MAD-FORCE: Targeting High-Breakdown Streaming Robust Correlation With Optional Drift Handling
Sooyoung Jang, Taehyun Park, Hyunbean Yi, Changbeom Choi · IEEE Access · 2026
Streaming robust-correlation estimators must keep bounded memory while resisting contamination. In Force, constant-state P2 quantile summaries use an interquartile range (IQR) scale, tying trimming to the classical 25% breakdown regime and limiting high-breakdown monitoring when stream storage is infeasible. We propose MAD-Force, which replaces IQR trimming with a two-stage streaming median absolute deviation (MAD) scale while preserving the Force pairwise correlation architecture. We analyze the core estimator under post-warm-up quantile-consistency assumptions, derive a finite-sample retained-contamination perturbation bound for standardized retained moments, and evaluate an optional acceptance-rate drift module through path-consistent ablations. Under these assumptions, the MAD path targets the classical $1/2$ MAD breakdown barrier, up to a transient $O(T_{0}/N)$ warm-up term. At $\varepsilon =0.25$ , it reduces stationary Frobenius RMSE from 0.0466 to 0.0307 on a 50-dimensional benchmark. Mechanism audits show lower retained contamination and more stable trimming scales. The streaming-only core reaches about 976 Hz at $p=20$ and 187 Hz at $p=50$ ; detector-enabled gains require informative acceptance-rate shifts. MAD-Force contributes an assumption-qualified constant-state high-breakdown scale substitution for dense streaming correlation, a retained-moment perturbation bound, an ablation-led empirical protocol, and an explicit operating-envelope analysis for settings where full stream retention is infeasible.