Forecasting material demand : Top-down hierarchical approach to scarce data
Kilpeläinen, Ossi (2024)
2024
Tuotantotalouden DI-ohjelma - Master's Programme in Industrial Engineering and Management
Johtamisen ja talouden tiedekunta - Faculty of Management and Business
This publication is copyrighted. You may download, display and print it for Your own personal use. Commercial use is prohibited.
Hyväksymispäivämäärä
2024-03-15
Julkaisun pysyvä osoite on
https://urn.fi/URN:NBN:fi:tuni-202403132823
https://urn.fi/URN:NBN:fi:tuni-202403132823
Tiivistelmä
Inventory demand forecasting is a crucial tool particularly for procurement aiming both to maximize customer satisfaction by ensuring the needed stock level while keeping said stock level as low as possible to minimize the costs associated with it. Inventory demand forecasting is characterized by having many forecastable time series, i.e. different materials in the warehouse, that are usually somehow grouped together. Often, some of these materials also exhibit demand patterns where the demand happens rarely, thus resulting in many zero observations in data. Hierarchical forecasting is one of the tools to overcome said problems, resulting in more easily forecastable time series and fewer number of zero observations.
This thesis focuses on applying a top-down hierarchical forecasting approach with autoregressive-based models. A literature review and an empirical part applying an approved Box-Jenkins methodological framework are placed to achieve the following goals of this thesis:
i) Identify possible models to predict future demand in a hierarchic inventory demand environment when the data available is limited.
ii) Predict the future component demand utilizing the models and tools identified before.
The hierarchy used in this study represents the distribution of materials to the product families consisting of multiple products. Autoregressive-based models are developed to predict the demand of said product families in a top-down hierarchical forecasting approach for the problem. The predictions from the product family level forecasts are then disaggregated to the bottom level of the hierarchy, i.e. material demand, by studying the effect of multiple windows of different moving averages.
The results of the product family level forecasts beat the naïve benchmark used in the study for all time series. However, another benchmark, a 4-step moving average, performs as well as the developed model for some of the series. Thus, the developed autoregressive model might not be ideal for all series, even though it does well against the naïve benchmark. After disaggregating the forecasts to material demand level, the developed approach results in a forecasting performance that beats the naïve benchmark in 96 % of the time series, based on RMSE-%. The forecasting performance level is found to be unrelated to the number of observations in the original training data. The results are in line with the literature review done in this study, indicating that hierarchical forecasting can be a valuable tool for solving problems with many grouped time series, even with scarce data.
This thesis focuses on applying a top-down hierarchical forecasting approach with autoregressive-based models. A literature review and an empirical part applying an approved Box-Jenkins methodological framework are placed to achieve the following goals of this thesis:
i) Identify possible models to predict future demand in a hierarchic inventory demand environment when the data available is limited.
ii) Predict the future component demand utilizing the models and tools identified before.
The hierarchy used in this study represents the distribution of materials to the product families consisting of multiple products. Autoregressive-based models are developed to predict the demand of said product families in a top-down hierarchical forecasting approach for the problem. The predictions from the product family level forecasts are then disaggregated to the bottom level of the hierarchy, i.e. material demand, by studying the effect of multiple windows of different moving averages.
The results of the product family level forecasts beat the naïve benchmark used in the study for all time series. However, another benchmark, a 4-step moving average, performs as well as the developed model for some of the series. Thus, the developed autoregressive model might not be ideal for all series, even though it does well against the naïve benchmark. After disaggregating the forecasts to material demand level, the developed approach results in a forecasting performance that beats the naïve benchmark in 96 % of the time series, based on RMSE-%. The forecasting performance level is found to be unrelated to the number of observations in the original training data. The results are in line with the literature review done in this study, indicating that hierarchical forecasting can be a valuable tool for solving problems with many grouped time series, even with scarce data.