The new AI forecasting system is able to reduce the forecasting mistakes with 20 to 30%. We’ve developed a further increase in accuracy of the forecast by including more grower-specific data.
Background and objectives
Kompany is a grower cooperative that supplies the market with high-quality fresh products,
including cucumbers, tomatoes, and strawberries. For the cucumbers, there are 31 growers
under the cooperation. Together they produce every year about 300 million cucumbers, with which Kompany has a market share of approximately 30% on the Dutch market.
One of the ways to better serve the market is a smarter, more accurate production forecast. The current overall forecast performance is relatively high, which is about 98% compared to the actual output. However, when the forecast number is analyzed on a detailed level (weight classes), Kompany is currently facing significant differences between the forecast production and actual production:
Per cultivation company, per sorting and cumulative on a weekly basis. In the current process, every week there is a forecast volume for the coming week per grower location submitted to a central platform. By 11am on Monday, the forecasting committee joins together and produces the final forecast based on the individual inputs. Afterwards, the feedback of forecast accuracy per grower company is provided. To summarize, the forecast activity is a manual process, which means the projection is made based on expert’s opinions, not any mathematical models. This leads to forecast error and brings a negative impact on the company.
The activities took place during three phases:
- Perform additions to the current AI model
- Prepare data, build and test new AI models
- Implement the new AI models in the Cloud environment
The current AI forecasting model takes the input from the forecasting committee and manages to already reduce the forecast error. By adding grower-specific data, the forecasting accuracy was further increased.