Demand Forecasting
Demand Forecasting is the discipline of predicting how much customers will buy — by SKU, channel, region, and time period — accurately enough to drive production, procurement, staffing, and inventory decisions. The two metrics that matter are Forecast Accuracy (1 − MAPE, where MAPE = Mean Absolute Percentage Error) and Bias (whether you systematically over- or under-forecast). World-class forecast accuracy at SKU/week granularity is 75-85%; mediocre is 50-65%; and many companies running 'gut feel forecasts' are at 30-40% — barely better than guessing. The methods range from simple (moving averages, exponential smoothing) to statistical (ARIMA, Holt-Winters seasonal) to ML (gradient boosting, neural nets) to causal (regression on price, promotion, weather, macro). KnowMBA POV: the choice of model matters less than (1) cleaning your data, (2) measuring accuracy at the level you make decisions, and (3) closing the loop so forecast error feeds back into safety stock. A 70%-accurate forecast with a known bias and a tight feedback loop beats an 85%-accurate forecast that nobody updates.
The Trap
The trap is forecasting at too aggregated a level. Total-company monthly revenue forecasts are often 95%+ accurate — and operationally useless. You don't make production decisions on 'total company revenue'; you make them on SKU-week-region. At THAT level, the same company is 60-70% accurate. Always measure forecast accuracy at the level of the decision it drives. The other trap is treating the forecast as a target. When sales teams sandbag forecasts to beat them, or when leadership pressures forecasts upward to match plan, the forecast becomes a political document instead of a probability statement — and inventory, capacity, and hiring decisions all degrade.
What to Do
Build a forecast accuracy review cadence: (1) Forecast at SKU-week-region. (2) Track MAPE and bias monthly. (3) Decompose error sources: were you wrong on baseline trend, seasonality, promotions, or new-product introduction? (4) Tie safety stock formulaically to forecast standard deviation — high-error items get more buffer, low-error items get less (don't carry safety stock for predictable items). (5) Run a Sales & Operations Planning (S&OP) cycle monthly that aligns the forecast across sales, finance, and operations — one number, owned, with explicit confidence bands. (6) Hold a quarterly bias review: if you're chronically over or under, fix the model rather than blaming reality.
Formula
In Practice
Zara built its competitive moat on demand forecasting that runs at the speed of fashion. Instead of a traditional 6-12 month seasonal forecast, Zara forecasts at the SKU-store-week level using daily POS data, store-manager qualitative input, and rapid small-batch production. New designs go from concept to store in 2-3 weeks. Each store gets twice-weekly shipments tuned to that store's actual sell-through, not a regional average. The result: ~85% of inventory sells at full price (industry average is 60-70%), markdown waste is 50% lower than competitors, and Zara's inventory turns ~12x annually versus 4-6x for Gap or H&M. Zara's lesson: forecast accuracy isn't a statistics problem — it's a feedback-loop problem. The shorter the loop between sale and replenishment, the less forecasting you actually need.
Pro Tips
- 01
Measure MAPE at the level you make decisions — SKU-week, not category-month. A 95%-accurate company-level forecast hides a 60%-accurate SKU-level forecast, and inventory pain happens at the SKU level.
- 02
Always track BIAS separately from accuracy. A forecast that's off by ±20% randomly is fixable with safety stock; a forecast that's off by 20% consistently in one direction is a model problem and the safety stock will be wrong too. Bias should oscillate around zero — if it doesn't, recalibrate.
- 03
Killing forecasting through speed beats improving forecasting through math. Zara, Shein, and Amazon all reduced forecast horizon (and thus forecast error) by compressing replenishment lead times. Every week you cut from your supply chain is a week of forecast you no longer need to make.
Myth vs Reality
Myth
“Better algorithms = better forecasts”
Reality
After basic statistical methods (Holt-Winters, ARIMA), the marginal accuracy gain from fancier ML is usually 2-5 points of MAPE — useful but not transformative. The biggest accuracy gains come from data hygiene, fixing promotion/event tagging, and forecasting at the right granularity. Companies adopting ML before fixing data engineering get expensive disappointment.
Myth
“100% forecast accuracy is the goal”
Reality
100% accuracy is impossible (demand is partially random) and even chasing it is wasteful — at some point another point of accuracy costs more than the inventory it would save. The goal is accuracy SUFFICIENT to make the operating decision, with safety stock sized to the residual error. Know when to stop chasing accuracy and start managing variance.
Try it
Run the numbers.
Pressure-test the concept against your own knowledge — answer the challenge or try the live scenario.
Knowledge Check
Your monthly company-level revenue forecast is 96% accurate, but stockouts and excess inventory are both rising. What's the most likely problem?
Industry benchmarks
Is your number good?
Calibrate against real-world tiers. Use these ranges as targets — not absolutes.
Forecast Accuracy at SKU-Week
Consumer goods, retail, and DTC at SKU-week-region granularityWorld Class (Zara, Amazon)
> 80%
Best in Class
70-80%
Average
55-70%
Below Average
40-55%
Guessing
< 40%
Source: Institute of Business Forecasting & Planning (IBF) benchmarks
Real-world cases
Companies that lived this.
Verified narratives with the numbers that prove (or break) the concept.
Zara (Inditex)
1990s-present
Zara built fast-fashion's dominant moat by treating demand forecasting as a feedback-loop problem instead of a math problem. Daily POS data from every store flows back to design and production teams in La Coruña. Designs go from sketch to store in 2-3 weeks. Each store gets twice-weekly shipments calibrated to that store's actual sales, not a regional forecast. Where competitors place 6-month seasonal bets and discount the misses, Zara places 2-week bets and reorders the winners. ~85% of inventory sells at full price versus 60-70% industry average.
Concept-to-Store Time
2-3 weeks (vs 6-9 months industry)
Full-Price Sell-Through
~85% (vs 60-70%)
Inventory Turns
~12x (vs 4-6x peers)
Markdown as % of Revenue
~50% lower than competitors
The fastest way to improve forecast accuracy is to need less of it. Zara compresses lead times so dramatically that the forecast horizon is short enough to be reliable. Compete on supply chain speed and you defang the forecasting problem.
Related concepts
Keep connecting.
The concepts that orbit this one — each one sharpens the others.
Beyond the concept
Turn Demand Forecasting into a live operating decision.
Use this concept as the framing layer, then move into a diagnostic if it maps directly to a current bottleneck.
Typical response time: 24h · No retainer required
Turn Demand Forecasting into a live operating decision.
Use Demand Forecasting as the framing layer, then move into diagnostics or advisory if this maps directly to a current business bottleneck.