What is demand forecasting?
Demand forecasting is the estimation of how much of a product or service customers will want over a future period. A buyer needs a forecast of units by item, store, and week, while a contact center manager needs call volumes by queue and half hour. The numbers feed inventory planning, production scheduling, purchasing, and staffing, so forecast errors later appear as stockouts, excess stock, or idle staff.
Demand forecasting is a branch of predictive analytics, and the technique underneath is usually time series forecasting applied to sales or order history. The difficulty specific to demand is that recorded sales understate what customers wanted. When an item is out of stock, its sales fall to zero while customers go elsewhere. Bulk orders, returns, and promotions distort the history in similar ways.
How demand forecasting works
History is cleaned first, with adjustments for stockouts, one-off orders, and changed product codes. Drivers are added next: price, promotions, holidays, weather, local events, and, for business customers, open quotes and contract renewals. A model then forecasts each item and location. Large catalogs suit machine learning models trained across many items at once. Short or stable series often do better with simple statistical methods. Planners review the output, add what the model cannot know, and approve the numbers that drive orders.
Demand forecasting in practice
A regional home-improvement chain with thirty stores forecasts weekly demand for seasonal goods such as snow shovels, ice melt, and patio heaters. Weather is a model input. As a storm approaches, short-range forecasts refresh daily so the distribution center can move stock to the stores in its path. Store managers may adjust their numbers, and the team measures whether those adjustments improve accuracy, a practice called forecast value added.
In Kirk Tech Solutions engagements, this work usually starts in the data platform. Order history from every enterprise resource planning (ERP) and e-commerce system is merged into one table with consistent item, customer, and unit definitions. This step matters most for companies that grew by acquisition, which often sell one product under several codes.
What to watch for
- Correct the history for stockouts before training, or the model will underforecast the items that sell out most often.
- Forecast new products from the history of comparable items, and review those forecasts more closely than the rest.
- Ask for a range around each forecast, because safety stock and staffing decisions depend on how uncertain the number is.
- Track bias alongside an error measure such as mean absolute percent error (MAPE). A forecast that runs consistently high fills the warehouse even when its average error looks small.