Numerai Competition Analysis

The Numerai competition involves a structured approach to data handling, feature selection, and model training. Below are the key points and processes involved:

1. Round-by-Round Data Usage:

   • The competition is structured in rounds, where data is provided and analyzed round by round.

2. Feature Set Selection:

   • Participants can choose between different feature sets: small, medium, or all features. This allows flexibility in the complexity and granularity of the analysis.

3. Data Organization by Eras:

   • The dataset is divided into eras, which are subsets of the data representing different time periods. This segmentation is crucial for understanding temporal patterns and variability.

4. Reading Data:

   • The selected features, along with their corresponding ERAS and TARGET columns, are read into the analysis environment.

5. Era-Based Traversal:

   • The analysis process involves traversing over every 4th era to ensure a representative and efficient sampling of the data.

6. Feature Elimination via Correlation:

   • Features are compared against the target using correlation metrics. Features with low correlation to the target are eliminated to streamline the model.

7. Model Training with LGBMRegressor:

   • The primary model used is LGBMRegressor. It is trained once using the entire dataset to ensure comprehensive learning.

8. Handling Missing Values:

   • Any missing values (NaN) in the dataset are filled with 0.5 to maintain consistency in the data.

9. Identifying Important Features:

   • The model.booster_() method is used to identify the most important features of the model. This step is crucial for understanding which features have the most predictive power.

10. Weekly Data Update and Comparison:

    • Every week, new data is added. The newly added features are compared with the previously identified important features. If they match, no retraining is required. Otherwise, the model is retrained to incorporate the new features.

11. Neutralizing Predictions:

    • To neutralize the predictions, a linear model is used. The predictions are adjusted by subtracting the influence of certain features, ensuring that the predictions are not overly biased by any single feature.

Additional Considerations:

12. Role of Eras:

    • Eras are critical in the analysis as they represent different time periods. Clustering can be performed based on the correlation values of different eras to identify similar patterns.

    • Correlation Analysis: The correlation between features and targets is analyzed era by era to understand temporal dependencies.

    • Data Efficiency: As an alternative to float, data can be loaded in int-8 format to save time and resources.

    • Target-Specific Models: Building different models for different targets can enhance prediction accuracy.

    • Era Statistics: There are 574 eras, with each era representing 5 days. Each target has 20 values, equivalent to 4 weeks. Thus, picking every 4th era is beneficial for a representative sampling.

Resource:

For further information and example scripts, visit the Numerai Example Scripts repository.