Contestants were traditionally required to forecast 18 points into the future.
11 monthly time series used as a small-scale pilot.
It is a standard historical benchmark in the forecasting community and is often included in modern research packages like the tscompdata R package on GitHub. nn3.zip
The historical data is typically provided in vertical columns of varying lengths.
The "masked" nature of the data (anonymized origin) ensures that models must rely on time series patterns rather than domain-specific knowledge. Practical Considerations The historical data is typically provided in vertical
For modern use, researchers often access these files through the NN3 Official Website or data science repositories like Kaggle . Critical Reception
The series vary in length (68 to 144 observations) and include seasonal, non-seasonal, and "difficult" patterns with outliers and structural breaks. Key Strengths Critical Reception The series vary in length (68
Includes a wide range of real-world business patterns (e.g., industry data), making it a robust test for model generalization.