Peramalan Inflasi Umum Indonesia Dengan Metode Ensemble LSTM (Long Short Term Memory) Bagging

Kondisi inflasi suatu negara mampu memberikan pengaruh terhadap ekonomi suatu negara serta pengambilan kebijakan moneter. Karena inflasi memengaruhi kebijakan moneter maka dalam membantu pengambilan keputusan tersebut dapat dilakukan peramalan terhadap inflasi. Peramalan variabel yang bersifat makroekonomi sebagaimana inflasi dimana proses inflasi memungkinkan berubah dari waktu ke waktu mengakibatkan model neural network akan memberikan peramalan inflasi yang lebih akurat. Metode LSTM (long short term memory) yang merupakan bagian dari neural network adalah jaringan pengembangan dari recurrent neural network. Model peramalan juga dapat mengalami bias yang tidak diketahui (idiosyncratic biases) dan menggabungkan peramalan (metode ensemble) dapat membantu mampu meningkatkan kinerja dari metode peramalan. Dalam penelitian ini menggunakan metode ensemble long short term memory bagging (LSTM-B) yang mengintegrasikan antara neural network LSTM dengan strategi pembelajaran ensemble bagging. Pada penggunaan metode neural network ini diperlukan variabel input yang berasal dari Lag AR yang signifikan Selanjutnya dibuat model FFNN, LSTM dan LSTM-B. pada model ARIMA, FFNN, LSTM dan LSTM-B yang terbentuk dilakukan peramalan outsampel 6, 12 dan 24 bulan dan dihitung akurasinya yang kemudian dibandingkan. Akurasi yang digunakan adalah symmetric Mean Absolute Percentage Error (sMAPE). Dari hasil penelitian didapatkan ensemble LSTM Bagging memiliki performa lebih bagus dalam peramalan outsampel inflasi 24 bulan daripada ketiga metode lainnya. Untuk peramalan outsampel 6 dan 12 bulan maka penggunaan metode FFNN dengan metode terbaik daripada ketiga metode lainnya. Dari hasil ini maka dapat disimpulkan peramalan menggunakan metode LSTM-B lebih bagus digunakan peramalan jangka panjang dari pada ketiga metode lainnya. Selain itu LSTM-B mampu meningkatkan akurasi dari hasil peramalan metode LSTM. ==================================================================================================== The inflation condition of a country is able to have an influence on a country's economy as well as monetary policy making. Because inflation affects monetary policy, it is possible to forecast inflation in helping to make these decisions. Forecasting macroeconomic variables such as inflation where the inflation process allows changing from time to time results in a neural network model that will provide more accurate inflation forecasts. The LSTM (long short term memory) method which is part of the neural network is the development network of the recurrent neural network. Forecasting models can also suffer from unknown biases (idiosyncratic biases) and combining forecasts (ensemble methods) can help improve the performance of forecasting methods. In this study, the Long Short Term bagging ensemble method (LSTM-B) integrates the LSTM neural network with the ensemble bagging learning strategy. In the use of this neural network method, it is necessary to have input variables originating from significant AR Lag. Furthermore, the FFNN, LSTM and LSTM-B models are made. In the ARIMA, FFNN, LSTM and LSTM-B models formed, outsample forecasting for 6, 12 and 24 months was calculated and the accuracy was calculated which was then compared. The accuracy used is symmetric Mean Absolute Percentage Error (sMAPE). From the results of the study, it was found that the LSTM Bagging ensemble had better performance in forecasting 24-month inflation outsamples than the other three methods. For forecasting out of sample 6 and 12 months, the FFNN method is the best method than the other three methods. From these results, it can be concluded that forecasting using the LSTM-B method is better for long-term forecasting than the other three methods. In addition, LSTM-B is able to increase the accuracy of the LSTM method forecasting results