Welcome to Avenue
Avenue trains gradient boosting machines (GBMs) that can be exactly represented as actuarial factor tables—no approximation or information loss.
The Problem
Generalized Linear Models (GLMs) are transparent and regulatory-compliant but require extensive manual feature engineering:
- Manual binning of continuous variables
- Trial-and-error for interactions and non-linearities
- Limited predictive power
Gradient Boosting Machines (GBMs) automatically discover patterns and achieve superior accuracy, but their complexity makes them unsuitable for regulatory filings.
The Solution
Avenue trains GBMs with specialized L0-like regularization that produces models exactly representable as factor tables. You get GBM-level predictive performance in a format suitable for regulatory filing.
Workflow
- Upload your dataset (Parquet format) from S3 storage
- Configure target variable, objective function (Poisson, Gamma, Tweedie, etc.), and features
- Training runs multi-objective optimization to build a Pareto frontier of models trading off accuracy vs. complexity
- Select a model from the Pareto frontier based on your accuracy/complexity requirements
- Review factor tables, feature importance, and test set performance
- Edit factor values directly in the interface (optional)
- Export factor tables as CSV or Excel
Key Features
Exact Representation: GBMs are exactly represented as factor tables—predictions match to machine precision. No approximation or distillation.
Multi-Objective Optimization: Training generates a Pareto frontier showing the complete trade-off between predictive performance and model complexity.
Standard Format: Factor tables identical to traditional GLM outputs with main effects and interactions. Suitable for regulatory filing.
Direct Editing: Modify factor values in the interface to apply business constraints, smooth patterns, or implement caps and floors.