Demand forecasting has been studied extensively because it serves as an input to other decision processes in an organisation. Imprecise forecasts can lead to stock-outs, lost-sales or overstocking, thus not meeting the service level targets. The case company, Scale AQ, experienced similar challenges with forecasting the demand of one of their crucial components. Scale AQ is a global supplier of technology and infrastructure for land- and sea-based aquaculture. The product of focus in this study is their best-selling circular sea-based fish farming cage, cage-P. The company faced component shortage, which is the brackets in the cage, especially during the peak production season, due to long replenishment lead-times involved. Consequently, reliable forecasts for atleast three-four quarters ahead were required.
To resolve the issue of limited component availability, we have chosen to improve the demand forecasts which will thereby reduce the uncertainty in demand and events of stock-outs. Since the demand of the component is dependent on the demand of the cage, we have chosen to perform demand forecasting for cage-P. Numerous demand forecasting methods, both qualitative and quantitative, have been researched since the past few decades, and the most popular and widely studied field was time-series forecasting. Thus, the main objective of this study is to investigate various time-series forecasting methods and choose a suitable method for forecasting the demand of cages. Various time-series forecasting models were identified using systematic literature review. Then the identified models that met the selection criteria were further shortlisted for quantitative modeling. Four traditional models: Seasonal naive, Holt-Winters (HW), State-space model (ETS) and seasonal autoregressive integrated moving average (SARIMA) and four advanced models: Prophet, Multilayer perceptron (MLP), Long Short Term Memory (LSTM) and Support vector regressor (SVR) were selected for the comparative forecasting analysis using short univariate time series data. Each model was optimized using grid search method where optimal parameters (or hyperparameters) were selected for each model configuration. The resulting model configuration was utilized to create multi-step ahead forecasts (for four quarters ahead) and was evaluated using two performance metrics, RMSE and R2.
The models were then compared against each other and against currently employed statistical forecasting model (seasonal naive). It was found that most of the traditional methods outperformed the advanced methods when dealing with short univariate times-series though LSTM was found to be the best overall performing model. It was identified that the forecasting performance of all the models, except SVR, surpassed seasonal naive model. The forecasting performance of LSTM model was found to be 51% better than seasonal naive. Whereas the SARIMA model (and its variants) resulted in an improvement of 29-48% compared to the seasonal naive model’s forecasting performance. Since traditional models performed better than advanced models on short time-series, it is recommended that Scale AQ employs SARIMA model for deriving statistical forecasts for cage-P and the other cages. It is also recommended that the statistical forecasts are complemented with managerial judgements since the domain knowledge of managers is vital in a complex environment.
This study contributes to both scientific community and the case company. With regards to scientific contribution, the study suggests an appropriate and optimized method for forecasting short univariate time-series. With regards to the case company, a suitable optimized model was recommended along with the implementation procedure to be used for the demand forecasting of cage-P as well as the other cages instead of the currently employed statistical forecasting model.
Keywords: demand forecasting, univariate time-series, traditional models, advanced models