Super Hosokawa trial: tofu and fried food waste at 0.20% vs 0.52% overall, using Imamura Shoji forecasts

In financial year 2021 Japan wasted 5.23 million tons of food, and business-related waste accounted for over half of that figure. Food logistics is split into manufacturing, wholesale and retail, each with its own demand forecasting data, which causes the bullwhip effect. Super Hosokawa's CEO wanted to cut waste while still uncovering latent demand, since driving waste loss to zero would steadily reduce sales volumes.

Seven companies took part in a government-commissioned trial in Oita and Fukuoka prefectures, in which Imamura Shoji built a demand forecasting system and Super Hosokawa supplied stores and ID-POS data. Using Azure Blob Storage, Snowflake and Azure Databricks, Imamura created a 60-pattern demand forecasting model that returns forecasts two days ahead, and used Azure OpenAI Service to create customer personas. The trial began in January 2024 in the tofu, fish paste and fried food sections of three Super Hosokawa stores, where orders to manufacturers were placed based on the forecasts.

In the January 2024 trial at three stores, orders placed by humans averaged 249.8 forecast errors a day, against 194.1 for the demand forecast model alone and 181.2 for humans using the model. The waste rate of tofu and fried products was 0.20 percent when ordered by combining forecasts with human ordering, against 0.52 percent overall, and fish paste products were 0.13 percent against 0.52 percent overall. The page says forecasts that account for latent demand also boosted sales.

• Combining demand forecasts with human ordering produced the fewest errors overall, so keep store staff in the loop while the model matures.
• Start with items that are delivered daily and expire quickly, since they are the most prone to waste.
• Share forecasts across manufacturers, wholesalers and retailers to dampen the bullwhip effect instead of forecasting in isolation.