Difference

COMPAS vs XGBoost Recidivism Models

A technical comparison of the legacy proprietary COMPAS algorithm against modern gradient-boosted XGBoost models for recidivism prediction, focusing on accuracy gains, feature importance transparency, and the application of fairness constraints.

Get in touch Learn more

Data engineer managing feature store on laptop, feature definitions visible, casual data engineering session.

THE ANALYSIS

Introduction

A data-driven comparison of legacy actuarial algorithms and modern gradient-boosted models for recidivism prediction, focusing on the critical trade-offs between interpretability, accuracy, and fairness.

COMPAS excels at providing a standardized, static risk score that has been deeply embedded in judicial workflows for over two decades. Its strength lies in its extensive legal precedent and the consistency of its 137-item proprietary questionnaire. For example, in the landmark ProPublica analysis, COMPAS demonstrated a 61% accuracy rate for general recidivism, a benchmark that, while contested, has defined the floor for algorithmic performance in pretrial and sentencing contexts. Its deterministic output offers a single, unchanging score that courts have come to rely on for high-stakes decisions.

XGBoost takes a fundamentally different approach by leveraging gradient-boosted decision trees to model complex, non-linear interactions between features. This results in a significant accuracy gain, with modern implementations often achieving an Area Under the Curve (AUC) of 0.75 or higher compared to COMPAS's typical 0.66-0.70. However, this performance comes with a trade-off: the model's complexity makes it harder to explain a specific prediction to a judge or a defendant, moving from a simple score to a complex feature importance map.

The key trade-off: If your priority is legal defensibility, workflow integration, and a decade of case law, choose COMPAS. If you prioritize predictive accuracy and the ability to apply state-of-the-art fairness constraints like adversarial debiasing, choose XGBoost. The decision hinges on whether your jurisdiction values a static, auditable score over a dynamic, high-performance model that requires a new layer of technical oversight.

HEAD-TO-HEAD COMPARISON

Core Architectural and Governance Feature Matrix

Direct comparison of key metrics and features for recidivism prediction models.

Metric	COMPAS	XGBoost Recidivism Models
Algorithm Transparency
Native Fairness Constraints
Feature Importance Explainability
AUC-ROC (General Recidivism)	0.65 - 0.70	0.73 - 0.78
Racial Bias (False Positive Rate Ratio)	1.45x (Black vs White)	0.95x (with Adversarial Debiasing)
Model Update Cycle	Static (Vendor Controlled)	Continuous (User Retrainable)
Constitutional Compliance Posture	Reactive Litigation	Proactive Auditing

COMPAS vs. XGBoost Recidivism Models

TL;DR Summary

A direct comparison of a legacy proprietary algorithm against modern gradient-boosted machine learning for predicting recidivism. The core trade-off is between legal precedent and auditability versus predictive accuracy and the ability to natively apply fairness constraints.

Choose COMPAS for Legal Defensibility

Established precedent: COMPAS has survived Loomis v. Wisconsin due process challenges, making it the default for jurisdictions prioritizing constitutional compliance over raw accuracy. This matters for court administrators who need to defend risk scores in adversarial sentencing hearings where algorithmic transparency is legally contested.

Choose COMPAS for Turnkey Deployment

Standardized workflow: COMPAS provides a fixed 137-item questionnaire with a validated, unchanging scoring engine. This matters for probation departments lacking in-house data science teams who require a vendor-supported, off-the-shelf tool that integrates directly into existing case management systems without custom model training.

Choose XGBoost for Predictive Accuracy

Superior AUC: Gradient-boosted models consistently outperform COMPAS by 5-15% on AUC metrics for predicting violent recidivism when trained on modern, jurisdiction-specific data. This matters for policy analysts optimizing resource allocation, where a reduction in false positives directly translates to fewer unnecessary pretrial detentions.

Choose XGBoost for Native Fairness Constraints

Adversarial debiasing: Unlike COMPAS's fixed logic, XGBoost architectures allow for the integration of fairness constraints during training (e.g., equalized odds post-processing). This matters for civil rights oversight bodies seeking to actively mitigate racial disparities rather than simply measuring them, addressing the core criticism of the ProPublica COMPAS analysis.

Choose XGBoost for Feature Transparency

SHAP explainability: Modern XGBoost pipelines provide instance-level explanations via SHAP values, showing exactly how each factor contributed to a specific score. This matters for judges who need to understand why a defendant is classified as high-risk, moving beyond the 'black box' critique to provide meaningful due process.

Choose XGBoost for Dynamic Updating

Concept drift mitigation: XGBoost models can be retrained on recent local data to adapt to changing crime patterns and policy shifts. This matters for state-level research units managing recidivism tools across diverse counties, ensuring the model reflects current realities rather than the static 1990s-era norms embedded in COMPAS.

HEAD-TO-HEAD COMPARISON

Predictive Accuracy and Fairness Metrics

Direct comparison of key metrics and features.

