Fuzzy Nearest Neighbour Method for Time-series Forecasting

This paper explores a nearest neighbour pattern recognition method for time-series forecasting. A nearest neighbour method (FNNM) based on fuzzy membership values is developed. The main aim of the forecasting algorithm is to make single point forecasts into the future on the basis of past nearest neighbours. The nearest neighbours are selected using a membership threshold value. The results include the mean absolute percentage error and the direction error for sales data of three real control products. The forecasts are compared to the actual values over a period of twenty months taking four years of monthly data in the estimation period. The results are very encouraging for further work on the development of fuzzy nearest neighbour methods in forecasting.