Agentic Trading: When LLM Agents Meet Financial Markets

arXiv:2605.19337v2 Announce Type: replace Abstract: Large Language Models (LLMs) combined with autonomous agent architectures are shifting quantitative finance from isolated predictive modeling toward closed-loop trading agents. These systems perceive multimodal market signals, maintain context and memory, reason about decisions, emit executable trading actions, and adapt to non-stationary market regimes. This survey provides a systematic synthesis of agentic trading through an Architecture--Capability--Adaptation (A-C-A) conceptual framework. Architecture covers perception (textual, numerical, and multimodal), memory (working memory, vector storage, and knowledge bases), reasoning (reactive, reflective, and planning-driven), and execution (order routing and friction management). Capability examines alpha discovery, portfolio optimization, and multi-tier risk control. Adaptation covers in-context learning, supervised fine-tuning, reinforcement learning, multi-agent coordination, and architectural self-evolution. Drawing on 185 systematically selected studies, including a core empirical subset of 80 systems evaluated under closed-loop market conditions, we analyze design spaces, operational trade-offs, and emerging paradigms. We also examine evaluation practices, including backtesting split schemes, transaction-cost realism, survivorship-bias handling, and execution fidelity, and distill methodological best practices for future research. By bridging artificial intelligence, market microstructure, and quantitative finance, this survey provides a roadmap for designing, benchmarking, and deploying reliable agentic trading systems.