With the ever-decreasing data acquisition and storage cost, more businesses are turning to data-driven decision-making. Forecasting is a primary precursor to business process planning and is thus essential for informed data-driven business decision-making. Forecasting is used to leverage the time lag between the awareness of an upcoming event and its occurrence for effective planning and decision. Effective business process planning and decision-making maximize revenue and minimize cost by efficiently utilizing the limited resources at the disposal of the business. This proposed research aims to develop algorithms based on multi-objective approaches that can provide more accurate forecasts than existing methods. The research objective is two-fold: (i) to develop a Linear Programming based bi-objective forecast generation algorithm, and (ii) to develop a bi-objective heuristic for forecast pooling.

One of the challenges in time series forecasting is to generate longer horizon forecasts that perform well on multiple accuracy measures using training data of short length (short time series data). In the first part of our proposed research, we aim to develop a Linear Programming-based time series Forecasting Algorithm – considering Multiple Accuracy Measures (LPFA-MAM) that can be used to forecast short time-series data. The proposed algorithm generates forecasts that are optimized for a pair of accuracy measures instead of just one measure, such as the Sum of Squared Errors. The algorithm is based on the ϵ - constraint-based multi-objective optimization method.

Forecast Combination is a widely accepted strategy to improve forecast performance by hedging against model uncertainties that arise from forecast selection. Over the last few decades, several forecast weighting schemes have been proposed and used in the literature for forecast combinations. The forecast weighting schemes generally assign lesser weights to poorer forecasts and more weights to better forecasts. When many base forecasts are available for combination, relying on forecast weighting schemes does not reduce the impact of poorer forecasts, as they can still add up with smaller weights. Forecast pooling is considered an effective alternative to forecast selection and forecast combination. Forecast combination with only a subset of all available forecasts is forecast pooling. Research literature attributes the reduction in errors after forecast combination to sufficient diversity in the forecast pool. In this proposal, we establish the need for an improved forecast combination heuristic that automatically identifies forecast pools that balance forecast performance and diversity.