Risk control models that reduce bad debt and survive audit
We design, validate and operate the credit, fraud and collections models that decide who gets credit, how much, at what price, and how to recover it when it goes wrong.
Most bad debt is a decision problem, not a data problem
Lenders rarely lack data. They lack a decision path that turns it into consistent, explainable and monitored outcomes. Scorecards go stale, policy overrides accumulate, and fraud adapts faster than the quarterly model refresh. The result shows up twelve months later as a worsening vintage curve.
Tosso rebuilds that path end to end. We start from your loss data and your approval targets, then deliver the features, model, policy rules, monitoring and documentation as one package your risk, compliance and engineering teams can each sign off on.
Four stages, one modeling stack
Device, identity, consortium and alternative-data signals combine into a pre-approval score that rejects synthetic identities and fraud rings before they enter the book.
Probability-of-default models with monotonic constraints, reject-inference calibration and champion/challenger deployment. Limit and pricing policy is expressed as versioned rules, not spreadsheet overrides.
Behavioral scores refresh every cycle. Population-stability and performance dashboards alert on shift, and early-warning triggers feed limit-management actions automatically.
Propensity-to-cure models rank accounts by expected recovery per contact, then match each one to the lowest-cost channel and treatment likely to work.
Deliverables in every engagement
Versioned, point-in-time-correct features so training and production never disagree.
Reason codes on every decision, mapped to adverse-action language your compliance team approves.
Development documentation, back-testing, sensitivity analysis and a monitoring plan aligned to SR 11-7 expectations.
Stability, drift and performance tracking with retrain triggers, so the model you launched stays the model you trust.
Models we build and operate
| Model | Decision it supports | Typical inputs |
|---|---|---|
| Application scorecard (PD) | Approve / decline, initial limit, pricing tier | Bureau, bank-transaction, alternative data, application attributes |
| Fraud & synthetic identity | Block, step-up verification, manual review | Device, identity graph, velocity, consortium signals |
| Behavioral score | Limit increase / decrease, cross-sell eligibility | Payment history, utilization, transaction patterns |
| Early warning | Pre-delinquency outreach, exposure reduction | Cash-flow trends, external stress signals, account behavior |
| Propensity to cure / LGD | Collections queue, channel and treatment selection | Delinquency history, contact outcomes, settlement data |
| Trade-based borrower score | Credit for cross-border sellers and importers | Marketplace settlements, shipment and delivery events, dispute ratios |
From data access to production in weeks
Loss analysis on your historical book, approval/loss trade-off curve, data-readiness review.
Feature engineering, model development, policy design, fairness and stability testing.
Run alongside the incumbent, compare decisions, complete the validation dossier for sign-off.
Go live behind a stable decision API or batch job; monthly monitoring reviews with your risk team.
Built for the people who have to defend the model. We assume every decision will one day be reviewed by an auditor, a regulator or a declined applicant. Fair-lending analysis (ECOA / Regulation B adverse action), model-risk documentation, role-based access and immutable audit logs are part of the standard delivery, not add-ons.
Bring us a portfolio
Tell us where loss is concentrated and what approval rate you need to protect. We come back within two business days with a scoped diagnostic.