Sales forecast: method, horizon, error
Builds a forecast from sales history using four methods, shows each method’s error on past data and suggests which is more reliable for your series.
At a glance
- What question does it answer?
- How much you will sell in the coming periods and how reliable the forecast is.
- What data is needed?
- Sales history by period, at least two years for seasonal apparel.
- How to read the result?
- Compare the error (WAPE) of the four methods on your data: the model suggests which is more accurate. A visible bias means the method systematically overstates or understates. Exclude anomalous months with the tick. Example on the default values: 24 months of history, the seasonal method, a 6-period horizon: the forecast is RUB 71,729,429, on average RUB 11,954,905 per period, 11.5% above the same periods a year ago. The method’s error on history (WAPE) is 8.3% and the bias −0.9%: an estimate on past data, not a guarantee for the future. If the method’s error is above your tolerance, forecast by category or by channel. Next: “Annual budget by month” and “S&OP: forecast bias, error and the balance of demand and supply”.
Why and how to reason
A forecast is not a prophecy but a continuation of patterns from history. The method is chosen by the nature of the data. If sales are stable, a moving average is enough. If there is a direction (growth), a linear trend is needed. If the latest dynamics matter, smoothing. And if the series is seasonal, as in apparel, trend and seasonality is used: forecast = trend line x the period’s seasonal index.
The seasonal index shows how many times a period is usually above or below the trend: December’s is above 1, January’s below. The trend is taken from deseasonalized data so the season is not mistaken for growth.
Method quality is tested blind: a forecast for past periods is compared with actuals. WAPE = sum of errors / sum of sales.
Example. Over two years revenue grew from RUB 133.6 m to RUB 151.0 m (+13%). The forecast for 6 months is RUB 71.7 m against 64.3 for the same months a year ago: +11.5%. WAPE error 8.3%.
When to use it
Use it for the sales plan, purchase budget and cash. For apparel take at least two full years so the method can see the season. It is unsuitable for new brands without history and for periods with one-off events (a store closure, a pandemic, a promo spike): fix or exclude such periods first. The manual adjustment is for when you know something that is not in the history: a new store, a new promotion.
How to read the result
Look at the error: a WAPE up to 10% for monthly data is usually considered good, but check against your own. Bias shows a systematic tilt: negative means the method slightly overstates. Compare methods: if a simple average errs no more than a complex one, take the simple one. Compare with last year (+11.5%): does it agree with what you know about the company? Decision: use the forecast as a base, add an adjustment for known events, and reconcile plan with actuals every month. It is worth writing the forecast assumptions next to the number so you can later see the cause of an error.
Common mistakes
Typical trap: season length not equal to 12 for monthly data; history with anomalies; too short a series; trusting the error measured on very few test periods. Also: using the revenue forecast as the purchase plan without accounting for stock and markdowns. Also: recomputing the forecast from scratch each time and losing comparability.
What data is needed?
- Sales history by period, at least two years for seasonal apparel.
- The horizon, method and, if needed, a manual adjustment (new stores, price rises).
- Where to take it: sales history by period from the sales report in 1C or the ERP excl. VAT, in equal periods (months), at least two years for seasonal apparel; mark periods with no stock and with abnormal promotions. Paste the history first, then choose the method and horizon, and enter an adjustment only for known events.
Inputs
| Field | Unit | Value in the example | Meaning |
|---|---|---|---|
| Forecast method | Trend and seasonality | For apparel with a strong season you need at least two years of history: then “Trend and seasonality” fits. For stable items smoothing or an average is enough. | |
| Periods to forecast | periods | 6 | How many periods ahead the forecast is built; the seasonal method needs a history of at least 12 periods. |
| Season length | periods | 12 | 12 for months in a year, 52 for weeks. |
| Average window | periods | 3 | How many recent periods the moving average averages; used only by that method. |
| Smoothing factor | 0–1 | 0.4 | The closer to 1, the more the forecast follows the latest sales. |
| Manual adjustment | % | 0 | For example +10 when new stores open or −5 when the range shrinks. Applied to the whole forecast. |
How it is calculated
| Step | Formula |
|---|---|
| Trend and seasonality | Forecast = (trend line on deseasonalised data) × the period’s seasonal index |
| Seasonal index | Average ratio of same-period sales to the trend line, normalised to a mean of 1 (5 refinements) |
| Smoothing | Level = α × latest sales + (1 − α) × previous level |
| WAPE | Σ |actual − forecast| ÷ Σ actual; each period is forecast from data before it only |
Worked numbers
Default values from the app; in the interactive version you enter your own.
- Forecast for 6 periods
- 71.7 m ₽
- Average per period
- 12 m ₽
- Versus the same periods a year ago
- 11.5 %
- Method error (WAPE)
- 8.3 %A blind test: each of the last 10 periods is predicted only from data before it. Sum of deviations ÷ sum of sales. The colour by 20 and 40% thresholds is a rule of thumb, not a standard.
- Percentage error (MAPE)
- 8.5 %The mean over periods of |actual − forecast| ÷ actual; periods with zero actual are left out. Small low-sales periods inflate the measure: compare it with WAPE.
- Bias
- −0.9 %Plus means the method overstates sales, minus understates. A visible bias means the method does not fit.
- Actual
- Forecast
Data as a table
| Measure | Actual | Forecast |
|---|---|---|
| Jan year 2 | 9.14m | — |
| Feb year 2 | 8.29m | — |
| Mar year 2 | 11.1m | — |
| Apr year 2 | 11.4m | — |
| May year 2 | 12.4m | — |
| Jun year 2 | 12m | — |
| Jul year 2 | 9.75m | — |
| Aug year 2 | 11.3m | — |
| Sep year 2 | 14m | — |
| Oct year 2 | 14.8m | — |
| Nov year 2 | 17.9m | — |
| Dec year 2 | 18.9m | 18.9m |
| Period +1 | — | 10.4m |
| Period +2 | — | 9.53m |
| Period +3 | — | 11.8m |
| Period +4 | — | 13.3m |
| Period +5 | — | 13.7m |
| Period +6 | — | 12.9m |
Accuracy of the methods: blind test on the latest periods
| Method | WAPE | Bias | Forecast for the horizon |
|---|---|---|---|
| Trend and seasonality ✓ | 8.3 % | −0.9 % | 71.7 m ₽ |
| Linear trend | 23 % | 6.3 % | 96.2 m ₽ |
| Smoothing | 18.9 % | −2.9 % | 99.9 m ₽ |
| Moving average | 21.8 % | −2.5 % | 103.2 m ₽ |
The chosen method gives the lowest error on your data: 8 %. It is no guarantee for the future: check the forecast against the promo and delivery plan.
Nuances and limits
- The forecast does not know about promotions, price or range changes: use the manual adjustment.
- The error is computed on past data: it is an estimate, not a guarantee.
- Months with stock-outs understate the history: demand was above sales there, mark them or replace with an estimate.