Traditional data analysis methods have limitations that result in many valuable tax insights being overlooked in routine tax data processing. However, with machine learning techniques, it is feasible to uncover insights that traditional methods fail to capture. Establishing a correlation between historical and future tax data has been a challenging topic, with no effective methods for resolution. Although various models, based on linear and non-linear data, exist for forecasting, they often produce significant errors when applied to tax data predictions. This paper addresses the non-linear characteristics and volatility of tax data, proposing an enhanced Long Short-Term Memory (LSTM) model. In contrast to the conventional LSTM model, this improved model boasts a higher prediction accuracy. The enhancement involves interpolating the input data and fusing the interpolated data back into the original dataset, aiming to augment the accuracy of the output. In our experimental phase, genuine tax data was used for forecasting, and the superiority of the enhanced LSTM model over the traditional one was visually demonstrated through charts. Upon predicting tax data for two companies and comparing the outcomes to actual scenarios, it was observed that the proposed enhanced LSTM model significantly reduced the Mean Absolute Error (MAE) and Root Mean Square Error (RMSE) by 13.65% and 14.49%, respectively, compared to the traditional LSTM model. This indicates the distinct advantage of the improved LSTM model in enhancing the accuracy of tax data predictions.