How 3S‑Trader’s Multi‑Agent Framework Optimizes Multi‑Stock Portfolios

Background Large language models (LLMs) have shown promise in single‑stock trading but lack the ability to reason about multi‑stock portfolios and adapt strategies to market volatility. Traditional quantitative methods (ARIMA, LSTM, Transformer) focus on time‑series prediction and ignore portfolio‑level risk control, while reinforcement learning suffers from high training cost and fragile reward design. The authors therefore propose a training‑free, self‑adjusting multi‑LLM framework called 3S‑Trader to directly build and iteratively improve portfolios from market signals.

Problem Definition

The task is weekly portfolio construction for a candidate set \(X = \{x_1, x_2, …, x_n\}\). The goal is to find a weight vector \(w_t\) that maximizes weekly return \(R_t = w_t \cdot r_t\), where \(r_t\) is the vector of individual stock returns. The main challenges are (1) efficiently integrating heterogeneous data (news, fundamentals, technical indicators) to enable multi‑stock comparison, and (2) dynamically adjusting the strategy to cope with market changes.

Method

3S‑Trader consists of four stages—data parsing, stock scoring, stock selection, and strategy iteration—implemented by six specialized LLM agents (news, fundamentals, technical, scoring, selection, strategy) built on GPT‑4o.

3.1 Market Analysis (Data Analysis)

Three dedicated agents parse multi‑source data and produce a comprehensive analysis report for each stock:

News Agent : extracts weekly news text and outputs a sentiment score \(\alpha_{i,t}^{news}\).

Technical Agent : processes the past four weeks of price and technical indicators (SMA, ATR, RSI, etc.) and outputs a trend score \(\alpha_{i,t}^{tech}\).

Fundamental Agent : reads the latest four quarters of financial statements and outputs a health score \(\alpha_{i,t}^{fund}\).

The three scores are combined into a multidimensional overview for each stock.

3.2 Stock Scoring

The Scoring Agent evaluates each stock on six dimensions (financial health, growth potential, news sentiment, news impact, price momentum, volatility risk) using a 1‑10 scale, producing a score vector \(s_{i,t}\) and a brief rationale.

3.3 Stock Selection

The Selection Agent combines the scoring reports \(s_{i,t}\) with the current strategy \(\pi_t\) (e.g., prioritize financial health, avoid high‑volatility stocks) to pick up to five stocks and assign weights that sum to at most one.

3.4 Strategy Iteration

The Strategy Agent analyses historical strategy trajectories, the current strategy \(\pi_t\), portfolio weights \(w_t\), weekly returns \(r_t\), and scoring reports \(s_t\) to identify common features of high‑ and low‑return stocks, then updates the strategy for the next week.

Experiments

4.1 Experimental Setup

Datasets : DJIA constituents (30 large‑cap US stocks), Nasdaq‑100 technology sector (44 stocks), S&P financial sector ETF (49 stocks), and S&P healthcare sector ETF (46 stocks). Data sourced from Alpha Vantage, covering 2022‑05‑16 to 2024‑05‑27 (training set 2012‑05‑01 to 2022‑05‑15).

Baselines : Equal‑weight (1/N), rule‑based models (SMA, MACD, BOLL), deep‑learning models (LSTM, Informer, Autoformer) that predict weekly returns and select top‑5 stocks, and multi‑LLM baselines (single‑step TradingAgent, reflective CryptoTrade).

Evaluation Metrics : Cumulative return (AR), Sharpe ratio (SR), Calmar ratio (CR).

Results and Analysis

Overall Performance : 3S‑Trader ranks in the top two across all sectors and achieves the best results on DJIA (AR = 131.83 %, SR = 0.31, CR = 11.84), with a consistently upward trend.

Comparison with Rule‑Based Baselines : On DJIA, 3S‑Trader’s AR (131.83 %) far exceeds SMA (70.63 %), MACD (46.00 %), and BOLL (20.09 %). The advantage stems from LLMs’ flexible integration of multi‑source signals, whereas rule‑based methods rely on static indicators and cannot adapt to dynamic market conditions.

Comparison with Deep‑Learning Models : In the technology sector, LSTM attains AR = 193.39 % but suffers low risk‑adjusted performance (SR = 0.21, CR = 5.81). 3S‑Trader achieves AR = 183.29 % with higher SR = 0.27 and CR = 11.81, indicating better balance between return and risk thanks to its multidimensional scoring and strategy reasoning.

Reflective Framework in Volatile Markets : In the healthcare sector, the reflective baseline (SR = 0.18, CR = 6.51) outperforms the single‑step baseline (SR = 0.12, CR = 3.11). However, in the financial sector the reflective approach over‑adjusts and performs poorly. 3S‑Trader’s explicit scoring provides clear guidance for strategy refinement, yielding AR = 84.93 % (SR = 0.21, CR = 7.57) and superior stability.

Conclusion

3S‑Trader demonstrates that a training‑free, multi‑LLM architecture can effectively construct and iteratively improve multi‑stock portfolios, achieving competitive cumulative returns while maintaining strong risk‑adjusted metrics across diverse market segments.