How to Measure Forecast Accuracy (MAPE, MAD, Bias)

Forecast accuracy measures how close your demand forecast came to what actually sold. The three standard metrics are MAPE (the average miss as a percentage), MAD (the average miss in units), and bias (the average signed error showing whether you forecast consistently high or low).

You cannot manage a forecast you do not measure, and "the forecast felt about right" is not a measurement. The three metrics below turn a pile of forecast-versus-actual numbers into a score you can track, compare across products, and act on. This post defines each one, walks a worked example by hand, explains bias and the tracking signal, and gives realistic accuracy targets by product type. It is educational and names no products.

What is forecast accuracy, and why measure it?

Forecast accuracy is a measure of how far your predicted demand landed from actual demand, expressed so you can track it over time. It matters because almost every downstream plan rides on the forecast. Set safety stock too low relative to your true error and you stock out; set it too high and you drown in finished goods inventory. The forecast error is the raw variability that sizing decisions are supposed to absorb, so you have to quantify it before you can size anything sensibly.

There is an important split hiding inside "accuracy." Two forecasts can miss by the same average amount for very different reasons. One misses randomly, sometimes high, sometimes low, with the errors canceling out. The other misses in one direction every time. The first is imprecise but honest; the second is biased, and bias is the more dangerous problem because it quietly builds up excess or shortage month after month. Good measurement separates the two, which is why you track both an error-size metric and a bias metric, not just one number.

How do you calculate MAPE, MAD, and bias?

You calculate all three from the same building block: the forecast error for each period, which we will define as forecast minus actual. A positive error means the forecast was too high that period; a negative error means it was too low. From those errors, MAD is the average of their absolute values, MAPE is the average of their absolute values as a percentage of actual, and bias is the plain average of the signed errors. Take four periods of one SKU.

Now the three metrics fall straight out of the totals:

• MAD = total absolute error ÷ number of periods = 45 ÷ 4 = 11.25 units. On average, the forecast missed by about 11 units in either direction.
• MAPE = average of the absolute percentage errors = 42.3% ÷ 4 = 10.6%. On average, the forecast was off by about 11% of actual demand.
• Bias = total signed error ÷ number of periods = −5 ÷ 4 = −1.25 units. The small negative value means the forecast ran slightly below actual on balance, a mild tendency to under-forecast, but it is tiny next to the 11.25-unit MAD, so this forecast is essentially unbiased.

The pairing is what makes this useful. MAPE tells you the size of the error in percentage terms, so you can compare a high-volume SKU against a low-volume one. MAD keeps the error in real units, which is what feeds safety-stock math. Bias tells you the direction catching the silent one-way drift that a size-only metric like MAPE would hide, because MAPE treats a plus-10 and a minus-10 as the same 10.

What is the tracking signal, and how does it catch bias?

The tracking signal is the running total of signed forecast error divided by MAD, and it is the standard early-warning that a forecast has gone biased. A single period's bias can be luck; a tracking signal that keeps climbing means the errors are stacking up in one direction and the forecast has a systematic problem. In the worked example, the cumulative error is −5 and MAD is 11.25, so the tracking signal is −5 ÷ 11.25 = −0.44, comfortably inside the usual control band of roughly plus or minus four. If it drifted past that band, you would stop and investigate before the bias did real damage.

How do you measure forecast accuracy across the whole catalog?

Measuring one SKU by hand is easy; the discipline is doing it consistently across hundreds of items so the number is comparable and actionable. The order below keeps it honest.

1. Freeze the forecast before the period. Lock the forecast you are grading before actuals come in, or you are scoring hindsight, not a forecast.
2. Pick one error convention and hold it. Decide that error is forecast minus actual (or the reverse) and use it everywhere, so bias signs mean the same thing on every report.
3. Compute MAD, MAPE, and bias per item. Score each SKU on its own, since a plant-wide average hides the items that are actually hurting you.
4. Weight the roll-up by volume. When you aggregate, use a volume-weighted MAPE so a tiny SKU with a huge percentage error does not swamp the number for the products that move real money.
5. Watch the tracking signal for drift. Flag any item whose tracking signal breaks the control band, because that is a systematic bias worth fixing, not random noise.
6. Feed the result back into safety stock. Use the measured error, not a flat percentage, to size buffers, so the items you forecast worst carry the protection and the items you forecast well do not.

That last step is the payoff. Accuracy measurement is not a report card you file; it is the input that sizes safety stock sets the master schedule, and decides how much finished goods you dare to run lean. It ties directly into disciplined production planning and the master production schedule downstream.

What is a good forecast accuracy target?

A good target depends entirely on the product, and any single number quoted without that context is a red flag. High-volume, stable staples are forecastable to a tight MAPE; short-lifecycle, seasonal, or promotional items are not, and holding them to the same standard just teaches planners to game the metric. The realistic ranges below come from published demand-planning benchmarks and should be read as broad bands, not promises.

For intermittent, low-volume items, MAPE itself misbehaves: a forecast of 2 against an actual of 1 is a 100% error on a rounding difference. That is why volume-weighted MAPE exists, and why the long tail of slow movers is usually better managed with simple min-max rules than chased with a forecast. Deciding which items even deserve a statistical forecast is a segmentation question that connects to broader lean manufacturing thinking about where to spend effort.

What do the standards and data say?

Context from standards bodies and primary references:

• MAPE, MAD, bias, and the tracking signal are defined in the forecasting and demand-management sections of the body of knowledge maintained by the Association for Supply Chain Management (ASCM/APICS).
• Academic operations texts present the same formulas: for example, the open supply-chain management materials published by the University of Arkansas lay out MAD, MAPE, and tracking-signal calculations with worked examples.
• The scale of the demand this governs is large: the Bureau of Labor Statistics reports roughly 13 million manufacturing jobs in the United States, at plants whose production plans all ride on some demand forecast.

The consistent message: the formulas are standard and simple; the value is in measuring consistently and feeding the result back into how much stock you hold.

Where forecast measurement breaks down

Forecast measurement breaks down not on the math but on the bookkeeping. The formulas take seconds; the hard part is having a clean, frozen record of what you forecast and what actually shipped, per item, period after period. In most plants the forecast lives in one spreadsheet, actuals live in the ERP, promotions and one-off orders are remembered by whoever ran them, and reconciling all of it into a trustworthy accuracy number is a monthly chore that slips. So the metric goes stale, and buffers get set by feel instead of by measured error.

Harmony is an AI-native layer that connects machines, software, and paperwork into one operational record, with no rip-and-replace, so the forecast, the shipment history, and the events that moved demand stop living in separate places and become one record you can actually score. AI search returns cited answers across those records, so a planner can ask which SKUs are biased high this quarter or how last month's accuracy compares to the same month last year and get a real answer instead of a spreadsheet rebuild. It is the same paper-to-digital move Harmony makes elsewhere on the floor (see the CLS case study). The same clean signal that lets you predict demand better also powers AI demand forecasting and Harmony's digital workflows keep the accuracy number current instead of a quarter behind.