A comparison of forecasting methods for medical device demand using trend-based clustering scheme

Traditional time series forecasting model considers the aggregated demand as target. However, not all hospitals have the same trend. Different demand trends offset each other and distort potentially useful information. In this study, we first apply linear regression to transform the time series of medical device demand into a vector to discover different demand trends. With the case company data, we discover three types of demand trends. Then hierarchical clustering method is employed to cluster hospitals based on the vector of trend. Next, five forecasting models are applied to forecast future demand of each cluster. The results reveal that the proposed trend-based clustering approach is able to segment hospitals into groups with meaningful patterns. Furthermore, comparing different models of each cluster, we find that there is no general model can consistently produce the most accurate results. Time series with different trends have corresponding forecasting models which can generate accurate results.