Getting Started
Avenue trains gradient boosting machines that are exactly representable as factor tables. This guide walks through the complete workflow from data upload to exporting factor tables.
Data Requirements
File Format: Parquet only
Target Variable: Must match your selected objective function
- Poisson: non-negative integer counts (e.g., claim frequency)
- Gamma: positive continuous values (e.g., severity)
- Tweedie: zero-inflated positive values (e.g., pure premium)
- Binary: binary outcomes (0/1)
- Regression: continuous values
- Huber: continuous values with outliers
Data Quality:
- No missing values in target column
- Consistent data types within columns
- Remove invalid records (e.g., negative exposures) before upload
Workflow
-
Upload Data
- Upload Parquet file from S3 or select previously uploaded dataset
- Only you can see your datasets
- Upload guide
-
Configure Model
- Select target variable
- Choose objective function (must match target distribution)
- Select feature columns
- Configuration guide
-
Training (automatic)
- Multi-objective optimization builds a Pareto frontier
- System explores hundreds of hyperparameter configurations
- Varies regularization penalties, tree depth, number of leaves
- Uses cross-validation to evaluate predictive performance
- May take several minutes depending on dataset size
-
Model Selection
- Review Pareto frontier showing accuracy vs. complexity trade-off
- Left side: simpler models (fewer tables) for easier interpretation
- Right side: more complex models for maximum accuracy
- Selection guide
-
Review Results
- Check feature importance rankings
- View factor tables
- Review test set performance if test data provided
- Results guide
-
Edit Tables (optional)
- Modify factor values directly
- Apply business constraints or smoothing
- Editing guide
-
Export
- Download as CSV or Excel
- Choose tables separated by main-effect and interaction (ANOVA-style) or consolidated
- Export guide