Transaction prediction in the financial field plays a crucial role in investors’ risk decision-making, management, and capital allocation. However, the complexity and time dependence of capital market data result in low accuracy of transaction prediction. This article applies Long Short-Term Memory Network (LSTM) to the construction of a capital market transaction prediction model. Historical transaction data can be collected from multiple financial databases, extracted through feature engineering, and model performance evaluated through cross validation. The research results indicate that the LSTM model performs well in predicting future price trends. Its accuracy is significantly higher than traditional time series models, with the highest prediction accuracy reaching 97.5%, providing investors with valuable market insights.