HYBRID METHOD SARIMAX-LSTM PERFORMANCE ANALYSIS FOR FORECASTING ON CROSS-SECTIONAL HIERARCHICAL TIME SERIES DATA (CASE STUDY: SALES)

• Main Author: Yasmin Sumardi, Annisa
• Format: Theses
• Language: Indonesia
• Online Access: https://digilib.itb.ac.id/gdl/view/86177
• Institution: Institut Teknologi Bandung

The advancements in technology have significantly facilitated various industrial sectors in accomplishing their tasks and aiding in critical decision-making for the future of companies, including those in the Food and Beverage (F&B) industry. In upper management, sales forecasting is essential as a foundation for making important decisions. This can be achieved through forecasting techniques, with SARIMA being one of the most effective models. However, despite its current performance, SARIMA has limitations: it cannot incorporate exogenous data into its predictions and fails to account for non-linear patterns in the data. Additionally, conventional forecasting processes typically focus on a single sequence type, whereas, in reality, multiple interconnected sequences often form a hierarchical structure. To address these issues, this study proposes a hybrid method, SARIMAX- LSTM. SARIMAX is an extension of SARIMA that can include exogenous data in its predictions, while LSTM handles the non-linear patterns of the data. This hybrid method is then applied to hierarchically structured data with a reconciliation process, which aligns the forecasting results within the hierarchical data structure. The findings demonstrate the improved performance of the hybrid SARIMAX-LSTM method on hierarchically structured data after reconciliation, compared to use SARIMA and SARIMAX alone. The hybrid method achieved a performance value of 10.2% (MAPE), representing a 4.0% improvement over the SARIMA model and 3.7% improvement over the SARIMAX model, using sales data from Brand B which is engaged in F&B.