Algorithmic trading systems face a fundamental dilemma: optimizing strategy parameters requires market data, but acquiring this data through live trading exposes capital to risk. We present ARTEMIS, a multi-agent architecture that resolves this through shadow mode—a dual-execution paradigm where rejected trading signals are simulated at zero capital cost, capturing counterfactual outcomes for continuous learning. ARTEMIS comprises five specialized agents (historical analyzer, real-time scanner, dynamic executor, closure agent, and Bayesian optimizer) that operate asynchronously on live cryptocurrency markets. Experiments on 60 days of live trading show that ARTEMIS improves win rate from 31.2% to 41.3% and turns cumulative PnL from negative to positive while performing twice as many parameter updates as real-only systems.
David Lopez Oñate (Mon,) studied this question.