Multivariate Time Series for Customs Revenue Forecasting Using LSTM Neural Networks

In the age of data, governments can use AI to achieve goals. One of the objectives is to find an appropriate way to accurately forecast international trade tax revenues, especially import duties. This is very important for state budget planning and achieving Key Performance Indicator target. Because if the planning misses, then the customs authority will increase the risk of failure in meeting the target. This research uses multivariate time series problem approach to produce a forecasting model with decomposition and recurrent neural network Long Short-Term Memory (LSTM). Our model employs a supervised mechanism and is trained on historical data with the seasonal component removed, resulting in a robust model that can be applied to various customs administrations. We have evaluated our proposed model on the real-world data. Results show promising performance, the LSTM Decomposition with the lowest RMSE value of 37,720 outperformed both the Seasonal ARIMA with RMSE 61,232 and the Holt-Winters Model with RMSE 47,703. When compared to LSTM without decomposition, our proposed model achieves a 17% improvement and a 38.4% reduction in error when compared to the seasonal ARIMA.