Forecast bias

Forecast bias is the direction of forecast error. A forecast that's chronically too high has positive bias. One that's chronically too low has negative bias. A good forecast has roughly zero bias — its over-shoots and under-shoots cancel out across many periods.

This is the diagnostic that separates "wrong on average" from "wrong on average in one direction." Both hurt; the second is worse because it compounds.

How forecast bias is calculated

The simplest measure is the mean error:

Mean Error = Σ(Forecast − Actual) / n

Sign matters. Positive mean error = forecast over-shoots. Negative = forecast under-shoots.

A more useful version is the bias percentage, normalised against the actual demand:

Bias % = Σ(Forecast − Actual) / Σ(Actual) × 100

A SKU running at +15% bias means the forecast over-shoots actual demand by 15% on average. That 15% extra is being bought, stored, and eventually marked down.

Bias is not the same as accuracy

Forecast accuracy measures how big the errors are. Forecast bias measures which way they lean. A forecast can be:

• High accuracy, low bias → the goal
• High accuracy, high bias → consistently wrong by a small margin in one direction (less common; usually a calibration miss)
• Low accuracy, low bias → noisy but unbiased — forecast is messy but not systematically off
• Low accuracy, high bias → noisy and leaning — the worst combination, usually a planning issue not a model issue

MAPE and accuracy metrics will tell you the magnitude of error. Bias tells you which direction those errors point.

What positive bias looks like in operations

Symptoms of a forecast that runs hot (over-predicts):

• Closing stock keeps growing relative to revenue
• Frequent end-of-season markdowns
• Increasing DIO over multiple quarters
• Ongoing LCNRV write-downs at audit (valuing stock at below what you paid for it)
• Buyers feeling like they "always over-order"

The structural causes: optimistic launch forecasts that aren't corrected after early data, sandbagging by sales teams whose targets reward over-supply, or marketing forecasts that assume every campaign will outperform.

What negative bias looks like

Symptoms of a forecast that runs cold (under-predicts):

• Stockouts on supposedly-forecast SKUs
• Frequent emergency reorders and air freight
• Backorder rates climbing
• Customer complaints about availability
• Buyers asking "how did we miss that?"

Structural causes: conservative finance teams pushing forecasts down to protect margin, persistent under-estimation of new-channel growth (especially Amazon and TikTok-driven viral moments), or simply not refreshing the forecast often enough.

Bias by SKU vs bias in total

A total bias near zero can hide bad SKU-level bias. Half the SKUs running +20% and the other half running −20% nets to zero on aggregate but represents two real problems: cash trapped in over-forecast SKUs and lost sales from under-forecast ones.

Always measure bias per SKU, then roll up to category, then to total. The aggregate number is reassuring; the SKU view is operational.

Fixing chronic bias

Bias doesn't fix itself. The forecast model doesn't know it's biased — only the comparison to actuals reveals it.

Three steps:

1. Measure it monthly per SKU. Build the mean-error and bias-% tables. Sort to find the worst offenders.
2. Apply a bias correction. If a SKU has been running at +15% for six months, dial the next forecast down 15% as a starting point.
3. Investigate the structural cause. A persistent bias usually has a human explanation — incentive misalignment, optimism in marketing, conservatism in finance. Fix the cause, not just the number.