Agentic Loop Engineering for Imbalanced Multi-Model and Time-Series Data in Large-Scale Enterprise Systems

Abstract

Rare events carry most of the risk in large-scale enterprise systems, yet the evidence for them is scattered across data modalities and drifts over time, so a classifier trained once for accuracy learns to ignore the minority class. This paper treats the handling of imbalanced multi-modal time-series data as an exercise in loop engineering: an agentic controller continually adjusts the class weight, the resampling ratio, and the decision threshold of a fused classifier in response to measured minority-class performance. We formalise the operating point and its proper metrics, and prove two results a practitioner can use. First, the agentic threshold-control loopis a contraction and therefore converges geometrically to a target minority recall. Second, multi-modal fusion cannot lower the achievable detection performance relative to any single modality. We propose the Agentic Imbalance Loop Engine, a three-timescale architecture with a fast threshold loop, a medium reweighting and resampling loop, and a slow drift-tracking loop. Implementing everything from first principles, we report reproducible measurements: the threshold loop converges within about ten iterations and lifts F1 from 0.31 to 0.44, multi-modal fusion reaches a precision-recall area of 0.68 against at most 0.30 for any single modality, the engine sustains a G-mean of 0.77 at a one percent minority fraction where a naive baseline collapses to 0.22, and under concept drift it holds a G-mean of 0.81 where a frozen model falls to 0.60. The results give a rigorous basis for engineering the feedback loops that keep rare events in view as enterprise data grow more imbalanced and less stationary.