Cainz: Using Vertex AI for demand forecasting to optimize sales-planning predictions
Successfully deployed the demand forecasting solution of Vertex AI across 209 stores, creating more accurate stock replenishment predictions
Through Cloud Run jobs, preprocessing time was reduced to 50 minutes across—regardless of the number of stores
Built and refined demand forecasting architecture twice with the support of TAP
Trained large amounts of data efficiently, improving overall development speed
Cainz cuts sales data preprocessing time across 209 stores
Japanese home improvement retailer Cainz improved its demand forecasting architecture with Vertex AI and the Google Cloud Tech Acceleration Program (TAP), which reduced data preprocessing time to 50 minutes regardless of the number of stores.
Japanese home improvement retailer Cainz improved its demand forecasting architecture with Vertex AI and the Google Cloud Tech Acceleration Program (TAP), which reduced data preprocessing time to 50 minutes regardless of the number of stores.
In 2022, Japan’s home improvement center industry was worth a sizable 3.342 trillion yen (US$22.01 billion). Cainz, a leading DIY home improvement company in the country, had struggled with accurate demand forecasting for years.
“The aim of AI-powered demand forecasting is to study sales patterns and optimize the ordering process per store. Sales predictions had previously been done by our experienced in-house distributors, but there was a significant difference in accuracy between experienced and new distributors,” says Michihiko Yaguchi, Demand Forecasting Group, Business Solution Unit, Information System Division at Cainz.
Due to multiple product categories that Cainz maintained, Yaguchi explains that they adopted a fixed-order-quantity system.
“Whenever inventory levels fell below a certain point, replenishment orders were automatically issued. The problem was, this demand forecast system used the moving average of the past few weeks. We could not accurately predict the demand for seasonal products or short-term sales trends,” he adds.
We previously used a fixed-order-quantity system. Whenever inventory levels fell below a certain point, replenishment orders were automatically issued. However, we could not accurately predict the demand for seasonal products or short-term sales trends.
Michihiko Yaguchi
Demand Forecasting Group, Business Solution Unit, Information System Division at Cainz
Exploring new innovations with Vertex AI
In 2021, the company launched an AI-powered project to improve its retail demand predictions. Today, Cainz has successfully built a demand forecasting model through Google Cloud’s Vertex AI and Google Cloud Run jobs.
“We initially used a third-party solution to train the model. But as we expanded the scope of demand prediction, we saw its limitations and replaced it with Vertex AI Forecast,” shares Yaguchi.
“Vertex AI Forecast can handle large training data, is capable of multi-horizon predictions across multiple variables, and has a superior algorithm. I also like the fact that I can define the system’s training time. Overall, it’s a user-friendly tool for a small team like ours.”
Today, the home improvement company uses sales data to train its demand prediction model once a week. The results are then fed back into Cainz’s core system.
Yaguchi adds: “In demand prediction, explanations for results are crucial. Vertex AI Forecast has a feature called Vertex Explainable AI that streamlines this process and saves time by identifying the most effective features and explanatory variables for our system.”
We initially used a third-party solution to train the model. But as we expanded the scope of demand prediction, we saw its limitations and replaced it with Vertex AI. It can handle large training data, is capable of multi-horizon predictions across multiple variables, and has a superior algorithm.
Michihiko Yaguchi
Demand Forecasting Group, Business Solution Unit, Information System Division at Cainz
When the AI-powered demand forecasting project went live in June 2022, Cainz limited the solution’s scope to several products and categories at its flagship store in Isesaki, Japan. The scope has then expanded to all categories of the applicable products in November 2022, and to all of the products in 6 stores in January, 2023. As of August 2024, the demand forecasting AI is implemented in 209 stores.
Building the optimal architecture through TAP
Yaguchi adds that Cainz really only achieved this feat with the help of the Google Cloud Tech Acceleration Program (TAP).
Through the program, the Japanese retailer was given access to Google Cloud’s in-house engineers, who helped to design a demand forecasting model for Cainz’s needs.
What we learned from TAP was the ability to refine our architecture to make it increasingly efficient. By using Cloud Run jobs, we were able to preprocess data in about 50 minutes, regardless of the number of stores.
Michihiko Yaguchi
Demand Forecasting Group, Business Solution Unit, Information System Division at Cainz
“We leveraged TAP twice. The first time was in December 2021, when our project had just started. We had already decided to use Google Cloud for the execution platform, but since our team had limited knowledge of it, we consulted the TAP team. The program lasted three days, from hearing to prototyping, but the architecture we built at that time was so well-made that it was in operation until early 2023. It was a great learning experience,” says Yaguchi.
At the time, the scope of the demand forecasting AI had expanded to forecasting of all categories for original products and store-by-store and product-by-product forecasting for six stores, and the plan was to expand this to 50 stores.
“A sharp increase in the number of stores would lead to a proportional growth in the volume of data handled.
Back then, it took three hours to preprocess data from six stores. A simple calculation showed that this would never be completed within a day,” says Yaguchi.
To tackle this, Cainz consulted TAP for an architecture that could process large volumes of data across its 200 stores. The TAP team proposed Cloud Run jobs, a new feature that allows Cloud Run to be executed in parallel.
Instead of performing preprocessing for each of Cainz’s six stores sequentially, the system uses parallel execution.
“What we learned from TAP was the ability to refine our architecture to make it increasingly efficient,” shares Yaguchi. “By using Cloud Run jobs, we were able to preprocess data in about 50 minutes, regardless of the number of stores.”
Demand forecasting AI as the basis for retail planning
With Cainz’s improved demand forecasting architecture, Yaguchi says the future is still rife with possibilities.
“For retailers like us, demand forecasting AI can be the foundation for planning. By improving its accuracy, we can optimize for inventory management, logistics, and product-planning. I hope to also challenge the optimization of supply chain management based on this demand forecasting.”
He adds, “I look forward to more Google Cloud innovations and services. I’m also interested in exploring Gemini Code Assist, a generative AI solution with real-time code suggestions. I hope to use this in future to make development more efficient.”
Cainz is a leading DIY home improvement company in Japan. The Company retails kitchen products, small appliances, bathroom products, drugs, apparels, furniture, gardening products, agricultural materials, pet supplies, construction materials, and other products.
Industry: Retail
Location: Japan
Products: Google Cloud, BigQuery, Cloud Run, Cloud Storage, Gemini Code Assist, Vertex AI, Workflows
