AlphaQuanter introduces a novel single-agent framework for stock trading that leverages reinforcement learning to optimize a tool-augmented decision-making process. Unlike existing multi-agent systems that suffer from inefficiency and inconsistent signals, AlphaQuanter uses a single LLM agent that iteratively reasons, acquires information via external tools (market data, fundamentals, sentiment, macroeconomics), and makes trading decisions (BUY/SELL/HOLD). The framework models the trading task as a Markov Decision Process and trains the agent using GRPO with a carefully designed reward function that combines outcome-based rewards (using exponentially weighted forward returns with asymmetric penalties) and process-based rewards (encouraging efficient tool use and proper reasoning format).

Extensive backtesting on five major stocks (GOOGL, META, MSFT, NVDA, TSLA) over a six-month test period demonstrates that AlphaQuanter achieves state-of-the-art performance, with the 7B variant reaching an average annualized return of 34.94%, Sharpe ratio of 0.65, and maximum drawdown of 24.93%. These results significantly outperform strong baselines including GPT-4o, multi-agent frameworks like TradingAgents, and traditional RL methods. The training dynamics reveal that the 7B model learns to selectively use tools and refine its reasoning, while the 3B model tends to converge prematurely. Human evaluation further confirms that AlphaQuanter produces more faithful and coherent decision traces, with better alignment between stated information needs and tool calls, stronger evidence grounding, and greater conciseness. The work highlights the importance of end-to-end RL training for enabling small LLMs to learn robust trading policies that surpass even larger zero-shot models.