Centurion · Credit · Model Foundry
A credit-risk model development factory.
Not a generic ML platform with credit recipes bolted on. Purpose-built around bureau data, the regulated lifecycle, and the documentation discipline that lets billion-dollar decisions stand up to model risk management, internal audit, and the regulator.
Delivered with a forward-deployed credit science team — not as a self-serve SaaS — because models this consequential are not built without the people accountable for defending them.
Why purpose-built
A general-purpose ML platform can train a model. It cannot defend one.
Built by practitioners who have led credit risk model development inside tier-1 card issuers, money-center banks, and the fintech lenders that have scaled the fastest — and codified the disciplines they ran there.
The people behind Model Foundry have built scorecards in production at one of the largest US card issuers, run model risk management functions for institutions with hundreds of billions in receivables, and led credit science at fintech lenders through every stage from launch to securitization. We're not learning credit risk from a textbook. We're writing it down.
And the goal is not the best AUC. The goal is the best AUC that is also explainable, calibrated, stable across vintages, defensible under disparate-impact testing, and reproducible two years from now. A model that wins on a holdout but cannot meet the rest of that bar is not a model that ships.
How it's delivered
Not software-as-a-service. A forward-deployed credit science team — with the factory behind them.
Models that decide who gets a loan, how much, at what price, and how to collect when it goes wrong are not the kind of work you self-serve from a SaaS dashboard. The exposure is too high, the regulatory accountability is too real, and the institutional knowledge required to build a defensible model is too specific.
Every Model Foundry engagement ships with a forward-deployed data science team — credit risk modelers, validation specialists, and model risk management practitioners who sit inside your workflow and drive the work end-to-end. They bring the discipline. The factory amplifies it. You get the model — and the people who built it can defend every choice in it.
Managed end-to-end
We stand up the modeling environment to your security and compliance posture, and we operate it. You set the controls. We carry the work.
Credit-risk practitioners
Modelers who have built scorecards, MRM teams who have defended them to OCC, validators who have written SR 11-7 documentation. Not generalist data scientists.
Accountable end-to-end
From the first sample-design conversation to the post-deployment monitoring review. The same people. Continuity is the point.
What we build models on
Bureau-native, hierarchical, time-trended.
The platform reads TransUnion, Experian, and Equifax data in the shapes they actually arrive in — consumer header, tradelines, inquiries, public records, collections — and exposes the hierarchy and trended history directly to the modeler. Thousands of attributes are pre-computed and pre-validated. No flattening, no information loss.
TransUnion CreditVision
Trended attributes, premium add-ons, link-level data — pre-parsed and pre-attributed.
Experian Premier / Trended 3D
Premier Attributes, Trended 3D, Income Insight, Clear Early Risk Score inputs.
Equifax ATB
Account-level Trended Behavior, OneScore inputs, ID:A consumer link.
24-month trended history
Balance, utilization, payment, limit and behavior trajectories — slopes, peaks, recencies, sustained-elevated counts.
Hierarchical tradeline data
Tradeline-level features stay tradeline-level until you choose how to roll them up. Bank-card, retail, auto, mortgage, student, installment — segmented natively.
Alternative and supplemental
Bank-transaction data, cashflow attributes, rent and telco trades, fraud-bureau signals — composable into the same hierarchy.
The lifecycle, codified
Eight stages. The same workflow your validators expect.
Every model that ships from the foundry passes through the same disciplined sequence. Each stage emits the artifacts the next stage — and the validators — need.
- 01
Sample design
Performance window, bad definition, inclusion rules, indeterminates, design weights.
- 02
Reject inference
Parceling, augmentation, fuzzy reclassification. Assumptions documented, impact quantified.
- 03
Feature engineering
Ratios, slopes, recency-weighted aggregates over trended history — versioned with economic rationale.
- 04
Variable selection & binning
WoE binning — monotonic, supervised, manual. IV, Gini, VIF, PSI at variable level.
- 05
Segmentation
Thin file, thick file, derog, NTC. Tree-discovered or expert-defined. Held out per segment.
- 06
Model fit
Logistic regression for the regulated workhorse. Constrained GBM where defensible. Monotonicity enforced.
- 07
Calibration
Platt scaling, isotonic, prior-correction. Calibration by segment, vintage, and score band.
- 08
Validation
Independent of development. Out-of-time, out-of-segment, override impact. Statistical and economic.
- 09
Compliance review
Adverse-action reasons mapped. ECOA disparate-impact testing. LDA search. Explainability artifacts.
- 10
Monitoring
PSI/CSI on inputs. Score PSI. Calibration drift. Vintage backtesting. Override drift.
The models built here
Every model in the credit lifecycle. Built with the same discipline.
One factory. One documentation standard. One monitoring layer. Whether the model decides who to lend to, how much to lend, what to charge, when to call, or whom to market to next.
Lead-scoring & prescreen
Rank prospects before the credit pull. Optimize marketing spend, channel mix, and pre-approval rates.
Application / origination
The accept/reject decision. PD models, custom scorecards, swap-in/swap-out analysis against the incumbent.
Risk-based pricing
APR-tier assignment. Calibrated PD into pricing curves, with profitability-aware tiering.
Line-assignment models
Initial credit-limit allocation. Loss-given-default and exposure-at-default integrated with PD into expected-loss optimization.
Loan-sensitivity / loan-amount
For installment lenders. The model that decides how much to lend given the applicant, the channel, and the elasticity of the segment.
Behavioral / account management
Ongoing risk re-scoring. CLI (credit-line increase), CLD (decrease), authorization strategies, reissue decisions.
Collections models
Propensity-to-pay, optimal-contact-time, settlement-amount, right-party-contact, and channel-treatment models for the full collections funnel.
Recovery & charge-off
Post-charge-off recovery scoring. Sale-vs-retain-vs-litigate decisioning. Net-present-value-aware treatment routing.
Loss-forecasting (CECL / IFRS-9)
Lifetime expected-loss models. PD / LGD / EAD term structures. Vintage-curve forecasting integrated with macro-scenario stress.
Audit-ready by construction
The documentation is the byproduct, not a separate project.
Every model build emits a complete audit trail — training data fingerprints, variable lineage, segmentation rationale, validation results, reason-code mappings, disparate-impact tests, and signed-off review artifacts. Aligned to the standards that examiners read against.
SR 11-7 / OCC 2011-12 model risk management
Development documentation, independent validation artifacts, ongoing monitoring records, and the inventory metadata MRM functions are built to consume.
ECOA and Reg B
Disparate-impact analysis using BISG / BIFSG proxies, less-discriminatory-alternative search, and the documentation trail that defends the model choice.
FCRA §615 adverse action
Reason-code attribution for every score, mapped to consumer-readable adverse-action language, with the score-factor logic auditors can reproduce.
Reproducibility
Every model artifact carries a hash of its inputs, code, and configuration. A model from two years ago can be rebuilt bit-for-bit and re-defended.
What's deliberately absent
No drag-and-drop. No notebook sprawl. No generic AutoML.
Drag-and-drop pipelines, notebooks, and AutoML are fine tools for exploration. They are the wrong tools for building a model on which credit decisions of consequence will be made. Model Foundry constrains the workflow to the disciplined sequence credit risk demands — because the discipline is what produces a model that holds up under scrutiny.
In production
Model Foundry is built for production credit programs where every score must be traceable, every variable documented, and every model defensible.