[Paper Review] ReLaMix: Residual Latency-Aware Mixing for Delay-Robust Financial Time-Series Forecasting

Tianyou Lai, Wentao Yue+6 more|arXiv (Cornell University)|Mar 21, 2026

Stock Market Forecasting Methods0 citations

TL;DR

ReLaMix is a lightweight Residual Latency-Aware Mixing network that recovers clean high-frequency financial signals under ZOH latency, achieving state-of-the-art accuracy with far fewer parameters than baselines and generalizing across assets.

ABSTRACT

Financial time-series forecasting in real-world high-frequency markets is often hindered by delayed or partially stale observations caused by asynchronous data acquisition and transmission latency. To better reflect such practical conditions, we investigate a simulated delay setting where a portion of historical signals is corrupted by a Zero-Order Hold (ZOH) mechanism, significantly increasing forecasting difficulty through stepwise stagnation artifacts. In this paper, we propose ReLaMix (Residual Latency-Aware Mixing Network), a lightweight extension of TimeMixer that integrates learnable bottleneck compression with residual refinement for robust signal recovery under delayed observations. ReLaMix explicitly suppresses redundancy from repeated stale values while preserving informative market dynamics via residual mixing enhancement. Experiments on a large-scale second-resolution PAXGUSDT benchmark demonstrate that ReLaMix consistently achieves state-of-the-art accuracy across multiple delay ratios and prediction horizons, outperforming strong mixer and Transformer baselines with substantially fewer parameters. Moreover, additional evaluations on BTCUSDT confirm the cross-asset generalization ability of the proposed framework. These results highlight the effectiveness of residual bottleneck mixing for high-frequency financial forecasting under realistic latency-induced staleness.

Motivation & Objective

• Formalize high-frequency forecasting under Zero-Order Hold (ZOH) stagnation and identify failure modes of existing models.
• Propose ReLaMix, a bottleneck-based residual mixing framework that suppresses latency-induced redundancy while preserving nonlinear market dynamics.
• Demonstrate state-of-the-art accuracy and cross-asset generalization with a compact model suitable for real-time deployment.

Proposed method

• Apply an Information Bottleneck-style bottleneck projection to compress inputs from D_in to a small latent dimension (d_b).
• Use a residual latency-aware mixing block with two submodules: Time Mixing (temporal linear mixing with residuals) and Expand–Compress Feature Mixing (bottleneck expansion and compression).
• Incorporate multi-level skip connections to improve gradient flow and information propagation across layers.
• Train end-to-end with MSE loss to recover future clean dynamics from ZOH-corrupted histories.
• Evaluate on second-resolution PAXGUSDT with simulated ZOH delays and generalize to BTCUSDT.

Experimental results

Research questions

• RQ1 How does latency-induced staleness via ZOH corruption affect high-frequency forecasting models?
• RQ2 Can a lightweight bottleneck and residual mixing architecture recover true market dynamics under delayed observations more effectively than existing mixer/Transformer baselines?
• RQ3 Does ReLaMix maintain accuracy with far fewer parameters, enabling real-time deployment, and does it generalize across assets?
• RQ4 What is the contribution of bottleneck compression and residual refinement to performance under ZOH delays?
• RQ5 Can the approach transfer its advantages from PAXGUSDT to BTCUSDT?

Key findings

• ReLaMix achieves state-of-the-art forecasting accuracy on the second-resolution PAXGUSDT benchmark across delay ratios (15%, 25%, 35%) and horizons (k=1,5,7,10).
• At 15% delay and horizon k=1, ReLaMix attains MSE 0.02928, MAE 0.06109, and R2 0.99103 with 13.93K parameters, outperforming TimeMixer and other baselines.
• Across delays, ReLaMix maintains the smallest parameter footprint (~13.9K–15.4K) compared to TimeMixer (~41K), PatchTST (~102K), and TimesNet (~229K).
• Ablation shows bottleneck compression is crucial for suppressing ZOH redundancy and residual refinement is essential for preserving predictive dynamics; removing either component degrades performance significantly.
• ReLaMix generalizes to BTCUSDT, achieving best metrics under 25% and 35% delays with k=5 or k=10 (MSE 0.0328/0.0302, MAE 0.0226/0.0293, R2 0.7727/0.7865).
• Multi-level skip connections improve training stability and information flow, aiding convergence under latency-corrupted inputs.