Classical and Deep Learning Time Series Prediction Techniques in the Case of Indonesian Economic Growth

Abstract

Abstract Gross Domestic Growth (GDP) as a proxy of an economic growth in one country is defined as the market value of services and goods produced by the country for a year. As the GDP of one country gets higher and higher through the years, this indicator points out that in general the economic growth of the country is growing; specifically, the values of services and goods accumulated in the market are increasing. To determine what indicators of a country that affects one’s economic growth remains an open question. Therefore, this research attempts to study those indicators and particularly utilize them to predict the economic growth. To answer those questions, this research employs diverse time series techniques ranging from classic time series analysis to machine learning and deep learning. Subsequently, our dataset comprises World Development Indicators (WDI) of Indonesia from 1962 to 2016. By measuring Root-Mean-Square Error (RMSE), we show Seasonal Autoregressive Integrated Average (SARIMA) and Convolutional LSTM give the the best performance from classical and deep learning techniques respectively. Our analysis shows SARIMA’s performance is boosted by its ability to capture trend and seasonality of the dataset; equally important, Convolutional LSTM equipped with convolutions as part of reading input into LSTM units significantly boost the performance over either Convolutional or LSTM networks. Furthermore, our analysis points out that the indicator which most strongly contribute to predict GDP is CO₂ emission. This result agrees to the fact that some countries with high CO₂ emission also has high GDP as well; also this finding should warn Indonesian government of the increasing CO₂ pollution.