AI demand forecasting is the use of machine learning to predict next period's sales by product, branch or customer from past sales, seasonality, promotions, prices, holidays and, where relevant, the weather. The model links the forecast to safety stock and reorder points and gives planners an explained order recommendation. Its value lies in better purchasing and production decisions.
Is planning still living in one person's spreadsheet?
In many businesses, the answer to "how much should we order?" sits in an experienced planner's spreadsheet and instincts. They look at the same month last year, decide "we will sell a bit more this time" and type in a number. That works while the range is small and the customer base stable. Once the range grows to thousands of items, the number of branches reaches dozens and promotions become frequent, two problems appear:
- Scale: one person cannot review hundreds of product-branch combinations every week. Attention goes to best-sellers, and the rest are planned as a copy of last week.
- Continuity: the knowledge lives in someone's head. When the planner goes on leave or moves on, the plan falters. It is a textbook case of outgrowing spreadsheets.
The data is often not ready either. According to TurkStat's ICT Usage in Enterprises Survey 2025, 28.3% of Turkish enterprises with 10 or more employees use ERP software, yet only 6.5% use business intelligence (BI) software. Sales data is recorded but not analysed for decisions. The structure that makes it analysis-ready is a data warehouse and ETL process.
Forecast errors cost money in both directions
A forecasting error loses money whichever way it falls:
- Overstock ties up capital on the shelf and fills warehouse space. For products with a shelf life it becomes waste; for fashion or seasonal items it becomes markdowns.
- Stock-outs lose sales and send customers to competitors. In manufacturing, a missing raw material stops the line or forces urgent, expensive purchasing.
In food and fresh produce, the loss is visible on a global scale.
FAO estimates that 13.3% of food was lost globally in 2023 after harvest, on farms and in transport, storage, wholesale and processing (before reaching retail), rising to 25.4% for fruits and vegetables. (FAO SDG 12.3.1 Global Food Loss Indicator)
The consumer end of the chain tells a similar story. According to the UNEP Food Waste Index Report 2024, 1.05 billion tonnes of food were wasted in 2022, with food services accounting for 28% and retail for 12%. Not all of this is down to forecasting, but decisions on how much to produce and order are among its direct drivers. For processors of farm produce, supply is uncertain as well as demand; estimating incoming volumes before harvest with drone crop yield estimation lets you see both ends of the plan. Other uses in the field are gathered in AI in agriculture.
Then there is data quality. Data quality expert Thomas C. Redman estimates the cost of bad data at 15% to 25% of revenue for most companies (MIT Sloan Management Review: Seizing Opportunity in Data Quality, 2017). In forecasting, this shows up as errors such as treating "zero sales" on out-of-stock days as zero demand, or failing to net returns off sales; we cover these problems in data quality and duplicate records.
Which forecasting method, and when?
| Method | How it works | Best suited to | Limitation |
|---|---|---|---|
| Moving average / same period last year | Averages past periods | Few products, stable demand | Blind to promotions and trends |
| Statistical time series (e.g. exponential smoothing) | Separates trend and seasonality | Products with a regular season | Limited handling of external factors (price, holidays, weather) |
| Machine learning models | Learns sales history together with price, promotions, holidays and weather | Many products, many branches, frequent promotions | Needs clean data with enough history |
| Hierarchical approach | Keeps forecasts consistent across total, category and item levels | Portfolios with many slow-moving items | Requires more design effort |
In practice, a good system does not commit to one method; it tests which method performs best on historical data for each product group. Because sales history already sits in the ERP, demand forecasting is one of the data-ready first projects to consider when starting with AI in business.
A practical framework: demand forecasting in seven steps
- Define the decision. Which decision will the forecast feed: weekly branch orders, monthly raw-material purchasing, production capacity? The horizon and frequency of that decision set the forecast level (item-branch-week or category-month).
