This study investigates the use of ChatGPT and DeepSeek for portfolio optimization in the financial investment sector. Based on a sample of constituent stocks from the CSI 300 Index, the study compares the portfolio construction characteristics and in...
This study investigates the use of ChatGPT and DeepSeek for portfolio optimization in the financial investment sector. Based on a sample of constituent stocks from the CSI 300 Index, the study compares the portfolio construction characteristics and investment performance of the two models, focusing on portfolio generation, asset weight allocation, mean-variance optimization, and prompt engineering.
The empirical results indicate that, during the in-sample period, DeepSeek exhibited relatively better performance in constructing a 10-stock portfolio, whereas ChatGPT showed better performance in the 20-stock portfolio. During the out-of-sample period, the 20-stock portfolio recommended by ChatGPT achieved a cumulative return exceeding that of the CSI 300 Index, while the portfolio generated by DeepSeek failed to outperform the benchmark.
In terms of prompt engineering, ChatGPT demonstrated stronger responsiveness to relatively ambiguous prompts and generated more favorable returns. By contrast, DeepSeek’s performance improved when the prompts were clear and contained both return and risk conditions.
These research results indicate that the effectiveness of generative artificial intelligence in portfolio optimization is not only dependent on the model itself, but also on portfolio size, prompt design, and the integration of mean-variance analysis. This study provides empirical evidence on the applicability of generative AI models to portfolio construction. The results also suggest that AI models should be used in combination with quantitative optimization methods rather than as standalone tools for portfolio selection.