Metric	COMPAS	XGBoost Recidivism Models
AUC-ROC (General Recidivism)	0.66 - 0.71	0.73 - 0.82
Racial Bias (False Positive Rate Ratio)	1.45x (Black vs White)	1.05x (with Adversarial Debiasing)
Feature Transparency
Dynamic Risk Factor Integration
Adversarial Debiasing Support
Proprietary Licensing
Constitutional Compliance Risk	High (Black Box)	Low (Explainable)

Contender A Pros

COMPAS: Pros and Cons

Key strengths and trade-offs at a glance.

Established Legal Precedent

Extensive case law acceptance: COMPAS has been cited and scrutinized in landmark cases like State v. Loomis, creating a well-defined legal framework for its use. This matters for court administrators who need defensible tools that have survived judicial review, reducing the risk of appeals based on novel algorithmic challenges.

Turnkey Deployment & Validation

Decades of field testing: With over 1 million assessments completed, COMPAS offers a standardized, off-the-shelf solution with established norms across diverse populations. This matters for probation departments lacking in-house data science teams, as it provides immediate, validated risk scales without the need for custom model training or local data infrastructure.

Composite Risk Profile

Multi-dimensional scoring: Unlike models predicting a single outcome, COMPAS generates separate scales for general recidivism, violent recidivism, and pretrial misconduct. This matters for judicial decision-making, as it allows for nuanced risk differentiation—distinguishing a high-risk property offender from a high-risk violent offender—which is critical for sentencing and supervision intensity.

Enabling Efficiency, Speed & Accuracy

Intelligent Analysis, Decision & Execution

We build AI systems for teams that need search across company data, workflow automation across tools, or AI features inside products and internal software.

Talk to Us

Search across company data

Give teams answers from docs, tickets, runbooks, and product data with sources and permissions.

Useful when people spend too long searching or get different answers from different systems.

Enterprise searchRAGPermissions

Automate internal workflows

Use AI to route work, draft outputs, trigger actions, and keep approvals and logs in place.

Useful when repetitive work moves across multiple tools and teams.

AI agentsWorkflow automationGovernance

Add AI to products and internal tools

Build assistants, guided actions, or decision support into the software your team or customers already use.

Useful when AI needs to be part of the product, not a separate tool.

AI integrationDecision supportModel routing

CHOOSE YOUR PRIORITY

Decision Guide by Stakeholder

COMPAS for Policy Advisors

Verdict: High risk for constitutional liability. COMPAS operates as a proprietary 'black box,' making it nearly impossible to satisfy the due process requirements established in State v. Loomis. The 137-item questionnaire includes factors that may act as proxies for race and socioeconomic status, exposing jurisdictions to disparate impact lawsuits under the 14th Amendment. Strengths: Long historical precedent and extensive case law for legal defensibility in some circuits.

XGBoost for Policy Advisors

Verdict: Superior for regulatory compliance and algorithmic transparency. Modern gradient-boosted models allow for adversarial debiasing and fairness constraints to be applied directly during training. Feature importance scores (SHAP values) provide a clear audit trail for every decision, enabling compliance with NIST AI RMF and emerging EU AI Act requirements for high-risk government systems. Strengths: Enables a 'glass box' approach to automated decision-making, allowing for meaningful human review and appeals.

THE ANALYSIS

Final Verdict

A data-driven comparison of legacy actuarial models versus modern gradient-boosted architectures for recidivism prediction, focusing on the accuracy-fairness trade-off.

COMPAS excels at legal defensibility and established precedent because it is a static, proprietary actuarial tool with decades of case law behind it. For example, its 137-item questionnaire produces a consistent decile-based risk score that courts and parole boards have standardized their workflows around, providing a known quantity in high-stakes sentencing hearings.

XGBoost models take a different approach by leveraging gradient-boosted decision trees to capture non-linear interactions between features. This results in a significant accuracy gain—often a 5-10% improvement in Area Under the Curve (AUC)—but introduces a 'black-box' complexity that is harder to explain to a judge than a simple decile score, creating a trade-off between predictive power and procedural transparency.

The key trade-off: If your priority is legal defensibility, immediate court acceptance, and a static risk score that requires no retraining, choose COMPAS. If you prioritize higher predictive accuracy and the ability to apply fairness constraints like adversarial debiasing to actively mitigate racial disparities, choose XGBoost. For a modern justice system aiming to reduce both recidivism and bias, a transparent XGBoost model with post-hoc explainability (SHAP values) and rigorous fairness audits represents the superior technical path forward, despite the higher implementation burden.

About the author

Prasad Kumkar

CEO & MD, Inference Systems

Prasad Kumkar is the CEO & MD of Inference Systems and writes about AI systems architecture, LLM infrastructure, model serving, evaluation, and production deployment. Over 5+ years, he has worked across computer vision models, L5 autonomous vehicle systems, and LLM research, with a focus on taking complex AI ideas into real-world engineering systems.

His work and writing cover AI systems, large language models, AI agents, multimodal systems, autonomous systems, inference optimization, RAG, evaluation, and production AI engineering.

Limited slotsGet a Free AI Consultation

How We Work

Custom AI workflows for your Business

One-fit-all AI don't work for modern businesses. At Inferensys, we aim to understand your business & custom requirements; which we use to define most efficient agentic workflows, the data, and the tools for your business.