Enhancing Stock Market Information Retrieval with Retrieval-Augmented Generation

Enhancing Stock Market Information Retrieval with Retrieval-Augmented Generation. By Huan Phong Thai and five others, from FPT University. Published on the second of November.

Natural language processing is already widely used in finance, to analyze and retrieve information about the stock market automatically. But traditional systems built this way often struggle on two fronts. They have trouble maintaining contextual accuracy, and they have trouble generating responses that stay relevant, especially when the financial environment keeps changing. A system that answers questions about markets has to be both correct about the context and useful in what it says.

The authors turn to retrieval-augmented generation, an approach that combines retrieval-based methods with generative capabilities in order to improve the quality of responses. They present a RAG system designed specifically to answer questions and generate responses about stock market information.

The system integrates the two halves. A retrieval side brings in relevant material, and a generative side produces the text of the answer. The stated goal of joining them is to raise both the accuracy and the relevance of financial text generation. The abstract does not go further into how the components are built, so the description stops there.

To judge the output, the authors use the ROUGE family of metrics, specifically ROUGE-1, ROUGE-2, and ROUGE-L. These measure lexical and structural similarity between generated text and reference text. The experiments use a dataset of one hundred queries related to the stock market. On that set, the system achieves high ROUGE scores, which the authors read as strong alignment with the reference documents.

Speed was measured too. Latency was assessed to gauge how responsive the system is, and the average response time came out at one point one six seconds.

Put together, the claim is that this RAG model can give stock market answers that are both accurate and timely. The authors present it as a promising approach for financial information retrieval. The evidence offered rests on overlap with reference texts across one hundred queries, plus a response time just over a second, and that is the scale at which the conclusion is drawn.