AI-Driven Demand Forecasting: Expectations vs. Reality
Demand forecasting is a critical aspect of management across various industries. With the advent of artificial intelligence (AI) and machine learning (ML) algorithms, businesses can now train models on historical data to predict future product demand. This capability allows for more accurate production planning, inventory management, and customer satisfaction. However, the implementation of AI in demand forecasting comes with its own set of challenges and realities. Let’s explore some real-world cases to understand the expectations versus the reality of AI-driven demand forecasting.
Case 1: Regional Bakery with Over 700 SKUs
Expectation: Daily demand forecasting by SKU with a one-month horizon for production planning.
Reality: The extensive assortment of the manufacturer lead to very sparse and intermittent demand by SKU, making it difficult to forecast using ML methods. Automating forecasting and planning is only feasible by aggregating data to the level of product categories to obtain more smoothed sales data. The excess assortment not only reduces forecasting accuracy but also decreases economic efficiency. Therefore, the marketing department was tasked with identifying high-margin items for subsequent assortment optimization to increase profits.
Reality: The extensive assortment of the manufacturer lead to very sparse and intermittent demand by SKU, making it difficult to forecast using ML methods. Automating forecasting and planning is only feasible by aggregating data to the level of product categories to obtain more smoothed sales data. The excess assortment not only reduces forecasting accuracy but also decreases economic efficiency. Therefore, the marketing department was tasked with identifying high-margin items for subsequent assortment optimization to increase profits.
Case 2: Gardening Products Manufacturer
Expectation: Daily demand forecasting by SKU with a two-month horizon in the retail sales channel, considering weather conditions.
Reality: Reliable weather forecasts are available with a horizon of 14 days. Weather forecasts for up to two months are based on historical observations and are not reliable. Therefore, predicting demand changes due to early spring, winter, or late autumn with high accuracy is not possible. In short-term forecasting for short delivery times, the weather forecast can be considered, and adding such a feature to the model improves forecasting quality.
Reality: Reliable weather forecasts are available with a horizon of 14 days. Weather forecasts for up to two months are based on historical observations and are not reliable. Therefore, predicting demand changes due to early spring, winter, or late autumn with high accuracy is not possible. In short-term forecasting for short delivery times, the weather forecast can be considered, and adding such a feature to the model improves forecasting quality.
Case 3: Alcoholic Beverage Manufacturer
Expectation: Demand forecasting by SKU and sales channels, considering competitors’ sales to assess the potential for sales growth.
Reality: To solve this problem, it is necessary to enrich the forecasting model with external data on competitors’ sales in sales channels. For subsequent automatic forecasting, competitors’ sales data must be regularly fed into the model. However, sales data is not publicly available, and obtaining it requires contacting data holders, such as retail chains. Since providing data is a paid service, this significantly increases the cost of forecasting and makes the solution unattractive in terms of “effort spent vs. result obtained.” It is recommended to start building the ML model on data that is available without additional costs. Even at this stage, the benefits of automatic forecasting and improved forecast quality can bring value to the company. Sales plans can be revised based on operational demand forecasting, applying analytics, for example, once a quarter.
Reality: To solve this problem, it is necessary to enrich the forecasting model with external data on competitors’ sales in sales channels. For subsequent automatic forecasting, competitors’ sales data must be regularly fed into the model. However, sales data is not publicly available, and obtaining it requires contacting data holders, such as retail chains. Since providing data is a paid service, this significantly increases the cost of forecasting and makes the solution unattractive in terms of “effort spent vs. result obtained.” It is recommended to start building the ML model on data that is available without additional costs. Even at this stage, the benefits of automatic forecasting and improved forecast quality can bring value to the company. Sales plans can be revised based on operational demand forecasting, applying analytics, for example, once a quarter.
Key Considerations for Implementing AI in Demand Forecasting
- Defining the Task and Setting Boundaries: Clearly define the task that AI will solve and its purpose, such as daily demand forecasting by current SKUs for the next four weeks for production planning.
- Data Quality and Availability: AI solutions rely on data, so the feasibility directly depends on the quantity and quality of the data. Organize systems for collecting, processing, and storing data to develop competitive advantages from their use.
- Increasing Staff Awareness of AI Solutions: Training employees on the principles of AI will help manage expectations and prevent misunderstandings, while also using the technology as effectively as possible.
Conclusion
Using AI for demand forecasting is a powerful tool that can help businesses make more informed decisions. However, it is important to understand that AI is not a universal solution and has its limitations. Educating realistic expectations, choosing the right method, and focusing on data quality will help successfully implement AI and achieve the desired results. While AI offers substantial benefits, understanding its limitations and continuously refining the approach will ensure that it remains a valuable asset in the long term.