FinRS: A Risk‑Sensitive Trading Framework for Real‑World Financial Markets

Background Large language models (LLMs) have shown strong reasoning abilities for financial trading, yet most LLM‑based agents focus on single‑step prediction and lack integrated risk management, performing poorly in volatile markets.

Problem Definition Existing methods suffer from oversimplified task modeling, insufficient risk awareness, shallow information processing, and ad‑hoc feedback, preventing effective risk handling in sequential multi‑stage trading.

Method

3.1 Market Perception and Analysis Module Inspired by institutional workflows, multiple analyst agents filter raw data (news, filings, macro narratives) for relevance and price impact, annotate reasons for traceability, and pass filtered signals to specialized agents. A hierarchical memory stores stable signals (e.g., annual reports) in deep layers and volatile signals (e.g., breaking news) in shallow layers, dynamically promoting profitable signals and demoting misleading ones.

3.2 Risk‑Sensitive Decision Module Two core agents generate concrete trade orders: a direction‑decision agent selects buy/sell/hold based on retrieved experience, market dynamics, and current positions, while a quantity‑and‑risk agent dynamically scales trade size using account exposure, memory cues, and risk constraints (scaled Kelly criterion, CVaR). This replaces fixed‑size assumptions of prior work and jointly outputs direction and position size, explicitly controlling capital risk and potential drawdown.

3.3 Multi‑Scale Reward Feedback Module Rewards combine short‑, medium‑, and long‑term momentum signals: for each timestep t, compute M_t^s = price[t+1]‑price[t], M_t^m = price[t+7]‑price[t], M_t^l = price[t+30]‑price[t]; aggregate M_t = M_t^s + M_t^m + M_t^l. The reward is defined as Reward_t = d_t * M_t - \lambda * q_t^2, where d_t is the trade direction and q_t the trade quantity. This multi‑scale feedback encourages alignment with future trends while penalizing passive positions during high volatility.

Experiments

4.1 Experimental Setup Real‑world daily OHLCV data from Yahoo Finance, news from Finnhub, and SEC filings were used for five representative stocks (TSLA, AAPL, AMZN, NFLX, COIN). Baselines included four LLM agents (FinGPT, FinMem, FinAgent, FinCon), three deep RL methods (A2C, PPO, DQN), two rule‑based strategies (MACD, RSI), and two market baselines (random, buy‑and‑hold). All LLM agents used GPT‑4o (temperature 0.7). Evaluation metrics were cumulative return (CR %), Sharpe ratio (SR), and maximum drawdown (MDD %).

4.2 Main Results FinRS outperformed all baselines on every metric, achieving cumulative returns above 50% while maintaining lower drawdowns. On high‑volatility stocks (TSLA, AAPL, AMZN) FinRS showed strong robustness; other LLM agents and RL methods occasionally suffered large losses or negative returns. For relatively stable stocks (NFLX, COIN) FinRS delivered more than double the returns of competing methods.

4.3 Ablation Study Removing any component (risk‑sensitive reasoning, financial‑insight prompting, market‑news integration, or multi‑time‑scale reward) caused noticeable performance drops. For example, omitting risk‑sensitive reasoning reduced TSLA’s cumulative return from 54.99% to 39.19% and worsened MDD from 42.34% to 82.89%. Excluding the financial‑insight prompt lowered TSLA’s return to 41.57% and SR from 0.67 to 0.49. Dropping market‑news integration cut TSLA’s return to 45.41% and increased MDD to 55.40%.