Ordering too much ties up cash and fills shelves with items that do not sell. Ordering too little loses sales and disappoints customers. Demand forecasting is the practice of estimating how much you will sell, so purchasing decisions rest on evidence rather than instinct. Many owners assume it requires a data science team. In reality, a company with a year or two of sales records in a POS or ERP already has the main ingredient. This article shows how to turn it into useful forecasts.
Check what data you already have
Forecasting needs a history of what sold, when and where. Most systems that record invoices or receipts store this already. Look for these fields:
- Product or item code, and its category.
- Quantity sold, with a date and, ideally, a time.
- Branch or channel, such as shop, online or sales team.
- Price and any discount applied.
- Stock levels over time, so you can tell low sales from no stock.
That last point is often missed. If an item was out of stock for two weeks, zero sales in that period do not mean zero demand. Mark those periods, or the forecast will learn the wrong lesson. Systems such as our Narmar POS & ERP, which keep accounting, inventory and sales in one place, make this kind of analysis far easier because the data is already connected.
Prepare the data before modelling
Clean the basics
Merge duplicate items, fix obvious data entry errors such as quantities that are a hundred times too large, and handle returns consistently. Decide what to do with one-off bulk orders: forecast them separately, or remove them from the baseline so they do not distort normal demand.
Choose a sensible level of detail
Forecasting every item at every branch for every day produces noisy data. Weekly figures by product and branch are a good starting point for most retail and distribution businesses. Slow-moving items can be forecast at the category level, or managed with simple reorder rules.
Capture events that change demand
Ramadan, Eid, the school year, exam seasons, summer travel, public holidays and price promotions all shape demand in Egypt. Record them as a calendar so a forecast can learn that a spike belongs to an event, not to a trend.
Start with a simple baseline
Before using machine learning, build the simplest forecasts that could work and measure them. They are quick to produce and often hard to beat.
- Last period. Predict that next week equals this week.
- Moving average. Predict the average of the last several weeks.
- Seasonal comparison. Predict the same week last year, adjusted for growth.
The baseline gives you a benchmark. A more advanced model is only worth its cost if it clearly beats these numbers on data it has not seen.
Measure accuracy honestly
Hold out recent data
Build the forecast using data up to a certain date, then compare it with what actually happened afterwards. Repeat across several periods. Judging a model on the same data it learned from will always look good and prove nothing.
Use error measures people understand
Average absolute error, expressed in units or as a share of sales, is easy to explain to buyers. Look at bias too: a forecast that is consistently too high or too low is a different problem from one that is simply noisy. Review accuracy by product group, because a model may work for fast movers and fail for slow ones.
When machine learning helps
Once a baseline is in place, models that use several signals can add value: day of week, holidays, promotions, prices, weather, and the sales of related products. They are most useful when you have many items, a long history and clear drivers of demand. Keep them explainable, so a purchasing manager can see why the forecast changed, and keep the process simple enough to run on a schedule without manual work.
Turn the forecast into decisions
A forecast has no value until it changes an order. Connect it to your purchasing process: suggested order quantities based on forecast demand, current stock, supplier lead time and safety stock. Show the suggestion to a buyer who can accept or adjust it, and record the reasons for changes. Those adjustments are useful feedback for improving the system.
For multi-branch businesses, forecasts also guide transfers between locations, helping to move stock to where it will sell. In our work on Narmar POS & ERP we see inventory, CRM and accounting as a base that analytics and an AI business advisor can build on, because insights depend on clean transactions.
Start small and keep a human in charge
Choose one product family and one or two branches. Run forecasts alongside your current ordering process for a few cycles and compare the results. Expand when the numbers show a real improvement. Forecasts support buyers; they do not replace their knowledge of suppliers, customers and the local market.
If you would like to see what your own sales history can tell you, the Narmar team can review your data and propose a practical first forecasting project. See our AI services or contact us to begin.
