Demand forecasting is estimating how much of each product your customers will buy in the coming weeks, from your sales history and known events. Every order you place bets on that number: too high ties up cash, too low empties the shelf. For a shop, the useful forecast is weekly, in units, and checked against what really sold.
This guide works seven steps on 12 weeks of one Kumasi mini-mart's soft-drink sales, in Ghana cedis. It then scores three forecasts against what sold, so you can see which one to trust.
What demand forecasting is, and what it is not
Demand forecasting predicts what customers will want to buy, not what you should hold. Turning that demand into stock levels and orders is a separate job, set out in inventory forecasting methods.
Demand and sales are not the same number. Sales are what you sold; demand is what customers wanted, including what they couldn't buy because the shelf was empty.
Demand planning comes after the forecast: deciding what to buy, make or promote to meet it. A workshop's version is the production plan.
Types and methods at a glance
Forecasts are grouped by how far ahead they look, and by whether they rest on numbers or on judgement. A small shop needs the short-term, number-based kind most weeks.
| Type | Looks ahead | Used for | Shop example |
|---|---|---|---|
| Short term | Days to a few months | Weekly orders and staff rotas | Soft drinks for the next four weeks |
| Medium term | A few months to a year | Seasons, promotions and supplier deals | Stock for the festive season |
| Long term | A year or more | A new branch or a new product line | Whether a second shop would sell enough |
| Family | Examples | Built on | Use it when |
|---|---|---|---|
| Quantitative | Moving average, exponential smoothing, seasonal index, regression | Your own sales history | You have at least 12 weeks of sales for the product |
| Qualitative | Your own judgement, the supplier's view, asking regular customers | What people know about the market | A new product, or a change your history has never seen |
Most shops need both. The numbers set the baseline, and judgement adds what the numbers can't know, such as a booked order.
How to forecast demand in 7 steps
The steps below turn raw sales into a forecast you can test. The worked example after them runs each step on one product.
Step 1: Choose what to forecast, and in what unit
Forecast the products that carry your sales first. Your A items deserve a weekly forecast each; the long tail can share one rough forecast per category.
Forecast in units, not money. A price rise lifts money sales without any change in how many bottles people want.
Look ahead at least as far as the supplier's lead time plus the days until your next order. A shorter forecast can't change what you buy.
Step 2: Pull at least 12 weeks of weekly sales
Twelve weeks show the level and the pattern within a month. A full year also shows the seasons, such as Christmas, Easter or school reopening.
Step 3: Clean the history
Raw sales hide three things that are not normal demand. Mark each one before you average anything.
- Weeks the shelf ran empty. Sales stop when stock does, so that week understates demand. Estimate the full week from the days you had stock.
- One-off bulk orders. A funeral, a wedding or a school event can buy a month's stock in one sale. Take it out, and plan such orders on their own.
- Promotions and closures. A price cut lifts a week; a closed day or a long power cut lowers it. Mark them so they don't set next month's level.
Step 4: Set the baseline
The baseline is a normal week's demand once the one-offs are out. The average of the clean normal weeks is enough to start.
A moving average or exponential smoothing keeps the baseline current as new weeks arrive. Both are worked in the inventory forecasting guide linked above.
Step 5: Add the patterns and the booked orders
Add what repeats and what is already booked, and nothing else. A pay week, a public holiday or school reopening repeats; a customer's confirmed order is booked.
The textbook Forecasting: Principles and Practice, by Rob Hyndman and George Athanasopoulos, says to use judgemental adjustments sparingly. Adjust only for important information the numbers don't hold, and write down why.
Put a booked order on its own line, not into the baseline. If the customer cancels, it comes out cleanly.
Step 6: Measure the error every week
The forecast error is what sold minus what you forecast, as the same book defines it in its chapter on forecast accuracy. A positive error means you forecast too low; a negative one, too high.
| Measure | Formula | What it tells you |
|---|---|---|
| Error | Actual sales − forecast | One week's miss, with its direction |
| MAE, mean absolute error | Average of the errors, ignoring the sign | The typical miss, in units |
| MAPE, mean absolute percentage error | Average of (error ÷ actual × 100), ignoring the sign | The typical miss as a percentage |
| Bias | Average of the errors, keeping the sign | Whether you lean high (negative) or low (positive) |
MAPE fails on slow sellers. The book notes it is infinite or undefined when a week's sales are zero, so use MAE for those products.
The benchmark is the naive forecast: next week equals last week. A method that can't beat it isn't worth the time.
Step 7: Hand the forecast on
A forecast earns its keep only when it changes an order. Turn it into reorder points that stop overstock and stockouts, and into cash by multiplying units by price.
Worked example: a Kumasi mini-mart's soft drinks
For example, a mini-mart in Kumasi sells a 1.5-litre soft drink at GHS 15 a bottle. In this example, the last week of each month is pay week, when many customers have just been paid.
| Week | Bottles sold | Owner's note | Demand used |
|---|---|---|---|
| 1 | 118 | 118 | |
| 2 | 112 | 112 | |
| 3 | 120 | 120 | |
| 4 | 156 | Pay week | 156 |
| 5 | 115 | 115 | |
| 6 | 68 | Sold out after Thursday | 119 |
| 7 | 121 | 121 | |
| 8 | 160 | Pay week | 160 |
| 9 | 117 | 117 | |
| 10 | 362 | Funeral order of 240 bottles | 122 |
| 11 | 119 | 119 | |
| 12 | 158 | Pay week | 158 |
Week 6 sold 68 bottles in four days, so its full-week demand is 68 ÷ 4 × 7 = 119. Week 10's funeral order comes out, leaving 122 bottles of normal sales.
