ARTEMIS: Adaptive Multi-Agent Trading System with Risk-Free Exploration via Shadow Mode

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 
(Adaptive Risk-free Exploration for Trading with Multi-agent 
Intelligent Selection), 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) deployed as AWS Lambda functions on live cryptocurrency 
markets. Experiments on 301 trades over 60 days show win rate 
improvement from 31.2% to 41.3% (+10.1pp, p<0.01 Wilcoxon), with 
2× more parameter updates than real-only systems via shadow 
exploration.

v1.1: Corrected shadow PnL formula, vol/momentum grid quantiles 
aligned with source code, Reporter activation updated to 
quantile-based filtering with Wilson LCB, BTCFOMO added as 9th 
executor filter.