Institutional infrastructure based on the ARTEMIS architecture. We solve the exploration-exploitation dilemma through counterfactual learning and multi-agent intelligence.
Experiments on 301 trades over 60 days demonstrate that ARTEMIS outperforms static baselines through continuous Bayesian optimization.
| Methodology | Win Rate | Cumulative PnL | Status |
|---|---|---|---|
| Fixed (Static Baseline) | 31.2% | -$45.30 | Underperforming |
| Random Updates | 28.7% | -$78.60 | Non-viable |
| Shadow-Only Learning | 35.1% | -$12.20 | Neutral |
| ARTEMIS (Full System) | 41.3% | +$18.70 | Active Alpha |
Acquiring market data through live trading exposes capital to risk. ARTEMIS resolves this via a novel dual-execution paradigm.
If filters become too restrictive, a system can stagnate. ARTEMIS features an automatic Starvation Mode for continuous adaptation.
Author: David López Oñate · Kinqo AI, Popayán, Colombia. Published on Zenodo. This work demonstrates decoupled exploration via architectural innovation in live cryptocurrency markets.
VIEW ON ZENODO@article{onate2026artemis,
title = {ARTEMIS: Adaptive Multi-Agent
Trading System with Risk-Free
Exploration via Shadow Mode},
author = {L{\'o}pez O{\~n}ate, David},
journal = {Zenodo Preprint},
year = {2026},
doi = {10.5281/zenodo.18565949}
}
Our research roadmap for 2026–2027 focuses on portfolio-level optimization and transfer learning across markets.
Correlation matrices and Kelly criterion sizing.
Expansion to Forex and Meta-learning adaptation.
Pre-trained transformer for universal market regimes.
Connect your Binance Futures via non-custodial API. Regulated by code.
REQUEST INTEGRATION