The nine normal weeks add up to 1,063 bottles, an average of 118. The three pay weeks average 158, so a pay week adds 40 bottles.
A customer has also booked 10 packs of 12, or 120 bottles, for a wedding in week 15. Here is the forecast for the next four weeks.
| Week | Baseline | Pay-week lift | Booked order | Forecast |
|---|---|---|---|---|
| 13 | 118 | 118 | ||
| 14 | 118 | 118 | ||
| 15 | 118 | 120 | 238 | |
| 16 | 118 | +40 | 158 | |
| Total | 632 |
The shop expects to sell 632 bottles in four weeks, worth GHS 9,480 at GHS 15 each. That figure feeds the cash plan as well as the order.
Which forecast was right? Three methods scored
Four weeks later, regular sales came in at 109, 125, 116 and 171 bottles, and the wedding order was collected. Here is the cleaned forecast against two shortcuts: last week's regular sales (the naive forecast), and the average of all 12 raw weeks.
| Week | Actual | Cleaned forecast (error) | Naive: last week (error) | Raw average (error) |
|---|---|---|---|---|
| 13 | 109 | 118 (−9) | 158 (−49) | 144 (−35) |
| 14 | 125 | 118 (+7) | 109 (+16) | 144 (−19) |
| 15 | 116 | 118 (−2) | 125 (−9) | 144 (−28) |
| 16 | 171 | 158 (+13) | 116 (+55) | 144 (+27) |
| MAE | 7.75 | 32.25 | 27.25 | |
| MAPE | 5.8% | 24.4% | 21.8% | |
| Bias | +2.25 | +3.25 | −13.75 |
The cleaned forecast missed by about 8 bottles a week; both shortcuts missed by 27 or more. The naive forecast misses on both sides of a pay week, and the raw average carries the funeral order into every week.
The raw average's bias of −13.75 is the costly one, because it over-forecasts week after week. Over these four weeks it would have bought 576 regular bottles against the cleaned 512: 64 too many, or GHS 768 at a GHS 12 cost.
The cleaned forecast's small positive bias is the cheap kind. A little safety stock covers a 9-bottle shortfall over a month.
How accurate should a demand forecast be?
Accurate enough to beat the naive forecast on your own products, with no lasting bias. No single percentage fits every shop, because a steady seller and a slow one miss by very different amounts.
Watch the bias more than the size of the miss. If the errors keep the same sign for four weeks running, your baseline has moved: reset it from the latest clean weeks.
Expect bigger misses on slow sellers, new products and longer horizons. Hold extra safety stock where the error is large, not across the whole shop.
Five mistakes that make a forecast worse
- Forecasting sales, not demand. Stockout weeks teach the forecast to order less, so the shortage repeats.
- Leaving one-off orders in the history. One funeral order of 240 bottles raised this shop's 12-week average by 20 bottles a week.
- Forecasting in money. A price rise looks like growth, and you order bottles nobody asked for.
- Adjusting by feel. An adjustment with no written reason can't be checked, so nobody learns from it.
- Never scoring the forecast. Without the error, you can't tell a good method from a lucky month.
Where ArmPOS fits: the history, not the forecast
ArmPOS does no demand forecasting. It records every sale by product and date, the history a forecast needs. It also warns you when stock falls to the level you set.
- Export the weekly sales. Open Reports › Product sales report, pick the dates, search the product and click EXCEL. Each line has the Date, the Qty sold and the Total (Sales reports).
- Pick the products to forecast. Top Selling Products ranks products by how many sales they appear in, with the money each one brought.
- Find the weeks to clean. In Stock Report, Reports on a product's row opens its sales, purchases, returns, transfers and adjustments (Stock reports). A late purchase often explains a low week.
- Mark the discounts. The Discount Summary Report lists each sale's discounts, with a chart over time. On plans with Promotions, its Usage Report lists each use by date (Promotions and promo codes).
- Turn the forecast into a warning. Enter your reorder level as the product's Stock Alert (Add your first products). Product Quantity Alerts then lists everything at or below it (Reports to check every day).
- Score the forecast each week. Run the Product sales report for the week just ended, and put the actual sales beside your forecast in the spreadsheet.
On plans with AI Reports, the fixed question This Week vs Last Week (Insights) says what changed and by how much (AI reports). It compares weeks already sold; it doesn't forecast.
Frequently asked questions
What are the main types of demand forecasting?
Short term (days to a few months), medium term (months to a year) and long term (a year or more). Each can be quantitative, built on sales history, or qualitative, built on judgement.
What is the difference between demand forecasting and demand planning?
Forecasting estimates what customers will buy. Planning decides what to buy, make or promote to meet it, within your cash, space and supplier lead times.
Can a small shop forecast demand without special software?
Yes. A spreadsheet with weekly sales per product, a cleaned baseline and a column for the error is enough for a shop's top sellers.
What is a good forecast accuracy?
One that beats the naive forecast on your own products, with no lasting bias. In the example above, the cleaned forecast missed by 5.8%; a slow seller will miss by far more.
Does ArmPOS forecast demand?
No. It keeps your sales history by product and date and exports it to Excel. It alerts you when stock falls to the level you set from your forecast.
A forecast is only as good as the sales history behind it. Create a free ArmPOS account and every sale is kept by product and date, ready to export each week. Shops in Ghana can start from ArmPOS in Ghana.
