Swin Transformer and dual-layer routing attention for enhanced financial accounting prediction

Abstract

With the rapid advancement of big data and artificial intelligence technologies, financial accounting is increasingly evolving toward greater intelligence and automation. However, the limited capacity of traditional methods to process complex data makes it challenging to address the rapid changes in enterprise dynamics within big data environments. To enhance forecasting accuracy and efficiency, this study proposes an intelligent financial accounting prediction model based on the Swin Transformer and a dual-layer routing attention mechanism. Experimental results demonstrate that the proposed model achieves substantial improvements in prediction performance compared with traditional Autoregressive Integrated Moving Average (ARIMA) and Long Short-Term Memory (LSTM) models. On the A-share dataset, the model attained a mean absolute error (MAE) of 1.47, representing improvements of 37.7% and 22.2% over ARIMA (2.36) and LSTM (1.89), respectively. Similarly, on the European and American enterprise dataset, the model achieved a root mean square error (RMSE) of 2.15, corresponding to reductions of 28.4% and 16.5% compared to ARIMA and LSTM. Furthermore, the model improved forecasting efficiency by reducing quarterly data processing time by approximately 15%. These findings highlight the potential of combining Transformer-based architectures with attention mechanisms for intelligent financial forecasting and underscore the broad application prospects of big data analytics in the financial domain.