A demand forecast can fail even when the model itself has not changed. Missing or duplicated transactions, incomplete stock records, and gaps in purchased data can distort its inputs. Data quality work starts before aggregation and continues after deployment: faulty training or prediction data can produce faulty predictions. section 6.7.1 section 14.3
Check quality at each point where data becomes a model input:
- Before aggregation, filter or preprocess missing and duplicated transaction, stock, and promotion records. Check externally cleaned data too; prior quality control does not remove the need for local checks. section 6.7.1
- During pipeline operation, watch for access failures, format changes, source corruption, and problems with external APIs. These may originate in processing rather than in the model. section 14.6
- In production, monitor missing data across the dataset and in important features separately. Compare it with a historical baseline that excludes known broken-pipeline periods. section 14.5.1
- When performance falls, check data quality before treating the change as genuine data or concept drift and retraining. section 14.4.4
Retail forecasting checks
Competitor prices cover only about 25% of products and contain gaps, while transaction, stock, and promotion records may be missing or duplicated. One pipeline aggregates transactions daily and can optionally rewrite the preceding two or three days to repair corrupted or incomplete records. For ongoing monitoring, missing-data percentages can trigger alarms at three historical z-scores for important features and four for others; Evidently AI test presets can supplement the checks. section 6.7.1 section 14.5.1
A quality alarm calls for investigation, not automatic retraining. Data-entry errors, schema changes, or an upstream model can explain a bad input. If reliable predictions cannot be maintained, stop and use a fallback until quality is restored. section 14.4.4 section 14.3
