Not every risky borrower looks risky at first glance. In a Latin American digital lending flow, JuicyScore identified compact applicant segments where NPL90 risk exceeded 70% — using signals hidden in the device, session behavior and connection environment.
The lender had an existing decisioning process, but needed a stronger digital risk layer to see what happens before a borrower enters the portfolio: which device is used, how the application is completed, what connection patterns appear during the session, and whether the technical environment looks consistent or suspicious.
The analysis showed how the full JuicyScore API can support deeper risk modeling, while the lightweight JuicyID vector can help validate simpler stop-marker rules for faster implementation.
Latin American Digital Lending Needs More Than Traditional Credit Data
Digital lending in Latin America continues to grow through faster onboarding, mobile-first journeys and short online application flows. But this growth also creates a familiar risk problem: lenders must process more applications with limited time, uneven borrower data and a constant need to separate real customers from high-risk or manipulated sessions.
In this environment, traditional data alone may not be enough. The way an application is submitted can carry critical risk information: device quality, browser setup, behavioral patterns, connection anomalies, repeated application signals and other technical indicators.
For online lenders, the challenge is not only to reject risky applications. It is to understand the full risk distribution of the flow and use this intelligence in a practical way: to filter the most toxic applications earlier, strengthen existing models and identify lower-risk segments that may support approval growth.
The Challenge: Improve Risk Separation Without Adding Friction
The lender wanted to improve the quality of credit decisions across its web channel without making the application process more complex for good customers.
The objective was to identify additional digital risk signals that could help the company:
- detect applications with significantly higher NPL risk;
- build clearer risk segmentation across the flow;
- strengthen existing decisioning rules and internal models;
- identify low-risk borrower segments that may support approval growth;
- validate simpler stop-marker rules that could be implemented quickly.
JuicyScore approached the task in two layers.
The full JuicyScore API was used to build a custom ranking model based on device, behavioral, connection and technical attributes. This helped show how deeply digital signals can separate risk across the lender’s own flow.
JuicyID, a lighter version of JuicyScore’s digital risk intelligence, was used to test simpler stop-marker logic. This helped identify compact, easier-to-implement rules that can deliver near-term risk control without requiring a full model redesign.
A Two-Layer Approach: Full Risk Model + Lightweight Stop Markers
JuicyScore built a custom ranking model around the lender’s historical application data. The model combined bucketed variables, high-risk markers, medium-risk markers and positive indicators from the JuicyScore API vector.
The analysis showed that the lender’s application flow could be divided into clear risk segments. On the model development sample, risk ranged from approximately 14% in the lowest-risk segment to around 55% in the highest-risk ranking segment. On the test sample, the model remained stable, with risk ranging from approximately 18% to 45% across the ranking segments.
This demonstrated that JuicyScore signals can be used not only as isolated fraud rules, but also as model components for broader credit risk segmentation.
At the same time, JuicyID was tested as a lightweight layer for quicker implementation. It focused on additional stop markers that combine technical anomalies, behavioral markers and connection-related signals. These rules can be easier to validate and deploy first, especially when a lender wants fast risk mitigation before moving to deeper model integration.
This became the key product advantage of the project: the lender could use JuicyScore for advanced custom modeling and JuicyID for practical, faster-to-implement risk rules.
Results of the Preliminary Analysis
Around 15% Gini on development data and around 13% on test data
The custom ranking model built with JuicyScore signals showed strong additional separation potential. The non-normalized Gini reached approximately 15% on the development sample and around 13% on the test sample, supporting the use of the model as a separate underwriting stage or as a component of the lender’s integrated scoring system.
5 risk segments identified across the application flow
The JuicyScore-based model divided the web-channel flow into five ranking segments. The analysis showed a clear difference between lower-risk and higher-risk applicants, helping the lender better understand where additional rules, review logic or model adjustments could be applied.
High-risk stop-marker segment with around 70%+ NPL90 risk
Additional JuicyScore stop markers identified a small but highly risky segment of approximately 1% of applications. This group showed NPL90 risk of around 70% on the development sample and above 70% on the test sample.
This segment can be considered for automatic decline, additional verification or stricter underwriting treatment after validation in the lender’s own decisioning environment.
Simplified stop markers also showed strong near-term potential
A reduced set of simpler stop-marker rules identified approximately 0.5–0.7% of applications with risk close to 70–75%. These rules are especially useful because they can be easier to validate and implement quickly.
