10.4 Managing Forecast Accuracy: Bias, Weighted Error & Forecast Value Added

Managing Forecast Accuracy: Bias, Weighted Error & Forecast Value Added

Computing MAD, MAPE, and a tracking signal is arithmetic. Managing accuracy is a supply management discipline: deciding what target applies to which item, distinguishing bias from noise, proving which process steps actually help, and converting the residual error into an inventory number. This is where Exam 2 separates planners from calculators.

---

Accuracy vs. Error — Define the Denominator

Organizations report "forecast accuracy" in incompatible ways, and the exam exploits it. The two common conventions:

$\text{Accuracy}_1 = 1 - \frac{|A - F|}{A} \qquad \text{Accuracy}_2 = 1 - \frac{|A - F|}{F}$

They give different answers. Actual = 80, Forecast = 100:

• Denominator = actual: error is $20/80 = 25%$, accuracy 75%.
• Denominator = forecast: error is $20/100 = 20%$, accuracy 80%.

Worse, dividing by actual demand explodes when actual approaches zero and is undefined at zero, which is why MAPE is unusable for intermittent items. A portfolio metric must state its denominator convention explicitly, or the number means nothing.

---

Bias vs. Random Error — the Distinction That Matters Most

The core exam point: safety stock is designed to absorb random error, not bias. Adding safety stock to a chronically under-forecast item buys temporary relief while the underlying bias keeps growing. The tracking signal exists precisely to catch this.

Where Organizational Bias Comes From

• Sales teams under-forecast when forecasts set quotas, because a low target is easier to beat.
• Sales teams over-forecast when forecasts secure allocation, because a high number reserves supply.
• Marketing over-forecasts new products because the business case depended on optimistic volumes.
• Operations under-forecasts to avoid being held to a capacity commitment.
• Finance flexes the forecast toward the budget rather than toward the likely outcome.

Each bias is rational for the individual and destructive for the organization. The structural remedies are to separate the forecast from the target (an unbiased best estimate is not a commitment), to measure and publish bias by contributor, and to reward forecast accuracy rather than forecast optimism.

---

Portfolio Metrics: Weighted MAPE

A simple average of item-level MAPE gives a $200-per-year C item the same influence as a $2-million A item — and because low-volume items have the highest percentage errors, the portfolio number is dominated by items nobody cares about.

Weighted MAPE (WMAPE) fixes this by weighting by volume or value:

$\text{WMAPE} = \frac{\sum |A_t - F_t|}{\sum A_t}$

Worked comparison.

• Simple average MAPE = $(5.0 + 15.0 + 100.0)/3 = \mathbf{40.0%}$ — a disastrous-looking number driven entirely by an item worth 0.2% of volume.
• WMAPE = $670 / 11{,}020 = \mathbf{6.1%}$ — the number that actually reflects the inventory and service consequence.

Report WMAPE for the portfolio and item-level MAPE only within a comparable segment.

---

Forecast Value Added (FVA)

FVA measures whether each step in the forecasting process makes the forecast better than doing nothing. The benchmark is the naive forecast: next period equals this period's actual (or, for seasonal series, the same period last year).

$\text{FVA of a step} = \text{Error of the previous stage} - \text{Error after the step}$

A positive FVA means the step added value; a negative FVA means the step made the forecast worse and should be eliminated.

Worked FVA stairstep (WMAPE at each stage):

Reading this table is the exam skill. The statistical baseline and the consensus process both add value. The sales override destroys value — it makes the forecast measurably worse than the statistical model it replaced. The correct action is not to abolish sales input but to restructure it: require documented reasons for overrides, apply them only where the salesperson has genuine customer-specific information, and hold override authors accountable for their measured FVA.

FVA analysis regularly shows that the most senior and most expensive steps in a forecasting process subtract value. That finding is the point of the technique.

---

Setting Accuracy Targets by Segment

A single organization-wide accuracy target is meaningless because achievable accuracy depends on the demand pattern, not on planner effort.

Exam anchor: the right response to persistently high error on an inherently erratic item is not a better model. It is to change the supply response — shorter lead times, postponement, capacity buffers, or a collaborative arrangement with the customer that reveals demand earlier.

---

Converting Forecast Error into Safety Stock

This is the link between forecasting and inventory that Exam 2 expects you to make explicitly. Safety stock should be sized on the standard deviation of forecast error over the lead time, not on the historical variability of demand itself — because what must be buffered is what you failed to predict, not what happened.

When the error distribution is approximately normal, standard deviation and MAD are related by:

$\sigma \approx 1.25 \times \text{MAD}$

Worked example. A component has a forecast MAD of 40 units per week and a supplier lead time of 4 weeks, and the buyer targets a 95% cycle service level ($Z = 1.645$).

1. Convert MAD to standard deviation: $\sigma \approx 1.25 \times 40 = 50$ units per week.
2. Scale to the lead time: $\sigma_{LT} = 50 \times \sqrt{4} = 100$ units.
3. Safety stock: $SS = Z \times \sigma_{LT} = 1.645 \times 100 = \mathbf{165\ \text{units}}$.

Now the business case writes itself. Halving forecast error halves safety stock. If a forecasting improvement takes MAD from 40 to 20, safety stock falls from 165 to about 83 units — a permanent working capital release, repeated across every item in the portfolio. That calculation is how supply management converts forecasting investment into a number the finance function will fund.