AI Trend Research

← AI glossary

AI glossary · Basic · Classic ML, recsys & search · Logistics & supply chain · Retail

Demand forecasting

Demand forecasting is estimating in advance which products will sell, when, where and in what quantities, based on past sales records and information such as weather, events and prices.

About this term
Level
Basic · Technical terms you often meet in AI news
Category
Industry AI
Added
2026-10-10
In plain words It is like a neighborhood bakery owner deciding how much dough to make from experience: 'less bread sells on rainy Fridays, and a lot sells on days with events at the nearby school.' AI demand forecasting calculates such rules of thumb from data for thousands of products and stores at once. However, events never seen before (sudden trends, disasters) are not in the data, so it does not predict them well.

How it works

First, past sales volumes are organized by product and store in time order. Information that affects demand, such as day of the week, season, holidays, discount promotions, price changes and weather, is attached to them. The model learns these relationships using time series forecasting techniques, gradient boosting or deep learning models, and outputs sales volumes for the next day, week or month. A good system reports not just a single number but also the uncertainty, such as 'it is likely to fall within this range'. As actual sales come in, forecasts are compared against them and continually corrected.

The demand forecasting loop

  1. 1. Collect data Sales records, weather, events, prices
  2. 2. Train the model Learn relationships that affect demand
  3. 3. Forecast sales Including range and uncertainty
  4. 4. Compare with actual sales Check errors and correct
Forecasting is not done once and finished; it is continually corrected by comparing against actual sales.

Why it matters

In logistics and retail, too much inventory increases storage costs and waste, and too little causes stockouts that lose sales and customers. Demand forecasting is the starting point for almost every operational decision, such as order quantities, warehouse inventory placement, and staff and vehicle allocation. Especially for products with short shelf lives, such as fresh food, forecast accuracy translates directly into the amount thrown away. The longer the supply chain, the more small forecast errors at the front grow further down the line, so accurate forecasts are even more valuable.

Things to know

Sales records from days when a product was out of stock show 'the amount sold up to available stock', not 'actual demand', so training on them as-is leads to underestimating demand. New products have no past records, so data from similar products must be borrowed. During periods when consumer habits change suddenly, models can quickly go off track due to data drift, so ongoing monitoring is needed. Also, rather than feeding forecast results directly into automatic ordering, many setups include a step where people make adjustments so that information only on-site staff know, such as promotion plans, can be reflected.

Example

A large supermarket chain forecasts next week's sales by store and product to set distribution center order quantities and store shelf quantities. For example, when hot weather is forecast for the weekend, the model raises demand estimates for ice cream and bottled water, and delivery vehicle scheduling and warehouse inventory placement are adjusted in advance to match. E-commerce companies use the same forecasts to decide which products to move ahead of time to distribution centers close to customers.

Where it came up on this site

Related terms

Time series forecasting Gradient boosting Data drift Route optimization Alternative data Human-in-the-loop