For risk teams, this creates a practical path: start with a compact set of high-impact rules, measure the effect, then expand usage into broader model integration.
JuicyID identified additional compact high-risk segments
The JuicyID analysis showed that even a lighter digital risk vector can help isolate meaningful high-risk groups. Additional JuicyID stop markers identified around 4% of applications with risk close to 55–57%, while the most conservative toxic-risk rules isolated less than 1% of the flow with risk around 70–75%.
This gives lenders a flexible way to use digital intelligence depending on their risk appetite: broader high-risk segmentation for monitoring and model enrichment, or stricter stop-marker rules for conservative early filtering.
Positive markers may support safer approval growth
The JuicyScore analysis also identified positive markers and lower-risk segments that may help the lender increase approvals safely. Around 7% of rejected applications were highlighted as a potential area for further validation.
For online lenders, this is an important part of the value: the same digital signal layer can help reduce losses on the high-risk side and reveal growth opportunities among applicants who may be safer than traditional rules suggest.
What the Signals Revealed
The analysis showed that several groups of digital indicators were especially useful for risk separation.
Technical and device anomalies
JuicyScore markers helped identify applications with suspicious or low-quality device environments, including unusual browser or device configurations, canvas-related anomalies, suspicious plugins, device quality indicators and other technical markers.
These signals are important because high-risk applicants often try to manipulate the environment from which the application is submitted.
Behavioral anomalies
User behavior during the application process also carried meaningful risk information. Time on page, cursor movement speed, number of corrections, scrolling patterns, hot keys and other behavioral attributes helped identify sessions that did not look like normal customer behavior.
These indicators can be particularly useful in digital lending because fraudulent or low-quality applications may be submitted too quickly, too mechanically or with unusual interaction patterns.
Connection and infrastructure signals
Connection-related markers also helped improve risk separation. Internet infrastructure quality, connection markers, IP behavior and mobile network-related attributes can help detect sessions with higher uncertainty or signs of manipulation.
For lenders operating across heterogeneous digital environments, this adds an important context layer beyond declared customer data.
Why JuicyScore and JuicyID Work Better Together
The project showed two complementary ways to use JuicyScore technology.
JuicyScore, the full product, is best suited for deeper risk analysis, custom model development and stronger integration into the lender’s decisioning strategy. It provides a rich set of device, behavioral, connection and technical attributes that can be used in scoring models, risk rules, segmentation logic and underwriting flows.
JuicyID, the lighter version, is useful when a lender needs a faster and more compact risk layer. It can help identify high-risk sessions using simplified digital markers and stop-rule combinations, making it easier to test, validate and implement quick risk controls.
- JuicyID fits for stop markers implementation and for quick risk mitigation;
- Use JuicyScore signals to build deeper segmentation and model uplift;
- Validate positive markers for safer approval growth;
- Expand from isolated rules to a stronger digital risk decisioning layer.
This makes the solution practical for both immediate operational use and long-term model improvement.
From Risk Control to Safer Growth
The preliminary analysis showed that JuicyScore signals can help the lender improve risk decisioning in two directions at once.
First, the lender can detect highly risky applications earlier by using technical, behavioral and connection anomalies as stop markers or additional review triggers.
Second, the lender can use positive markers and low-risk segments to explore safer approval growth, especially among applicants who may have been rejected by existing rules but show healthier digital risk patterns.
For digital lenders in Latin America, this balance is critical. Growth cannot come at the cost of portfolio quality, but excessive conservatism can also block good borrowers. A real-time digital risk layer helps lenders move beyond binary approval logic and build more precise, data-driven risk strategies.
A Stronger Digital Risk Layer for Latin American Lending
This anonymous case study shows how JuicyScore can help Latin American digital lenders strengthen credit risk models with device intelligence, behavioral data, connection signals and technical risk markers.
By building a custom model around the lender’s own application flow, JuicyScore identified clear risk segmentation, high-risk stop-marker groups and potential low-risk segments for further validation.
By adding JuicyID as a lightweight layer, the lender also received a practical path for faster implementation of compact digital stop markers.
For online lenders, this is not only about fraud detection. It is about building a more precise, scalable and risk-aware approval strategy — one that can reduce exposure to toxic applications while preserving growth potential among good borrowers.
