Intermittent Demand Forecast and Inventory Reduction Using Bayesian ARIMA Approach

Natural calamities (e.g., hurricane, excessive ice-fall) may often impede the inventory replenishment during the peak sale season. Due to the extreme situations, sales may not occur and demand may not be recorded. This study focuses on forecasting of intermittent seasonal demand by taking random demand with a proportion of zero values in the peak sale season. Demand pattern for a regular time is identified using the seasonal ARIMA (S-ARIMA) model. The study proposes a Bayesian procedure to the ARIMA (BS-ARIMA) model to forecast the peak season demand which uses a dummy variable to account for the past years intermittent demand. To capture uncertainty in the B-ARIMA model, the non-informative prior distributions are assumed for each parameter. Bayesian updating is performed by Markov Chain Monte Carlo simulation through the Gibbs sampler algorithm. A dynamic programming algorithm under periodic review inventory policy is applied to derive the inventory costs. The model is tested using partial demand of seasonal apparel product in the US during 1996-05, collected from the US Census Bureau. Results showed that, for intermittent seasonal demand forecast, the BS-ARIMA model performs better and minimizes inventory costs than do S-ARIMA and modified Holt-Winters exponential smoothing method.