Forecast accuracy calculator

Forecast accuracy formulas

For every row, the error is the forecast minus the actual. The four measures summarize those errors in different ways:

WAPE = Σ |actual − forecast| ÷ Σ |actual| × 100%

MAPE = average of ( |actual − forecast| ÷ |actual| ) × 100%

Bias = Σ (forecast − actual) ÷ Σ actual × 100%

MAE = Σ |actual − forecast| ÷ number of rows

• WAPE is the total absolute error as a share of total actual sales. Large rows weigh more, in proportion to their volume.
• MAPE averages the percentage error of each row, so every row counts the same whatever its size. Rows with an actual of zero have no percentage error and are left out.
• Bias keeps the sign. Positive means the forecast runs high (excess stock); negative means it runs low (stockouts). Errors in opposite directions cancel.
• MAE is the average error in units. It is easy to read for one SKU but cannot be compared between SKUs that sell very different volumes.

Many reports also quote forecast accuracy = 100% − WAPE. The calculator shows it, floored at 0% when WAPE is above 100%.

Why WAPE is the better default at SKU level

At SKU level, many items sell a handful of units a month. A forecast of 10 against an actual of 4 is a 150% error on 6 units; a forecast of 1,100 against 1,200 is an 8.3% error on 100 units. MAPE averages the percentages, so the 6-unit miss counts for more than the 100-unit one. WAPE weights each row by its volume, so it measures the error in the units you actually buy and stock.

• MAPE explodes on low-volume SKUs. One unit of error on a SKU that sold 2 is 50%. A long tail of slow sellers can push MAPE above 100% even when the units that matter are forecast well.
• MAPE is undefined for zero actuals. Intermittent SKUs have months without sales. A forecast of 3 against 0 cannot be expressed as a percentage, so MAPE has to drop the row, although those 3 units are real overstock. WAPE still counts them.
• MAPE favors forecasting low. An under-forecast can be at most 100% wrong (forecasting 0), while an over-forecast has no upper limit. A team judged on MAPE learns to forecast low and runs out of stock.
• WAPE adds up across levels. Sum the absolute SKU errors and the SKU actuals of a category and you get that category's SKU-level WAPE, consistent with the SKU and company figures.

MAPE is fine when every series is large and of similar size, such as monthly sales per region. For a catalog with a long tail, report WAPE and bias.

Worked example

Six SKUs, one month. The calculator above is loaded with these rows.

• WAPE = 214 ÷ 2,174 = 9.8%.
• MAPE = (8.3 + 7.5 + 20.0 + 75.0 + 150.0) ÷ 5 = 52.2%, with C-600 left out because its actual is 0.
• Bias = (2,188 − 2,174) ÷ 2,174 = +0.6%: almost no lean overall.
• MAE = 214 ÷ 6 = 35.7 units.

MAPE says the forecast is off by half; WAPE says it is off by a tenth. By volume, WAPE is right: A-100 and A-200 make up 92% of the units sold and were forecast within 8.3%. C-400 and C-500, which sold 24 units between them, supply 225 of the 260.8 percentage points behind the MAPE.

Read bias next to WAPE

WAPE tells you how big the error is; bias tells you which way it leans. A WAPE of 30% with bias near zero is noise, which safety stock has to absorb. A WAPE of 30% with a bias of +25% is a forecast that is systematically too high: fixing it cuts stock without hurting service. Over-forecasting shows up as excess stock and markdowns; under-forecasting as stockouts and lost sales. To size a buffer for the noise, use the safety stock calculator with the standard deviation of forecast error.

Common mistakes

• Measuring only at an aggregated level. Errors of different SKUs cancel when you forecast or measure the total, so a category can look accurate while its SKUs are not. Measure at the level you order: SKU, and location if you stock several.
• Measuring the wrong forecast. The forecast that drove the order is the one made a lead time ahead. For a three-month lead time, compare actuals with the forecast made three months earlier, not last week's update.
• Counting stockout months as actual demand. Zero sales in a month with zero stock is not zero demand. It penalizes a forecast that was right and rewards one that predicted the stockout.
• No benchmark. Compare with a naive forecast, such as the same month last year or a three-month average, on the same SKUs and months. A method that does not beat it adds no value.

A forecast for every SKU

The calculator measures a forecast you already have. Varox builds one: a monthly quantity forecast for every SKU, with a base method picked per SKU from the shape of its history. Stable history gets a statistical model, short history borrows the level and seasonal shape of its parent group, and inactive SKUs stay near zero. Each forecast shows which model produced it, and the Delivery Planner orders from it directly.

Questions

What is a good forecast accuracy?

There is no universal target. Error falls as volume rises, so a category total is always more accurate than the SKUs inside it, and a fast seller more accurate than a long-tail item. Compare your forecast with a naive one (last year's same month, or a moving average) for the same SKUs and months, and track the trend over time instead of chasing an industry figure.

What is the difference between WAPE and MAPE?

MAPE averages the percentage error of each row, so a SKU selling 4 units counts as much as one selling 4,000, and rows with zero actual sales cannot be used. WAPE divides the total absolute error by total actual sales, so each row counts in proportion to its volume.

Is WAPE the same as weighted MAPE?

Yes, in the usual definition. WAPE is MAPE with each row weighted by its actual volume, and it equals the mean absolute error divided by the mean actual, which is why it is also called the MAD/mean ratio.

How do I calculate forecast bias?

Subtract total actual from total forecast and divide by total actual. A positive result means the forecast runs high, a negative one that it runs low. Errors with opposite signs cancel, so a bias near zero does not mean the forecast is accurate; read it next to WAPE.

How do I measure accuracy when actual sales are zero?

Use WAPE or MAE, which still count the error on a row with zero actual sales. MAPE is undefined for those rows, so it has to leave them out, even though a forecast of 3 against 0 is 3 units of real overstock. For SKUs that sell rarely, also check whether a zero was missing demand or a stockout.

What does forecast accuracy as a percentage mean?

Most reports define it as 100% minus WAPE, so a WAPE of 25% is reported as 75% forecast accuracy. When WAPE is above 100% the result would be negative; this calculator shows 0% in that case.