Indonesian Inflation Forecasting with Recurrent Neural Network Long Short-Term Memory (RNN-LSTM)

Abstract

Listen

Translate article

This study forecasted inflation in Indonesia using the Recurrent Neural Network Long Short-Term Memory (RNN-LSTM) model, ideal for nonlinear, complex time series data. It evaluated the effects of different activation functions, such as Logistic, Gompertz, and Hyperbolic Tangent (tanh); and weight update methods, such as Stochastic Gradient Descent (SGD) and Adaptive Gradient (AdaGrad) on RNN-LSTM performance. Monthly inflation data from January 2005 to December 2023 underwent preprocessing, including normalization and autoregressive lag-based input selection. Model accuracy was assessed with Root Mean Squared Error (RMSE) and Symmetric Mean Absolute Percentage Error (SMAPE). The findings indicated that the RNN-LSTM model with the logistic activation function and SGD optimization achieved the highest accuracy, outperforming traditional models such as Exponential Smoothing (ETS), Autoregressive Integrated Moving Average (ARIMA), Feedforward Neural Network (FFNN), and Recurrent Neural Network (RNN). Additionally, optimal learning rate and epoch values were identified, enhancing model stability and precision. In conclusion, the study confirms that the RNN-LSTM model is effective for inflation forecasting when optimized with specific activation functions and optimization methods. It recommends further exploration of neuron configurations and alternative models, such as the Gated Recurrent Unit (GRU), to improve forecast accuracy.