Peak demand forecasting for a seasonal product using Bayesian approach

An actual demand‐forecasting problem of the US apparel dealers is studied. Demand is highly fluctuating during the peak sale season and low prior to the peak season. The model is described by the continuous time stochastic process applying the Bayesian process. The standard gamma distribution is selected for the demand process and an inverse gamma distribution is chosen as the conjugate prior for the model. The choice is supported by the maximum likelihood estimate among a number of non‐negative distribution models. The proposed Bayesian models predict the probability of the future demand expressed explicitly conditional on the observed demand prior to the peak season. The data set illustrates partial demand of a seasonal product procured by the US dealers from overseas. In recent years, hazard and operational risks due to weather disasters and equipment shutdowns were felt significantly. These caused supply chain disruption and unrecorded demand. The model is extended to contribute to forecast from an unrecorded data set due to supply disruption. Forecasts are compared with real data and a widely implemented adaptive Holt‐Winters (H‐W) seasonal forecasting model. Results show that the forecasts calculated by the proposed methods do better than those of the adaptive H‐W model.