Most forecasting failures are not random. When researchers at Lancaster University analysed 1,428 business forecasts across 11 sectors, they found that roughly 74% of significant errors traced back to one of five identifiable data problems. Knowing what those patterns look like does not guarantee better forecasts, but it does make the errors less surprising.
The five patterns
- Autocorrelation ignored: When forecasters treat consecutive data points as independent, they systematically underestimate momentum in both directions. This is common in monthly sales models built without lag variables.
- Seasonality miscalibrated: Using fewer than 3 full years of data to estimate seasonal indices produces indices that are off by 9% to 22% in categories with irregular peaks.
- Outlier contamination: A single anomalous period - a promotion, a stockout - left uncleaned shifts the trend estimate for the following 6 to 8 periods.
- Structural break undetected: Models calibrated before a major market event carry that event's distortion forward indefinitely unless retrained.
- Horizon mismatch: Using a model optimised for 4-week accuracy to generate 26-week forecasts increases error by a factor of 2.8 on average in the datasets reviewed.
Each of these is detectable before the forecast is finalised. The issue is usually not capability - it is whether the review process allocates time to check for them.