Counterfactual Explanations

The core mechanism identifies the closest possible world where the outcome differs. Unlike feature importance scores, counterfactuals define a precise vector of change. For a denied loan, it pinpoints exactly how much additional income or debt reduction is required for approval, optimizing for sparsity and actionability.

Counterfactuals bridge the gap between explanation and intervention. They answer 'What can I do to change this outcome?' by suggesting realistic, mutable features.

• Fraud Example: 'If the transaction amount was below $5,000 and the shipping address matched the billing address, the block would be lifted.'
• Contrast with SHAP: SHAP tells you why it was blocked; counterfactuals tell you how to unblock it.

A naive counterfactual might suggest impossible states (e.g., 'be 5 years younger'). Advanced generation uses causal constraints and density-weighted metrics to ensure plausibility.

• Manifold Guidance: Ensures the counterfactual lies within the distribution of real data.
• Causal Reasoning: Prevents changes that violate physical or logical laws, ensuring the explanation is credible to a human auditor.

Generating a counterfactual is a balancing act between competing goals formalized as a loss function:

• Validity: The new prediction must match the desired target class.
• Proximity: The distance between the original input and the counterfactual must be minimized.
• Sparsity: The number of features changed should be as low as possible.
• Diversity: Providing multiple distinct counterfactuals (e.g., 'reduce amount' vs. 'change merchant category') gives the user options.

Humans naturally think in contrasts ('Why P instead of Q?'). Counterfactuals formalize this cognitive process. They don't require opening the black box; they probe the decision boundary directly. This makes them inherently model-agnostic—the same generation logic works for gradient-boosted trees, neural networks, or logistic regression, as long as the model can be queried.

Regulators like the OCC and frameworks like SR 11-7 require adverse action reasons. Counterfactuals naturally translate into Adverse Action Reason Codes by identifying the top features whose modification flips the decision. They provide a mathematically rigorous foundation for generating the 'principal reasons' for denial in credit and fraud models, moving beyond simple sensitivity analysis.