Forecasting starts with the data you already own
Independent retailers with 8 to 12 weeks of sales history for each active SKU already have enough information to begin forecasting demand. That information is usually sitting in the point of sale system, an ecommerce platform, a spreadsheet, or a mixture of all three. The work is to organize it into a regular process that helps you decide what to buy, when to buy it, and how much stock to hold.
This is different from asking a system to predict the future perfectly. Demand forecasting is a structured projection from what has happened before. It gives you a measured expectation, along with a way to see when actual sales begin to differ from that expectation. That distinction matters. A forecast can reduce avoidable stockouts and excess inventory, but it cannot tell you with certainty what a customer will want next Thursday.
For an independent retailer, the value is often practical rather than theoretical. You want to know whether a product is selling at a steady rate, whether a seasonal lift is beginning, and whether the stock on hand will last until the next delivery. Those are questions that a useful forecast should answer in plain language.
What demand forecasting is, and what it is not
A forecast starts with a history of sales and looks for repeatable patterns. It may identify a stable run rate, a gradual increase or decline, a recurring seasonal effect, or a sudden change connected to an event. The output is an estimate of future demand for a defined period, such as the next seven, 30, or 90 days.
It is not a crystal ball. Historical sales can be incomplete or distorted. A stockout may make demand look lower than it really was because customers could not buy the item. A promotion may make one week look unusually strong. A new competitor, a change in local footfall, or a product discontinuation can make old patterns less relevant.
It is also not the same as knowing what will sell. Forecasting reduces risk by making assumptions visible and consistent. It does not eliminate risk. The buyer still needs to apply commercial judgment, especially for products with little history, irregular demand, or a short selling window.
The useful question is not, “Is this forecast always right?” No forecast is. The useful questions are, “How was it calculated?”, “How far wrong is it usually?”, and “What should I do when actual sales move away from it?”
The minimum viable inputs
You do not need years of perfectly clean data to start. For many products, 8 to 12 weeks of sales history is a reasonable minimum for an initial forecast. More history is helpful, especially when the product has a clear annual pattern, but a shorter period can still support a basic run rate.
Start with these inputs:
- Sales by SKU, with the date and quantity sold.
- At least a weekly sales cadence, although daily data is better for products that move quickly.
- Current stock on hand and, if possible, stock already on order.
- Supplier lead time, meaning how long it takes from placing an order to receiving it.
- Promotions, holidays, closures, and other events that affected sales.
- Separate sales by channel when the channels behave differently, such as a shop, website, and marketplace.
Channel separation is important. If a product sells mostly online but is also stocked in a shop, combining all sales into one number may hide a useful pattern. The total demand may be accurate while the allocation is wrong. You could have enough units across the business and still run out in the channel where customers expect to find them.
Clean data does not mean perfect data. It means you understand the main exceptions. Mark weeks when an item was unavailable, record unusual promotions, and distinguish a genuine zero sale from a product that was not listed or not yet launched. These details prevent the model from treating missing opportunity as weak demand.
Three simple ways to estimate demand
Moving average
A moving average takes sales from a recent number of periods and calculates the average. If a product sold 10, 12, and 14 units in the last three weeks, the three week moving average is 12 units per week. The estimate then moves forward as each new week replaces the oldest one.
This method is easy to explain and useful for products with fairly stable demand. Its weakness is that it can react slowly when sales are rising or falling. A long average smooths out noise, but it may also hide a recent change. A short average responds faster, but it can be more affected by one unusual week.
Exponential smoothing
Exponential smoothing also uses past sales, but it gives more weight to recent observations. In practice, this means a strong recent week has more influence than a much older week. The method is still understandable, even though the calculation is more structured than a simple average.
This approach is useful when demand is relatively stable but changes gradually. It can follow a product that is building momentum or slowing down without allowing one isolated transaction to dictate the entire forecast. At Stockagile, we use exponential smoothing for stable SKUs, then recalibrate the forecast weekly.
When a simple method is not enough
A moving average or smoothing method may be sufficient for a year round product with a regular sales pattern. It becomes less suitable when demand depends on several factors at once, such as holidays, local events, weather sensitive categories, promotions, or different customer behavior by channel.