- Collect and clean the data. At least two years of sales, returns, out-of-stock days, prices and the promotion calendar. Days when nothing sold because there was nothing on the shelf must not be recorded as "no demand". Anomaly detection methods help filter out one-off bulk orders and data-entry errors.
- Add external factors. Public and religious holidays, school terms, promotions, competitor activity and, where relevant, the weather. In Türkiye, religious holidays move through the calendar each year, and models built on fixed calendar months often miss them. Using the forecast to set the price itself is the subject of AI dynamic pricing.
- Backtest. Train the model on part of the history, have it forecast the following period and compare with actuals. Race it against your current method, including the planner's own numbers, over the same period.
- Choose the right metric. Alongside item-level error, track volume-weighted error and bias (whether the forecast is consistently high or low). Percentage errors can mislead on slow-moving items.
- Link the forecast to a stock rule. The forecast and its uncertainty feed safety-stock and reorder-point calculations, and the output is an order recommendation. For products with a shelf life, this rule must work alongside FIFO and FEFO inventory methods.
- Keep the planner in the loop. The model provides the starting figure; the planner adjusts it for what they know (a new customer, a supply issue). Adjustments are recorded, and the next period shows whether they improved the forecast.
The same method answers other questions too. We show how forecasts turn into order and inter-store transfer decisions in retail inventory optimisation, and how sites predict their hourly energy needs in electricity load forecasting.
How we set up AI demand forecasting at Digital Bridge
Our demand forecasting and predictive analytics service starts not with the model but with the decision the forecast will change. However accurate it is, a forecast that does not change a purchasing or production decision is just a report.
- Data review: we go through your ERP, POS, e-commerce and warehouse records with you and identify missing or inconsistent data. If data is scattered, we first build the foundations with a data warehouse and data governance and quality work.
- Model and backtest: we build models that combine your sales history with seasonality, promotion effects, public holidays and, where relevant, the weather, and compare them with your current method over a past period. We review the results with you.
- Tying it to decisions: we link the forecast to safety-stock and reorder-point calculations and post order recommendations to your ERP. The planning team does not have to move to a new program; they see the recommendation and its reasoning on BI dashboards.
- The warehouse side: a forecast only makes sense against a system that knows what is really on the shelf. If your stock records do not match reality, we use warehouse and inventory management (WMS) to capture every movement by barcode as it happens, so the forecast runs on true stock. Our warehouse management system guide covers this in more detail. For other AI uses in the warehouse, see AI in logistics and warehousing.
- The production side: for manufacturers, we connect the forecast to capacity and shift planning through production planning and scheduling (APS). How the forecast becomes a schedule is explained in AI production scheduling.
The scenarios we meet most often are branch-level ordering in chain stores, fresh products and waste in food production, dealer shipments in wholesale distribution, pre-promotion stock building in e-commerce, and spare-parts demand in service businesses. Our retail and chain stores page looks at branch-level planning in particular.
How will you measure the value?
A forecasting project should be judged by well-chosen KPIs, not model error alone: stock turnover, number of out-of-stock days, waste and write-off value, number of urgent orders, and the time planners spend on weekly planning. Record baseline values for these before the project starts, or it becomes hard to prove the gain. For the method, see our guide to measuring AI project ROI.
We also recommend measuring forecast value added: checking whether each step (statistical model, AI model, planner adjustment) actually improves on the one before it. In some product groups a simple method beats a complex model, and some manual adjustments make the forecast worse. Measuring this directs effort to where it really adds value and grounds trust in the model on evidence.
In retail, adding store traffic to sales data can strengthen a forecast; our guide to a retail people counting system explains how to collect footfall data.
Next step
Start with the product group that causes you the most trouble: the one with the most waste or the most frequent stock-outs. Export two years of sales, returns and promotion records for it. Then get in touch; we will review your data together, run a backtest on that product group and show in numbers whether the model beats your current method. For AI use cases beyond forecasting, browse our complete Artificial Intelligence guide.