A deep learning framework for automated transaction classification and budget prediction in personal financial management

With the popularization of mobile payment, personal financial data has experienced explosive growth, and traditional manual recording methods can no longer meet the needs of intelligent financial management. Existing research faces challenges in processing Chinese financial texts, including insufficient semantic understanding, inadequate capture of temporal patterns, and independent processing of classification and prediction tasks. This study proposes a Deep Learning-based Financial Tracking System (DFTS) that innovatively combines the Transformer multi-head attention mechanism with Bidirectional Long Short-Term Memory (BiLSTM) network and designs a Temporal Feature Fusion Module (TFFM) to achieve collaborative optimization of transaction classification and budget prediction through a multi-task learning framework. The DFTS model employs a pre-trained Chinese BERT model to extract semantic features, introduces an amount-weighted attention mechanism to enhance perception of high-amount transactions, and captures global dependencies and multi-scale local patterns through the complementary collaboration of BiLSTM and Temporal Convolutional Network (TCN). On a self-collected Chinese dataset containing 52,847 real transaction records, the DFTS model achieves 96.3% accuracy and 95.8% F1-score in transaction classification tasks, improving by 2.2 and 2.3 percentage points compared to the strongest baseline (BERT+LSTM, achieving 94.1% accuracy and 93.5% F1-score); in budget prediction tasks, the Mean Absolute Percentage Error (MAPE) is only 6.2%, reducing by 8.6 percentage points compared to traditional ARIMA models and 3.6 percentage points compared to LSTM networks. Ablation experiments validate the effectiveness of core components including amount-weighted attention, BiLSTM, TCN, Focal Loss, and multi-task learning, while attention visualization analysis enhances model interpretability. Supplementary classification experiments on a publicly available multilingual financial transaction benchmark further confirm that the core DFTS architecture generalises effectively across language and dataset boundaries. This research provides a new technical solution for developing intelligent personal financial management systems and has significant application value in the financial technology field.