Discounting Window-aggregated ARIMA model for time series forecasting

This project proposes a new model for time series forecasting, namely, Discounting Window-aggregated ARIMA model or the DWA model which was developed from the MASA model proposed by Paisit Khanarsa in 2017. The difference between the DWA model and the MASA model is the method of computing cumulative discount to calculate a ratio of data of each window. The DWA model is applied using the aggregate technique to reduce sway in the time series similar to the MASA model by dividing the sequence of data into groups called a window, then constructing a new time series resulting from the sequence of aggregate data in each window. This project examined 2 different data sets. In the algorithm process, the data was split into 2 sets. The first one is in-sample data for training and building the DWA model and the rest is out-sample data for evaluating the model. After the data is divided, the in-sample data is analyzed by using the method of Box-Jenkins to build an ARIMA model for forecasting, by selecting the best model with the lowest AIC value and using the forecast values obtained from the best fit to disaggregate the forecasted aggregate group value into individual future values using ratio from in-sample data and compare the model's performance of the DWA model, the ARIMA model, and the MASA model by using symmetric mean absolute percentage error (SMAPE). The result of this project revealed that SMAPE of the DWA model is better than the ARIMA model and the MASA model for the highly frequency time series data having a complex pattern.