Browsing by Author "Tiainen, Matias"
Now showing 1 - 2 of 2
Results Per Page
Sort Options
Item Forecasting Seasonal Demand at the Product Level in Grocery Retail(2021-05-17) Tiainen, Matias; Latva-Pukkila, Niilo; Perustieteiden korkeakoulu; Kivelä, MikkoThe grocery retail industry is highly competitive with razor-thin margins. Retailers are inclined to cut their costs in any way possible to remain competitive. Forecasting sales and product demand precisely have profound impacts, and it holds high strategic importance for retailers. It helps retailers to decrease waste, optimize order sizes from suppliers and the required workforce. This study examines whether advanced machine learning methods such as random forest, recurrent neural network, or Bayesian hierarchical modeling can outperform a simple linear regression model at forecasting seasonal demand. The study aims to find out if product-level predictions yield better results than product-group-level predictions. Further, this study demonstrates the challenges of predicting over 300 000 unique products' seasonality and how well different evaluation metrics define and justify the results. The study results indicate that an increase in the level of complexity of models does not always guarantee superior results. The subtleties such as inspection level, clean point of sales data, and well-chosen model for a problem play a more critical role in the success. The random forest is the best-performing model at the product level. This study also revealed that product group-level forecasts are more accurate than product-level forecasts. Obtained results imply that predicting the seasonal demand of over 300 000 unique products creates many corner cases to consider. Understanding them requires an extensive dive into the data. Other concerns lie in data quality and examining the importance of features. Traditional evaluation metrics such as mean absolute percentage error and Spearman rank correlation coefficient do not fit the purpose of predicting the seasonal demand, especially if the objective is not clearly defined. However, they are an excellent proxy to infer the performance between individual models. This study suggests companies to craft domain-specific features and define clear goals of the seasonal demand to find proper evaluation metrics. It also recommends that they tackle the imbalance between predictions of best- and worst-selling products to remain competitive and fulfill customer satisfaction.Item Luokittelumenetelmiä ennustettaessa NHL-ottelun voittajaa(2018-12-18) Tiainen, Matias; Saramäki, Jari; Sähkötekniikan korkeakoulu; Turunen, Markus
