基于预训练模型的文本-时序双模态股票收益率联合预测

打开文本图片集

中图分类号：TP18；F831.5 文献标识码：A 文章编号：2096-4706（2025）16-0168-06

Abstract：Stock price fluctuations exhibitrandomnessandcomplexitymaking ita challenging problem in financial markets toimprovetheaccuracyofstock pricepredictions.This paperproposesadual-modaljointpredictionmethodframework based on pre-trained models，which includes twocomponents ofLarge Language Models （LLMs）for predicting stock price movement intervals andtime-series prediction models for predicting dailyreturns.The final predicted returns are obtained by integrating theresultsof these two predictions.GPT-4and Chronos are employedasthe pre-trained LLMand time-series predictionmodel，respectively，inthejointfrmework.Comparedtosingle-modaltime-seriespredictiomodels，teroosed modeldemonstratessuperiorperformanceacrosmultipleevaluationmetrics，achieving higheraccuracyinstock priceprediction. This research approach notonlyreduces the dependenceonlarge volumesoftraining databut alsoenances textcomprehension capabities.Moreover，itseffctivenesscanbefurtherimprovedthroughfine-tuning，providingneisightsandmethodologies for stock return prediction.

Keywords：Large Language Model; stock market prediction; pre-trained time series prediction model; accuracy

0 引言

股价是市场对公司价值的即时评估，精确的股价预测对于投资者评估市场风险和监管机构分析宏观经济政策具有重要意义。（剩余9206字）