The Inference-Compute Frontier and a Latency-Efficient Architecture for Limit Order Book Prediction

A new study finds a power-law relationship between computational work and predictive accuracy in limit order book forecasting, and introduces a latency-efficient architecture that outperforms published benchmarks on the FI-2010 dataset [1]. The research, posted to arXiv on 24 June 2026, examines whether a scaling-law-style inference-compute frontier emerges in limit order book (LOB) prediction [1]. Using the FI-2010 dataset and a suite of models spanning small decision trees to neural LOB architectures, the authors report that the empirical frontier of predictive loss versus structural forward work is well summarized by a power law [1]. When the MLPLOB architecture family is held out, a power-law fit to the low- and mid-compute frontier extrapolates across multiple orders of magnitude and attains an R² of 0.941 on the excluded high-compute MLPLOB target [1]. The study then tests whether latency can substitute for structural work. A CPU-measured single-observation experiment replacing work with latency reaches only R²=0.468 on held-out MLPLOB observations [3]. The authors note a reordering of architectures in latency space, concluding that latency is not merely noisy compute but a distinct factor [3]. This gap motivates FastBiNLOB, a dense axis-separable LOB mixer built from hardware-friendly temporal and feature mixing operations [1]. In a five-seed experiment, FastBiNLOB exceeds the published y₁₀ macro-F1 target at lower latency than the MLPLOB anchor, with a median decrease of approximately 23% in single-batch inference time [3]. Its H120 taper variant exceeds the published y₁₀₀ target at lower latency than the TLOB anchor, with a median decrease of approximately 60% [3]. Under the published MLPLOB/TLOB comparison setting, the H120 taper variant posts selected-horizon state-of-the-art FI-2010 macro-F1 scores on both targets at notably lower latency [3]. Deep learning architectures have been applied across fields including computer vision, speech recognition, and natural language processing, where they have produced results comparable to and in some cases surpassing human expert performance [6]. Neural networks consist of connected layers of artificial neurons that process signals through weighted connections, with training accelerated by graphics processing units and large datasets [7]. The field of artificial intelligence, founded as an academic discipline in 1956, saw a substantial increase in funding and interest after 2012 when GPUs began accelerating neural networks, and growth accelerated further after 2017 with the transformer architecture [8]. The paper makes two linked contributions: evidence that structural forward work is a useful empirical capacity coordinate for heterogeneous LOB predictors on FI-2010, and a demonstration that measured latency is a separate design target [3]. The practical implication is that model capacity and serving latency should be optimized as separate objects in LOB prediction [3]. Scaling-law fits can identify whether useful computation exists, but they do not indicate whether that computation will be cheap to serve [3]. FastBiNLOB places useful temporal and feature mixing in dense operations that the serving runtime executes efficiently, suggesting that latency-efficient LOB modeling should be designed around where the computation sits, not only how much computation is counted [3].