The Inference-Compute Frontier and a Latency-Efficient Architecture for Limit Order Book Prediction
C. Evans Hedges
Read on arXiv →
Key claim
FastBiNLOB outperforms existing models with lower latency.
In plain English
Imagine you're trying to predict stock market orders based on historical data. You might use various models, from simple decision trees to complex neural networks, but the challenge is figuring out how much computational power you need to minimize prediction errors. Traditionally, people have struggled to find the right balance between the complexity of their models and the accuracy of their predictions, leading to inefficiencies and missed opportunities. This is what's called the inference-compute frontier, and it can be tricky to navigate because not all models scale the same way with increased compute power.
In this paper, the authors explore this frontier specifically for limit order book predictions, which are crucial for trading strategies. They discover that the relationship between predictive loss and the amount of compute used can be described by a power law, meaning that as you increase compute, you can expect diminishing returns on prediction accuracy. This insight helps clarify how to allocate resources effectively when building predictive models. However, they also find that when it comes to latency — the time it takes to make predictions — the relationship isn't as straightforward, indicating that simply throwing more compute at the problem won't always yield better results.
To address these challenges, they introduce FastBiNLOB, a new model designed to mix features and temporal data in a way that's efficient for hardware. In tests, FastBiNLOB not only meets but exceeds existing benchmarks for prediction accuracy while operating at lower latency. This means that for builders in finance looking to optimize their trading algorithms, FastBiNLOB offers a compelling option that balances speed and accuracy better than previous state-of-the-art models.
The paper introduces a new approach to limit order book prediction using power-law relationships, which extends existing methods.
The findings are supported by empirical results and a solid experimental setup, including multiple models and a high R^2 value.
Deep reliability assessment
The methodology supports a claim that, on FI-2010 and for the tested model families, realized loss-vs-structural-work frontier points are well fit by a power law that extrapolates to held-out MLPLOB operating points. It does not support a general scaling law for LOB prediction across markets, time periods, datasets, hardware, or deployment stacks, and the SOTA latency/F1 claims are hard to fully assess from the provided prose without exact macro-F1 tables.
Reproducibility
Dataset appears reproducible because FI-2010 is a public benchmark; no code repository is mentioned in the provided abstract, introduction, results, limitations, or conclusion snippets.
Key figure
The key figure likely plots predictive loss against structural forward work on log-log axes, showing a power-law frontier fit from lower-compute non-MLPLOB models extrapolating to held-out high-compute MLPLOB points, contrasted with a weaker latency-based frontier.