Beyond random walks: exploring the learnability threshold of AI agents in algorithmic markets
Expert Systems with Applications, vol.315, 2026 (SCI-Expanded, Scopus)
- Publication Type: Article / Article
- Volume: 315
- Publication Date: 2026
- Doi Number: 10.1016/j.eswa.2026.131776
- Journal Name: Expert Systems with Applications
- Journal Indexes: Science Citation Index Expanded (SCI-EXPANDED), Scopus, Compendex, INSPEC, Public Affairs Index
- Keywords: Adaptive decision systems, Agent-based computational economics, Algorithmic trading, C45, C63, D83, G14, G17, Learnability threshold, Market complexity, Market efficiency
- İstanbul Ticaret University Affiliated: Yes
Abstract
Financial markets contain statistically detectable patterns whose economic exploitability remains uncertain. This study introduces the Learnability Threshold the boundary beyond which detectable patterns cannot yield positive net-of-cost returns for AI agents. This study compares a rule-based heuristic with Proximal Policy Optimization (PPO) agents (trained tabula rasa and via imitation) in simulated markets with transaction costs. To ensure robustness, aligned-path evaluations are conducted and richer observation spaces are tested. Results show DRL agents consistently fail to exploit long-memory dynamics, converging to inactivity or loss-making behavior, whereas the heuristic delivers stable risk-adjusted returns. The findings formally distinguish statistical detectability from economic exploitability and reposition DRL as a diagnostic decision-support tool for identifying unexploitable market regimes rather than a standalone profit engine.