Forecast Troubleshooting

Demand Planning and Forecasting Mistakes: How to Catch a Bad Number Before It Ships

The mistakes that quietly wreck a demand plan are rarely the formula itself. They are wrong units, hidden bias, and dirty history. Here is how to spot each one and the number that proves you fixed it.

The most expensive forecasting mistake is trusting a good-looking accuracy number that hides directional bias. Symptom: MAPE reads a respectable 18 percent but you are always short at month end. Root cause: absolute error metrics cancel plus and minus swings, so a plan that runs 12 percent low every period still scores well. The fix is to run Forecast Bias alongside Forecast Accuracy and hold bias inside plus or minus 5 percent of demand. If mean bias exceeds that band for three consecutive months, the model is structurally low or high, not just noisy, and no safety stock tuning will save it.

Mixing units and time buckets is the classic silent error. Symptom: your supply plan overstates need by roughly 4x with no obvious cause. Root cause: forecasting in eaches while the BOM and MPS run in cases of 4, or comparing a weekly forecast against monthly actuals so a 4.33 week month inflates every ratio. The fix is one canonical unit and one canonical bucket, declared before anyone touches Forecast Accuracy. Convert 10,000 eaches to 2,500 cases explicitly, and reconcile 52 forecast weeks to 12 actual months, or your error percentage is measuring the calendar, not the model.

Averaging percentage errors across fast and slow movers produces a number that describes nothing. Symptom: portfolio MAPE looks fine at 20 percent, yet A items overshoot warehouses while C items stock out. Root cause: unweighted MAPE lets a single slow SKU forecasting 2 units versus 1 actual post a 100 percent error that swamps a high volume line off by 3 percent. The fix is volume weighting: multiply each SKU error by its unit share, so a 10,000 unit line off 3 percent carries 40 times the weight of a 250 unit line. Run Demand Variability first to segment SKUs by coefficient of variation before you pool anything.

Sizing safety stock from a gut feel instead of measured error burns cash on both ends. Symptom: 60 days of finished goods on hand and still a 4 percent stockout rate. Root cause: buffers set to a flat two weeks ignore that error scales with the standard deviation of forecast miss, not with average demand. The fix is to drive the Inventory Buffer from Forecast Error calculator: a SKU with a forecast error standard deviation of 200 units and a 95 percent service target needs about 1.65 times 200, or 330 units, not an arbitrary round number. Recompute quarterly, because error distributions drift as products age.

Feeding raw shipment history into a statistical model without cleaning it teaches the model to repeat your past failures. Symptom: the forecast keeps a phantom demand spike alive for a year. Root cause: a one-time 3,000 unit promo or a stockout that capped true demand at the shelf gets read as baseline. The fix is outlier scrubbing before fitting: cap any period beyond 3 standard deviations, restore censored demand where you sold out, and flag promo weeks. A single uncorrected 3,000 unit event on a 500 unit baseline can lift the annual forecast 5 percent and misstate every downstream Supply Demand Gap.

Treating the S&OP cycle as a monthly report instead of a decision gate lets errors age unchallenged. Symptom: gaps surface only after commitments are locked, so every fix is expedite freight. Root cause: a cycle that takes 25 days end to end means you are always reacting to demand signals five weeks stale. Measure it with S&OP Cycle Time and target a 10 to 12 day close so the consensus number is fresh. Pair it with Demand Plan Attainment tracking, because a plan you never grade is a plan nobody trusts by the third miss.

Ignoring the capacity side of the plan makes a perfect demand number useless. Symptom: the forecast is signed off but the line still misses ship dates. Root cause: nobody checked load against available hours, so a validated 40,000 unit demand plan hit a cell rated for 34,000. The fix is to run MPS Load and the Capacity Demand Gap together every cycle, flagging any period above 90 percent utilization as a risk. A 6,000 unit shortfall found at planning costs a subcontract quote, the same shortfall found at execution costs premium overtime at 1.5 times labor rate.

The last trap is never pricing the error, so improvement never gets funded. Symptom: leadership treats forecasting as free and staffs it thin. Root cause: no line item connects a 20 percent error to dollars. The fix is Forecast Error Cost: multiply your miss by the marginal cost of being wrong, roughly holding cost near 25 percent annually on the overs plus lost margin on the unders. On 100,000 units at 15 dollars, cutting error from 25 to 18 percent can free six figures, which is the business case for a better process.

Published 2026-07-01.