Toshiba Develops Sales Forecasting AI Using Transaction Big Data from ‘SmartReceipt’

Kawasaki, Japan – Toshiba Corporation has developed a sales forecasting AI that uses transaction big data accumulated by the electronic receipt service ‘SmartReceipt’ ^*1 to realize highly accurate sales forecast for new product launches and in other scenarios. The technology is expected to be used to forecast product sales in a wide range of consumer goods sectors, including food and daily necessities.
In recent years, consumer preferences have become increasingly diverse, and market conditions have also become more volatile. Attempting to predict the purchasing behavior of individual consumer posed a challenge, as it required performing prediction calculations for each person, resulting in a significant increase in computational load.
The technology uses two AI systems to process big data and improve forecasting accuracy. The first uses clustering AI^*2, independently developed by Toshiba for transaction big data, to group consumers with similar purchasing behavior. Grouping that data makes it possible to reflect increasingly diverse consumer preferences in forecasts while limiting the computational load. The second uses generative AI to make highly accurate forecasts for ambiguous variables such as consumer preferences. Toshiba has improved forecasting accuracy with a newly designed “response score” and having generative AI forecast purchases for each group.
A further major feature of the technology is that it uses transaction big data^*3 from ‘SmartReceipt’, which has more than 3 million members^*4. Consumption trends change from day to day, affected by factors that include the weather, events, trends in competing products, and consumer sentiment. Combining the two AI systems and continually updating forecasts using transaction big data accumulated by ‘SmartReceipt’ over the long term realizes more up-to-date and accurate predictions than conventional methods while limiting the computational load (Figure).
An evaluation that Toshiba carried out using transaction big data from ‘SmartReceipt’ confirmed that the technology reduced the gap between forecast sales volumes and the actual sales volumes by approximately 23% compared with a conventional forecasting method^*5 when forecasting sales of products in a specific category^*6.
Details of the technology will be presented at the 40th Annual Conference of the Japanese Society for Artificial Intelligence (JSAI2026), which will be held from June 8 to 12, 2026. 

When launching a new product, it is important to gauge market reactions in advance. As consumer preferences become increasingly diverse and market conditions grow more volatile, conventional forecasting methods have had difficulty accurately reflecting purchases based on preferences, resulting in lower forecasting accuracy.
Against this backdrop, expectations are rising for the application of generative AI in sales forecasting. Sales forecasting using generative AI predicts sales volumes based on past transaction data. However, methods that process all data at once involve enormous amounts of data, so forecasts must be made by extracting only a portion of the data. As a result, the data used as the basis for forecasts tends to be biased toward frequently appearing data, making it difficult to fully reflect the finer details of purchasing behavior. The alternative methods that predict individual consumer purchasing behavior can reflect individual purchasing behavior in detail, but they create the issue of an enormous increase in the amount of computation in proportion to the number of people covered.

Toshiba focused on developing a technology that forecasts overall market sales volumes without requiring enormous computation. It did this by combining clustering AI and generative AI and grouping consumers with similar purchasing behavior.
The clustering AI automatically groups consumers with similar purchasing behavior, based on transaction big data from ‘SmartReceipt’. The generative AI then calculates a response score for each cluster, expressing the degree of interest in purchasing a new product as a score from 0 to 100. Sales volumes for the overall market are calculated by aggregating the response scores, with consideration for the number of people in each cluster. This approach realizes forecasts that reflect purchases based on increasingly diverse consumer preferences without predicting individual purchasing behavior. In addition, because processing is performed by cluster, the technology achieves forecasts within a practical amount of computation even when using big data. By calculating forecast values according to the target period, the technology can be used flexibly, including forecasting sales volumes for a specified period and for continuous forecasting over long periods.
In an evaluation experiment using transaction big data from ‘SmartReceipt’, Toshiba confirmed that the technology reduced the gap between forecast sales volumes and actual sales volumes for a specific category by approximately 23% compared with a conventional forecasting method. Conventional forecasting uses frequently occurring transaction data, so it is unable to sufficiently learn purchases based on preferences, and forecast sales volumes also tend to be close to the average value for that category. The new technology uses transaction big data accumulated by ‘SmartReceipt’ over the long term, and can capture diverse consumer preferences and market changes, thereby producing forecast results that are closer to actual sales volumes.

Going forward, Toshiba will continue validation while expanding the target categories and periods, and will broaden the scope of application. It will also use the technology to provide new insights in the transaction data analysis services it offers.

• ‘SmartReceipt’, an electronic receipt service developed and operated by Toshiba Tec Corporation with operational support from Toshiba Corporation, digitizes itemized receipts for purchased products, which are typically provided on paper at checkout, and manages and provides them as data through an electronic receipt center. Customers can check their purchase history on their smartphones at any time without keeping paper receipts, improving convenience when shopping, while helping participating stores reduce the cost of issuing paper receipts and reduce paper consumption.
  Official website: https://www.smartreceipt.jp/
• A technology that automatically classifies products with common consumer purchasing patterns (purchasing characteristics) into the same product groups (clusters) based on transaction big data.
  Analysis example: https://www.global.toshiba/jp/news/data-corp/2024/09/20240911.html
• Purchase data is used with the consent of ‘SmartReceipt’ users.
  Details: https://sr-mobile-apps2.smartreceipt.jp/srsw/signup/terms_to_blank
• As of June 2026
• A method that processes all transaction data at once to forecast sales volumes. Due to limits on the amount of data that can be input into generative AI, a subset of the purchase data is extracted for forecasting.
• The experiment was conducted using one year of transaction big data from approximately 100,000 people for 580 products in a specific category. The comparison between the conventional forecasting method and the proposed method was performed using the ratio obtained by dividing the cumulative sales forecasting gap by the total sales amount.