Contestants were traditionally required to forecast 18 points into the future.
A review of typically refers to the dataset from the NN3 Forecasting Competition (2006–2007), a seminal event in neural networks and computational intelligence for time series forecasting. This file usually contains a collection of 111 monthly time series drawn from empirical business data. Dataset Overview
The historical data is typically provided in vertical columns of varying lengths. nn3.zip
The "masked" nature of the data (anonymized origin) ensures that models must rely on time series patterns rather than domain-specific knowledge. Practical Considerations
111 monthly time series, including the 11 from the reduced set. Dataset Overview The historical data is typically provided
The series vary in length (68 to 144 observations) and include seasonal, non-seasonal, and "difficult" patterns with outliers and structural breaks. Key Strengths
The NN3 competition was designed to evaluate how modern neural network (NN) and computational intelligence (CI) methods compare to traditional statistical benchmarks like those used in the M3 competition. Composition: The series vary in length (68 to 144
Includes a wide range of real-world business patterns (e.g., industry data), making it a robust test for model generalization.
