ReLaMix addresses the challenge of high-frequency financial time-series forecasting under realistic latency-induced observation staleness, where signals are corrupted by a Zero-Order Hold (ZOH) mechanism that produces repeated stepwise artifacts. Unlike standard missing-value settings, ZOH stagnation introduces statistical bias and redundancy that mislead forecasting models. The proposed framework integrates an Information Bottleneck compression module that projects noisy inputs into a compact latent space to filter out stale redundancy, followed by stacked Residual Latency-Aware Mixing Blocks. Each block combines temporal mixing via linear transformations along the time dimension with feature mixing through an expand-compress mechanism (32→64→32 dimensions) to capture nonlinear market dynamics while maintaining efficiency. Multi-level skip connections further enhance gradient flow and information preservation.

Experiments on second-resolution PAXGUSDT and BTCUSDT data from Binance demonstrate that ReLaMix consistently achieves state-of-the-art accuracy across multiple delay ratios (15%, 25%, 35%) and prediction horizons (k=1,5,7,10), significantly outperforming strong baselines including TimeMixer, TimesNet, PatchTST, LSTM, and GRU. Notably, ReLaMix attains this with only ~14K parameters, an order of magnitude fewer than Transformer-based models. Ablation studies confirm the critical roles of both bottleneck compression and residual refinement: removing compression increases parameters and degrades accuracy, while removing residual refinement causes catastrophic performance collapse. The framework also shows robust cross-asset generalization to BTCUSDT under severe delays. Overall, ReLaMix provides an efficient and principled solution for delay-robust high-frequency forecasting, balancing recovery accuracy with computational efficiency suitable for real-time trading environments.