Stockagile adds a gradient boosted machine learning layer for seasonal and event driven demand. The purpose is not to hide the calculation behind a mysterious score. Buyers see outputs such as run rate, days to stockout, and suggested reorder quantity. The forecast is recalibrated weekly so that recent sales and known changes can influence the next recommendation.
Whatever method you use, keep the output understandable. A buyer should be able to see why a reorder is suggested and decide whether the underlying assumptions still make sense.
Seasonality changes the question
Some products sell at a steady level throughout the year. Others have a trend, a seasonal pattern, or both. A retailer may be selling more of a category because it is growing, because the calendar is approaching its normal peak, or because a one time event has created extra demand. These causes should not be treated as the same thing.
One useful way to think about seasonality is to separate the underlying trend from the recurring calendar effect. The trend answers, “Is this product becoming more or less popular over time?” The seasonal component answers, “Does this product usually sell more or less during this part of the year?”
Consider a shop that sells school supplies. Sales may rise every August and September, then fall in October. If you ignore seasonality, a forecast based on the quiet months could recommend too little stock before the school term. If you calculate the August increase as permanent growth, it could recommend too much stock later in the year.
Seasonality requires enough history to see a repeated pattern. With only a few weeks of data, you may be able to identify a current run rate, but you cannot confidently estimate an annual cycle. In that situation, use category knowledge and supplier constraints as context, and treat the seasonal assumption as something to review rather than a fact.
A practical example: three shops and 2,000 titles
Imagine an independent bookshop with three locations in Madrid. It carries about 2,000 titles across a year round catalog, plus seasonal categories such as school books, calendars, and holiday gift books. The owners have been buying largely from experience. That experience is valuable, but each location has developed a different spreadsheet, and replenishment decisions often happen after a manager notices an empty shelf.
The first step is not to forecast every title in the same way. The team separates sales by location and groups products by behavior. A popular year round title with regular sales can use a simple smoothing method. A calendar has a short selling window and needs an expected seasonal curve. A new author event may require a manual adjustment because the store has no comparable history.
For each established title, the team reviews the last 12 weeks of weekly sales, current stock, open purchase orders, and supplier lead time. They flag weeks when a title was unavailable. If one location sold zero copies because it had no stock, that zero is not treated as evidence that local demand disappeared.
Suppose one title is selling at about 15 copies per week across the three locations, and the supplier usually takes two weeks to deliver. The buyer does not need a precise prediction of every daily sale. The practical question is whether the combined stock will cover the lead time and the expected demand until the next review. If the forecast shows 30 units of demand during that period and only 18 units are available, a reorder deserves attention.
The team also reviews the forecast against actual sales every week. If a recommendation was consistently too high for one location, the reason might be a local preference, a display change, or a channel shift. If it was too low before a seasonal peak, the seasonal assumption needs adjustment. Over time, the process replaces unrecorded intuition with measured intuition. The buyer still makes the decision, but the decision has a clearer starting point.
Forecast error is part of the process
A forecast should be judged by its errors, not by how confident its presentation sounds. Compare the forecast with actual sales over time. Look for bias, where estimates are repeatedly too high or too low, and for volatility, where errors vary widely from one period to the next.
A bad model with false confidence is worse than an honest intuition. If a spreadsheet produces a single exact number without showing the assumptions behind it, the number may encourage overbuying or delay a necessary order. Calibration is the goal. You want a forecast that is appropriately cautious when the data is thin and more specific when the pattern is stable.
There are clear boundaries. A new SKU has no sales history. A product may be discontinued before its pattern can be learned. A one time event can create demand that will not repeat. Forecasting cannot resolve these cases by itself. They need a buyer’s decision, a supplier conversation, or a deliberate business rule.
To begin, choose 20 regularly sold SKUs, export their last 12 weeks of sales, and mark every stockout or promotion before calculating a weekly average. Review that list with your next purchase order. That small exercise will show you where the data is reliable and where judgment remains essential.
