Background
Anyone can recite the MAPE formula. What separates a planner from a spreadsheet is knowing why the business should care about each metric, and when each one lies to you. These are my working notes on the eight forecast KPIs that come up constantly in demand planning — with the formula, the business meaning, and the trap hidden inside each one.
Notation used throughout: $A_t$ = actual demand in period $t$, $F_t$ = forecast for period $t$, $n$ = number of periods.
MAPE — the metric everyone knows (and misuses)
$$MAPE = \frac{1}{n} \sum_{t=1}^{n} \left\vert \frac{A_t - F_t}{A_t} \right\vert \times 100%$$
What it tells the business: on average, how far off were we, as a percentage? It’s the easiest error metric to explain to a non-technical stakeholder, which is exactly why it’s everywhere.
The trap: it divides by actual demand, so it breaks completely on any period where $A_t = 0$. It also punishes over-forecasting harder than under-forecasting — an over-forecast can produce an error above 100%, while an under-forecast error tops out at 100%. A planner optimising for MAPE will quietly learn to forecast low.
Use it for: high-volume products with steady, non-zero demand, and any report headed to people who don’t want to see a formula.
WMAPE — MAPE that respects volume
$$WMAPE = \frac{\sum_{t=1}^{n} \vert A_t - F_t \vert}{\sum_{t=1}^{n} A_t}$$
What it tells the business: the same “how far off” question, but weighted by volume, so a 300% miss on an item that sells 4 units a year doesn’t wreck the scorecard for a category that’s otherwise well forecast.
The trap: the weighting cuts both ways. A low-volume item that’s strategically critical — or carries most of the margin — can have terrible accuracy that WMAPE hides completely behind the big movers.
Use it for: aggregate reporting across a diverse portfolio: a product category, a region, a whole network.
Bias — the direction of your error
$$Bias = \frac{1}{n} \sum_{t=1}^{n} (A_t - F_t)$$
What it tells the business: whether the errors lean one way. Consistently positive means you’re under-forecasting (stockouts, expedited freight). Consistently negative means over-forecasting (dead stock, tied-up capital).
The trap: positive and negative errors cancel. A bias of zero doesn’t mean a good forecast — it can mean you were badly wrong in both directions in equal amounts.
Use it for: detecting systemic behaviour, like a sales team that always over-promises. Never read bias alone; always pair it with an absolute metric like MAD or MAPE.
MAD — error in physical units
$$MAD = \frac{1}{n} \sum_{t=1}^{n} \vert A_t - F_t \vert$$
What it tells the business: the average miss in actual units — “we’re off by 500 bolts a month.” That’s a number you can plan warehouse space and safety stock around.
The trap: it doesn’t scale across products. Being 50 units off is a disaster on an item selling 100 a month and a rounding error on one selling 10,000.
Use it for: SKU-level operational planning. If you’re managing long overseas procurement lead times, MAD gives you the unit buffer that feeds directly into safety stock calculations.
RMSE — the metric that hates big misses
$$RMSE = \sqrt{\frac{1}{n} \sum_{t=1}^{n} (A_t - F_t)^2}$$
What it tells the business: the standard deviation of forecast errors. Squaring before averaging means one huge miss hurts far more than several small ones.
The trap: the output isn’t intuitive for business audiences. It’s a diagnostic tool, not a boardroom number.
Use it for: evaluating ML forecasting pipelines — it’s a standard loss function in libraries like scikit-learn and XGBoost. In business terms, reach for RMSE when a single massive error costs disproportionately more than many small ones.
Forecast Value Add — does each step actually help?
$$FVA = \text{Accuracy of process step} - \text{Accuracy of naive baseline}$$
What it tells the business: whether a step in the forecasting process improved anything. The naive baseline is usually “next period equals this period.” If a planner manually overrides the statistical forecast, FVA shows whether that override added value or destroyed it.
The trap: it demands disciplined tracking of every forecast version (baseline, statistical, planner override, consensus). And politically, it can be awkward — FVA frequently reveals that human overrides make the forecast worse.
Use it for: justifying the ROI of planning software, and identifying which manual touches in your S&OP process should be eliminated.
Forecast Accuracy % — the positive spin
$$Accuracy = \max(0,\ 100% - MAPE)$$
(Many companies substitute WMAPE here.)
What it tells the business: the same information as MAPE, framed as a win instead of a miss.
The trap: there’s no industry-standard definition. Your 85% and a competitor’s 85% could be built on entirely different formulas, so the comparison is meaningless without the underlying maths.
Use it for: executive dashboards and vendor scorecards — anywhere a positive framing is expected. Just document which formula sits underneath it.
Tracking Signal — the smoke alarm
$$TS = \frac{\sum (A_t - F_t)}{MAD}$$
What it tells the business: whether errors are random noise or a systematic drift. When the signal breaches a threshold (typically ±4), the forecast model is broken and needs review.
The trap: one anomaly — say, a single bulk order from one customer — can trip the alarm without any real shift in baseline demand.
Use it for: automated exception alerts. It’s the early-warning system that tells a planner to review an SKU before it turns into a stockout or an overstock write-down.
Quick reference
| Metric | Best for | Watch out for |
|---|---|---|
| MAPE | Stakeholder communication, steady demand | Zero-demand periods; penalises over-forecasting |
| WMAPE | Portfolio-level reporting | Hides poor accuracy on strategic low-volume items |
| Bias | Detecting systematic over/under-forecasting | Errors cancel out; never read alone |
| MAD | SKU-level planning, safety stock | Doesn’t scale across products |
| RMSE | ML model evaluation, big-miss sensitivity | Not intuitive for business audiences |
| FVA | Auditing the forecasting process itself | Heavy tracking overhead; political friction |
| Accuracy % | Dashboards and scorecards | No standard definition — always state the formula |
| Tracking Signal | Automated exception alerts | False alarms from one-off anomalies |
Takeaway
No single metric is honest on its own. MAPE without bias hides direction. Bias without MAD hides magnitude. WMAPE without SKU-level review hides your strategic items. The planners who get hired — and the ones who keep inventory off the write-down list — are the ones who know which metric to reach for, and which lie it’s about to tell them.