A Low-Latency Multi-Agent Architecture for Statistical Arbitrage, Market Making, and Volatility-Conditioned Leverage Scaling Across Asset Classes

Abstract

This institutional research memorandum presents the architecture, theoretical design, and empirical evaluation of a systematic, multi-asset intraday trading system designed for statistical arbitrage and market making across cryptocurrency perpetual derivatives, equity index derivatives, and CME futures. The system combines a four-agent signal ensemble, hardware-normalized Level-3 data pipelines, and a capital-allocation meta-learner governed by a Regime-Penalized Proximal Policy Optimization (RP-PPO) framework. Aggregate portfolio exposure is conditioned on the observed variance risk premium and governed by a deterministic, time-decaying leverage schedule that materially reduces overnight jump-diffusion exposure. The production specification was frozen prior to the 2025 holdout period. Across the untouched 2025 out-of-sample calendar year, the architecture recorded a mean daily net return of 2.51%, an intraday Sharpe ratio of 7.90 (HAC-adjusted 6.42), and a daily close-to-close Sharpe ratio of 4.81, within an empirical capacity ceiling of USD 15.0M. The memorandum also documents a 41.7% maximum drawdown driven by a 14-minute exchange matching-engine outage on March 14, 2025. These results are derived from an out-of-sample historical simulation under modeled market frictions and do not constitute a live-trading track record. Exact algorithmic formulations, network weights, and parameter manifests are proprietary and withheld.