“Operating costs were now decoupled from portfolio size growth.”
The client
A major North American consumer finance company with a large store network, serving sub-prime and near-prime customers.
The lender held a large amount of predictive data. But its risk management, underwriting and customer management had been built around the stores, so most of the work was manual. Its operational controls were strong, with mature identity checks, and customer satisfaction was high thanks to a strong relationship culture.
Product range
- Sub-prime: APR 42%–116%, terms up to 5 years, loans of $500–$5,000
- Near-prime: APR 26%–38%, terms of 3–5 years, loans of $5,000–$15,000
- Products: short-term loans, instalment loans, lines of credit, car finance, savings accounts, overdrafts and insurance
Regulation was changing and the market was under pressure. The lender wanted to:
- automate underwriting across all stores and online (a robo-underwriter);
- reduce losses and improve control;
- improve the performance of the existing book;
- lower operating costs, and stop them rising with the size of the portfolio;
- grow online and mobile lending profitably, in more regions;
- improve its product mix, targeted marketing and customer acquisition.
It wanted to automate quickly, and at the same time move its products from the stores to digital channels.
How the work was done
Zoral started with a short proof of value, then put the results into production in phases, with training and monitoring.
Zoral analysed the store and online portfolios. Standard models from the Zoral model library set a baseline for the lender’s results and estimated the uplift a robo-underwriter could deliver.
Offline loan performance
Default models from the Zoral model library, using only the lender’s application data for new and returning store customers, were applied to the portfolio. They showed which segments were underperforming, and where the underwriting process was inefficient.
Model Library default model for new offline clients
Model Library default model for returning offline clients
Note: each segment represents 5% of the portfolio.
At first the store portfolio was much larger than the online one. An online default model was then applied to the online portfolio, and it also exposed underperforming segments.
Model Library default model for online clients
Next, credit bureau and bank statement data were added to the application data. This let the models work more fully, and their predictive power rose significantly.
Model Library default models with additional data, offline clients
Model Library default models with additional data, online clients
The models also found more fraud in both portfolios. Fraud models from the Zoral model library were put into use, and fraud fell. This cut fraud losses, and gave the lender a reliable digital platform for high-volume growth.
Model Library fraud scorecard for online clients
Model Library fraud scorecard for offline clients
The fraud models also uncovered fraud by staff at some stores. To address it, automated underwriting moved loan approvals from store staff to the lender’s credit policy, applied by the decision engine.
Selected store locations with the highest fraud rate
Affordability checks and models were added through Zoral Decision Engine, online and in the stores. They improved underwriting decisions, reduced fraud and helped the lender meet its regulatory requirements.
Affordability Estimation
Consumer’s net income
+ $6,321
- Consumer’s debt obligations− $273
- Consumer’s housing costs− $1,531
- Consumer’s basic living expenses− $732
Amount of spare cash to repay new loan = $6,321 − $273 − $1,531 − $732 = $3,785
Limit models were also put into use for every product, online and in stores. They calculated the largest affordable monthly amount for each loan.
The robo-underwriter
Default and fraud models fed the robo-underwriter. It ranked and grouped customers by expected profit. The lender also wanted to keep acceptance as high as possible. So the robo-underwriter combined default, fraud, affordability and limit models for each product line. The uplift was significant.
Revenue and average write-off after robo-underwriter optimisation
| Income bracket | <20k | 20k–28k | 28k–36k | 36k–42k | 42k+ |
|---|---|---|---|---|---|
| Current revenue | 180 | 258 | 339 | 415 | 653 |
| Robo-underwriter revenue | 242 | 333 | 406 | 494 | 753 |
| Current average loss amount | 60 | 85 | 120 | 122 | 207 |
| Average loss amount with robo-underwriter | 40 | 62 | 84 | 89 | 158 |
Into production, in three phases
Zoral fOS components
ZDE
Zoral Decision Engine
Model execution and real-time decisioning
ZPS
Zoral Portal Studio
Digital product personalization: mass personalization of financial, digital products
ZBM
Zoral Behavior Matrix
Customer behavior and click-stream data: real-time behavioral data and features
Model Library
Zoral Intelligence Layer
Tuned decision models
ZDF
Zoral Data Fabric
Feature engineering, creation of production-strength analytical models
Your loan management system
Connected to Zoral fOS
Adapts to multiple financial products, intelligent workflows and back-end business processes
Dashboards
Phase I: improve the stores
- Zoral fOS was first connected to the lender’s existing loan management system, to avoid disrupting the business.
- More third-party data sources were added, so the full models could be used.
- The robo-underwriter went live across the lender’s regions and stores.
Phase II: grow online
- The online journey and workflows were built in Zoral Portal Studio to fit the lender’s brand, checks, data capture and processes.
- Lead evaluation was added, to grow online customer acquisition at a sensible cost.
- The online robo-underwriter went live across the lender’s regions.
- Zoral fOS was kept in step with the loan management system, so stores and online worked together.
- Zoral Behavior Matrix was added to the online journey.
Phase III: tune and improve
- Conversion models were tuned and used in digital marketing workflows, across all channels.
- Retention models were used to target and keep profitable existing customers.
- Behavioural signals helped reduce abandoned online applications.
- Variations of the customer journey were tested and tuned, which further improved acceptance rates, profit and customer satisfaction.
During Phase III, online revenue for some products began to exceed store revenue. The lender could then move faster on other parts of the business, such as new products. Operating costs no longer rose with the size of the portfolio, so the lender grew its market share while containing costs and growing profit. It continues to add models and automation.
Results
- Profit up 28%
- Portfolio write-offs down 19%
- Quarterly volume growth up 12%
- Online straight-through processing up from 50% to 87% (the 74% increase in the results box)
- Customer acquisition costs down 32%
- Customer retention rate up 9%
- Online conversion rate up 31%
New clients - dynamic of default rate
Returning clients - dynamic of default rate
Customer application journey - conversion rate dynamics
At launch, with behaviour capture
Conversion rate +15%Using behavioural signals and conversion models
Conversion rate +14%With email marketing automation, remarketing, live chat and a chatbot
+44 20 3963 2002