.
50%UnverifiedThe authors conducted a chronological holdout test on Coinbase's LOB data, and the proposed variant remained competitive on the uncorrupted BTC and altcoin test intervals.
50%UnverifiedStochLOB achieves improvements on stress diagnostic metrics in realized tail windows, indicating that the model performs more reliably under extreme market windows.
50%UnverifiedThe A-share transfer part of the paper remains exploratory in nature, with limited samples and data coverage, and StochLOB's generalization ability still awaits more independent replication tests.
50%UnverifiedStochLOB proposes a training-only score matching framework that changes the corruption law of the forward diffusion, while the predictor at deployment time remains deterministic.
50%UnverifiedThe paper conducted an exploratory Feishu A-share transfer study to examine whether the probabilistic outputs remain usable under trading cost and trading rule constraints.
50%UnverifiedStochLOB's code and configuration files have been open-sourced on GitHub under the repository name Penny.
50%