A linear programming-based bi-objective optimization for forecasting short univariate time series

This paper proposes a Linear Programming-based Bi-Objective time series Forecasting Algorithm that helps forecast sub-annual short univariate time-series data. The proposed algorithm generates forecasts optimized for a pair of accuracy measures instead of just one. The algorithm is based on the ϵ-constraint-based multi-objective optimization method. The accuracy measure pairs used in this paper are the Mean and Maximum of Absolute Errors and Absolute Percentage Errors. We compare the performance of the proposed algorithm with several industry-standard forecasting methods using accuracy measures commonly reported in the literature for three forecast horizons: long-term, medium-term, and short-term. The proposed algorithm performs the best in the long- and medium-term horizon for the short-time series studied in our paper. Across all forecast horizons, the proposed algorithm has the least maximum errors, reducing the over-and under-forecast errors. The proposed algorithm can yield interpretable linear forecasting models and is quite flexible.