A Multi-Objective Optimal Forecast Combination for Predicting Intermittent Demand

Demand forecasting methods usually adopt the approach of minimizing certain forms of error or maximizing a likelihood function. However, for an end user, such objective functions are often the means to achieve optimum results over various key performance indicators in respective domains. A production manager will likely weigh a high product fill rate and low average inventory resulting from a given forecast more than a low absolute error. Especially in the case of intermittent demand forecasting, error minimizing often results in poor values of such key performance indicators. We propose an algorithm that considers the user’s preferences across multiple objectives for generating an optimized combination of different forecasting methods. Since every user may perceive the relative importance of each objective differently, the proposed algorithm interacts with the user and computes a set of optimal weights assignable to the n distinct objectives. Using these optimal weights, we define a metric called the Multi-Objective Value, which is subsequently maximized to obtain the Adaptive Multi-objective Optimal Combination (AMOC) of m diverse forecasting methods. We demonstrate its application with (m, n) = (5, 4) on an intermittent demand time series for four users over two scenarios: single-period and multi-period inventory models. Additional experiments are included to exhibit the robustness and optimality of the AMOC