Demand Forecasting for Small Business: A Practical System for Buying the Right Stock

Small businesses rarely lose cash because they cannot create a spreadsheet. They lose cash because the spreadsheet does not change the buying decision. Demand forecasting should answer one practical question: how much are you likely to sell during the period in which you must decide what to buy, make or reorder? You do not need an AI platform to start. A clean weekly process with a simple baseline, known business events and forecast-error tracking can already improve decisions. Demand forecast is not a sales target NumberMeaningUse Sales targetWhat you want to sellMotivation and planning Demand forecastWhat customers are likely to buyInventory and capacity Purchase planWhat you should orderSupplier action If your target is 1,000 units but the evidence supports 700 units, ordering 1,000 only because that is the target can turn ambition into dead stock. Start with the decision horizon Your forecast horizon should reflect lead time. A local supplier that replenishes in seven days needs a different process from an imported component that takes 90 days. For each SKU, record supplier lead time, minimum order quantity, review frequency and desired safety stock. This tells you how far ahead the forecast must be useful. Method 1: simple moving average If weekly demand is reasonably stable, average the last few comparable periods. Suppose weekly sales were 92, 105, 98 and 109 units. Four-week moving average = (92 + 105 + 98 + 109) ÷ 4 = 101 units That gives you a baseline of 101 units for the next comparable week. Do not stop there. Ask whether the coming week has a promotion, festival, stockout recovery, price change or unusual event. Method 2: seasonal comparison For products affected by festivals, weather, school calendars or wedding seasons, compare with the same period last year, then adjust for current growth and known changes. If last year's Diwali-week demand was 500 units and the business is running 12% above last year's comparable baseline, a starting forecast might be 560 units. Then adjust for price, promotion, channel mix and availability. Method 3: weighted recent demand If recent weeks are more relevant, give them higher weights. Example: last three weeks sold 80, 100 and 120 units. Apply weights of 20%, 30% and 50%. Forecast = 80×0.20 + 100×0.30 + 120×0.50 = 106 units This responds faster to a trend than a plain average. Separate demand from sales when stockouts happen Sales data can understate demand. If you sold 100 units because you ran out of stock on day 20, that does not mean monthly demand was 100. Flag stockout periods. Use waiting lists, lost-order notes, marketplace availability data or daily run rates to estimate constrained demand. Keep the estimate separate from actual sales so you can audit it later. Measure forecast error A forecast is useful only if you compare it with actual demand. AWS Supply Chain documents both MAPE and WAPE as forecast-accuracy metrics. For a small business with many SKUs, WAPE can be easier to interpret at an aggregate level: WAPE = Sum of absolute forecast errors ÷ Sum of actual demand × 100 Suppose three SKUs have actual demand of 100, 200 and 50. Forecasts were 90, 230 and 40. Absolute errors are 10, 30 and 10. WAPE = 50 ÷ 350 × 100 = 14.3% Track the number over time. Do not chase a universal benchmark. Different categories have different volatility. Track bias too A forecast can have similar average error while being systematically high or low. Oracle's documentation notes that consistent over-forecasting can build excess inventory, while consistent under-forecasting can consume inventory and hurt service. Use a simple signed error: Forecast error = Actual demand - Forecast demand If the number is positive for many periods, you are repeatedly under-forecasting. If it is negative, you are repeatedly over-forecasting. Turn the forecast into a reorder decision A forecast by itself does not tell you what to buy. Use: Planned order = forecast demand during lead time + target safety stock - available inventory - confirmed inbound inventory Example: Expected demand during lead time: 400 units Target safety stock: 100 Available stock: 220 Confirmed inbound: 80 Planned order = 400 + 100 - 220 - 80 = 200 units. Then apply supplier minimum-order quantities and pack sizes. Create ABC forecast discipline Do not spend equal time on every SKU. ClassTypical importanceReview AHigh revenue, margin or stockout impactWeekly BModerate impactEvery 2 weeks CLow impact or long tailMonthly This keeps the process manageable. Build an event calendar Your model cannot know everything. Add a calendar for festivals, promotions, price changes, influencer campaigns, marketplace events, distributor orders, shutdowns and supplier disruptions. Record the adjustment separately from the baseline. If the model forecast is 1,000 and you add 200 for a promotion, save both numbers. After the event, check whether the adjustment helped. When AI becomes useful AI and machine learning can help when you have enough clean history across many SKUs, stores or channels and meaningful external drivers. Modern forecasting systems can use product attributes, promotions, prices and other demand drivers. But automation does not repair bad inputs. Before buying software, check whether SKU IDs are consistent, returns are separated, stockouts are flagged, promotions are recorded, cancelled orders are removed and supplier lead times are maintained. A weekly 45-minute forecasting routine Refresh actual sales and inventory. Flag stockouts and unusual orders. Generate the baseline forecast. Review A-class SKUs. Apply documented event adjustments. Calculate WAPE and bias for the previous period. Convert forecast into purchase recommendations. Approve orders only after checking cash availability. A 30-day implementation plan WeekAction 1Clean SKU history, lead times and stockout flags. 2Create baseline forecasts for top 20 SKUs. 3Add event adjustments and purchase logic. 4Measure WAPE and bias, then refine the method. Mistakes to avoid Using targets as forecasts.Ignoring stockout periods.Forecasting every SKU with the same method.Changing forecasts without recording why.Judging accuracy only after a major seasonal event.Ordering from the forecast without checking cash and supplier constraints. FAQs How much history do I need?Start with what you have. More history helps identify seasonality, but even 8 to 12 clean weeks can support a simple baseline for stable items. Is MAPE always the best metric?No. MAPE can behave poorly when actual demand is zero or very small. WAPE, MASE and other measures may be more useful depending on the data. Should I forecast revenue or units?For inventory decisions, forecast units first. Revenue can change because of price even when physical demand does not. What to do next Pick your 20 most important SKUs. Forecast the next four weeks with the simplest reasonable method. Save the forecast before the period starts. Then compare it with actual demand. That one discipline is more valuable than a sophisticated model you never measure. SourcesAWS Supply Chain demand planning accuracy metricsAWS forecast accuracy metricsOracle forecast bias and accuracy documentation