Wiki V1.0
Powered by

# Forecast Accuracy

You can't manage what you don't measure, and demand forecasts are no exception. Forecast accuracy quantifies how close your demand plan was to what actually happened — and more importantly, it tells you where to focus improvement efforts. 📊

# Why It Matters

Poor forecast accuracy cascades through the entire business:

  • Inventory — Over-forecast and you build excess stock; under-forecast and you stock out.
  • Service levels — Customers don't care about your forecast process, they care about availability.
  • Financial planning — Revenue, margin, and cash flow projections all start with a demand number.
  • Supply operations — Production scheduling, procurement, and logistics capacity all depend on a reliable demand signal.

Improving forecast accuracy by even a few percentage points often delivers outsized returns in working capital and service.

# Key Metrics

# MAPE (Mean Absolute Percentage Error)

The most widely used forecast accuracy metric. Expresses error as a percentage of actual demand.

\displaystyle \text{MAPE} = \frac{1}{n} \sum_{i=1}^{n} \left| \frac{A_i - F_i}{A_i} \right| \times 100

Where A = actual demand and F = forecast. Simple and intuitive, but has a well-known flaw: it's undefined when actuals are zero and overweights low-volume items.

# Weighted MAPE (wMAPE)

Weights each item's error by its share of total volume, giving high-volume products proportionally more influence. This is often more useful than simple MAPE for business decisions.

\displaystyle \text{wMAPE} = \frac{\sum_{i=1}^{n} |A_i - F_i|}{\sum_{i=1}^{n} A_i} \times 100

Rule of thumb: If you can only track one accuracy metric, make it wMAPE. It avoids the distortions of simple MAPE and reflects the items that matter most to the business.

# Bias

Measures whether your forecast systematically over- or under-predicts demand. A forecast can have a low MAPE yet still be heavily biased — and bias is often more actionable than error.

\displaystyle \text{Bias} = \frac{\sum_{i=1}^{n} (F_i - A_i)}{\sum_{i=1}^{n} A_i} \times 100

Positive bias = over-forecasting (builds inventory). Negative bias = under-forecasting (risks stockouts). Target is zero.

# Tracking Signal

Detects when a forecast has drifted systematically out of control. It's the ratio of cumulative error to mean absolute deviation (MAD), functioning like a quality control chart for your forecast.

\displaystyle \text{Tracking Signal} = \frac{\sum_{i=1}^{n} (A_i - F_i)}{\frac{1}{n} \sum_{i=1}^{n} |A_i - F_i|}

When the tracking signal exceeds +/- 4 to 6 (thresholds vary by organization), the forecast model likely needs recalibration.

# Forecast Value Added (FVA)

Measures whether each step in the demand planning process — especially human overrides — actually improves accuracy. Compare accuracy at each stage against a naive or baseline forecast.

\displaystyle \text{FVA}_{\text{step}} = \text{Error}_{\text{before step}} - \text{Error}_{\text{after step}}

If FVA is negative, that process step is making the forecast worse. This happens more often than people expect, particularly with sales overrides that introduce optimism bias.

# Summary Table

Metric Formula What It Tells You When to Use
MAPE Mean of |(A-F)/A| Average percentage error per item Quick, intuitive reporting; avoid for intermittent demand
wMAPE Sum |A-F| / Sum A Volume-weighted error across a portfolio Default accuracy KPI; best for mixed portfolios
Bias Sum (F-A) / Sum A Directional tendency to over/under-forecast Root cause analysis; inventory health diagnostics
Tracking Signal Cumulative error / MAD Whether the forecast has drifted out of control Model monitoring; triggering recalibration
FVA Error before - Error after Whether a process step improves accuracy Process improvement; justifying overrides

# Where to Measure

# Aggregation Level

Forecast accuracy varies dramatically depending on where you measure:

  • SKU-location — Most granular, highest error. This is where execution happens.
  • Product family — Errors partially cancel out. Useful for capacity planning.
  • Category / Business unit — Smoothest signal. Relevant for financial planning.

Always report accuracy at the level where decisions are made. Measuring only at the aggregate level can hide serious problems at SKU level.

# Time Horizon

  • 1-month lag — How good was last month's forecast? Your most frequent check.
  • 3-month lag — The horizon most relevant to S&OP/IBP decisions (lead time coverage).
  • 6+ month lag — Strategic accuracy. Expect lower precision, focus on bias.

# Common Pitfalls ⚠️

  • Measuring at the wrong level — Aggregate accuracy looks great while individual SKUs are wildly off. Always decompose.
  • Ignoring bias — A 20% MAPE with zero bias is very different from 20% MAPE with 15% positive bias. Track both.
  • Not tracking FVA — Without FVA, you can't tell if your consensus process adds value or just adds noise.
  • Gaming the metric — If planners are incentivized on accuracy alone, they may sandbag forecasts or avoid adjustments. Pair accuracy with bias.
  • Comparing across segments — A 10% MAPE on stable A-items is mediocre; 10% MAPE on erratic C-items is exceptional. Benchmark within segments.

# Further Reading