Bento is widely known as a part of Japanese food culture. There are many companies in Japan that produce and sell thousands of boxed lunches every day. Bento companies have multiple menus. In this study, we focus on the catering lunchbox industry, which handles everything from the production of lunchboxes to their delivery.

The catering lunchbox industry needs a demand forecasting model that accurately predicts. Catering companies produce and sell boxed lunches every day. The sales data of catered lunches are time-series data. The number of boxed lunches sold fluctuates daily depending on customer schedules and the day of the week. It is difficult to accurately forecast the demand for catered lunches. On the other hand, if demand forecasting is not done accurately, a large amount of food loss will occur. Creating a demand forecasting model that accurately predicts can reduce food loss.

In a study of demand forecasting for boxed lunches, the model's forecasting accuracy becomes worse during periods of rapid changes in the number of boxed lunches sold. The machine learning model used in previous studies cannot respond unless there have been similar rapid changes in the number of boxed lunches sold in the past. This study aims to improve the accuracy of forecasts during periods of rapid changes in the number of boxed lunches sold.

The phenomenon in which the data distribution changes over time is commonly referred to as concept drift. Concept drift has been extensively studied. Four types of concept drift exist. This study analyzes two of the four types. The first type is "A new concept occurs within a short time". The second type is "An old concept may reoccur after some time".

This study establishes criteria for detecting sudden changes in the number of boxed lunches sold. In addition, we propose a demand forecasting model with high forecasting accuracy even after changes. Two criteria were established for detecting sudden changes in the number of lunches sold. The first criterion is "Number of consecutive days of rapid changes in the number of boxed lunches sold". The second criterion is "Number of types of lunchboxes”. The post-change demand forecasting model emphasizes the most recent information to make forecasts. The most recent information available is important to respond quickly to rapid changes.

Numerical experiments using real data showed that the proposed model improved prediction accuracy compared to the conventional model. The usefulness of the proposed model was demonstrated. The usefulness of the proposed criteria for detecting sudden changes in the number of boxed lunches sold was also demonstrated.

Yasunori Iwata

Affiliation

Conference Tracks

Refine Results

Predicting Demand for Catering Lunchboxes Using Machine Learning to Respond to Rapid Changes in Bento Sales