Inflation forecasting using hybrid ARIMA-LSTM model

Abstract

Prediction of time series is one of the most demanding research areas due to the nature of various time series i.e., stocks, inflation, stock indexes etc. Various methods have been used in the past to forecast such time series, however, Machine Learning (ML) methods have been suggested in the academic literature as alternatives to statistical ones for time series forecasting. Yet, scant evidences are available about their relative performance in order of their accuracies and computational requirements. In this thesis, a hybrid model consisting of ARIMA and LSTM is proposed and compared with individual models ARIMA, LSTM, and PROPHET for inflation forecasting. Two Scale-dependent metrics namely mean absolute error (MAE) and root mean square error (RMSE), one Percentage-error metric, mean absolute percentage error (MAPE) and coefficient of determination (R 2 ) are used to evaluate the variance between dependent and independent parameters for inflation forecasting in developed and developing countries. Consumer Price Index (CPI) data is collected monthly to reflect the effect of price inflation at consumer level. Most of the central banks depend on inflation forecast to inform their respective monetary policy makers and to enhance the efficacy of monetary policy. The publicly available CPI data is presented for analysis and evaluation of price inflation effects on developed and developing countries. For this research work, six developed countries (Canada, United States, Australia, Norway, Poland and Switzerland) and six developing countries (Colombia, Indonesia, Brazil, South Africa, India and Mexico) with different durations are targeted to evaluate the performances of proposed machine learning model and the individual models to forecast inflation (CPI). The proposed HYBRID model with one-step ahead forecasting outperformed every other model for forecasting inflation (CPI) of developed and developing countries regardless of duration. The best performance was observed by taking 90% training data and 10% testing data. All iv forecasting models performed better on data of six developed countries with overall average errors of 1.023796 in MAE, 0.009648 in MAPE and 1.222454 in RMSE when taking 10% as test data. While in the case of developing countries overall average errors of MAE, MAPE and RMSE was 1.361308, 0.011847, and 1.562288 respectively. Also, in the case of 20% and 30% test data, the performance of all models on developed countries data was better than developing countries in terms of least errors in MAE, MAPE and RMSE.