ANALYZING THE IMPACT OF HISTORICAL DATA LENGTH IN NON SEASONAL ARIMA MODELS FORECASTING

Abstract

Different values of minimum data requirement for ARIMA models have been proposed. It also proposed to use as much data as they are available in formulating ARIMA models. This paper studied the impact of the size of the historical data on ARIMA models in forecasting accuracy. The study used 286 weekly records of amount of solid waste generated in Arusha City to formulate four ARIMA models using different data lengths or size. The first model, M1 used 30 observations, the second model, M2 used 60 observations, the third model M3 used 120 observations and the fourth model, M4 used 260 observations all of which are the most recent. A total of 26 observations were held out for validation. The precision in forecasting was tested using MAPE, RMSE and MAD.  The results indicated variation in precision. M3 performed best in one-week ahead and 9 – 12 weeks ahead while M4 did best in 2 – 8 weeks and also for 13 weeks  and above. M1 was the worst model in forecasting.

Keywords: ARIMA models, MAPE, RMSE, MAD, Forecasting