Lenders, banks and insurers are embracing machine learning to automate processes, improve underwriting decisions and optimize marketing efforts.

Machine learning is transforming financial services in a variety of ways. The financial services industry is a data intensive business and banks, lenders and insurance companies have a wide range of customer data (such as credit reports, application information and account histories) that they can leverage to make better business decisions. In today’s market, strong data management and analytics capabilities can be critical drivers of sustainable competitive advantage, allowing financial institutions to take advantage of next-generation tools such as machine learning and artificial intelligence. This post explores some of the most powerful use cases of machine learning across lending, banking and insurance.

Consumer & Small Business Lenders

Consumer and small business lenders are adopting machine learning models for a variety of powerful use cases, including:

1. Default Rate Forecasts

Predictive models that estimate the probability of default for a loan, either for new originations or existing portfolios. The result is typically either a probability of default (percentage) or a numeric risk score which can be used to approve or deny applications, set risk-based pricing or manage existing accounts.

Data Used to Train Model:

  • Predictive variables such as Loan Amount, Credit Inquiries, Open Accounts, DTI, FICO, Public Records, Home Ownership, Credit History, etc. (Note: for thin-file underwriting, utilize alternative data such as rental history and utility payments)
  • The historical result: did the loan default?

What the Model Predicts:  

  • The percentage chance that a loan will default

2. Prepayment Estimates

Predictive models that determine the probability a loan will be prepaid in full, typically within a given timeline (e.g. in the first XX months after issuance). Prepayment models can be built for new originations or existing portfolio monitoring and the result is typically a percentage probability of prepayment.

Data Used to Train Model:

  • Similar predictive variables to default rate models (see above)
  • The historical result: did the loan prepay?

What the Model Predicts:  

  • The percentage chance that a loan will be prepaid in full

3. Income Predictions

Estimates the current income of an applicant or existing borrower. The result is a numeric income value. The predictive income is often compared to the applicant’s stated income to help verify the accuracy of an application and prevent income fraud.

Data Used to Train Model:

  • Similar predictive variables to default rate models (see above), plus factors such as Residence Location, Age and Job Type. The most powerful predictive variables are often related to total outstanding debt and maximum credit limits
  • The historical result: what was the verified income?

What the Model Predicts:  

  • The individual’s monthly or annual income

4. NSF / No-Pay (Missed Payments in Servicing)

Models that estimates the probability that an existing borrower will miss an upcoming payment. The result is typically the percentage chance a payment will be missed, which is compared to rule-based thresholds that set required actions. This can help internal teams prioritize their workflow and outbound contact attempts.

Data Used to Train Model:

  • Similar predictive variables to default rate models (see above), plus factors related to payment history such as Number of Payments Made, Remaining Loan Amount, State, Loan Type, NSFs, Max Days Past Due, etc. The most powerful predictive variables are typically related to recent payment history.
  • The historical result: did an NSF/no-pay occur?

What the Model Predicts:  

  • The percentage chance that an upcoming payment will be missed

5. Charge-Off Recovery Amounts

Models that estimate the dollar amount that will be recovered in a collection effort, to help determine the next-best action for charged-off loans (e.g. collect, sell or litigate). The result is typically a dollar amount (recovery estimate), which would then be compared with the sale price or cost of litigation to decide on the best course of action for the account.

Data Used to Train Model:

  • Similar predictive variables to default rate models (see above), plus factors related to recent payment history such as Number of Payments Made, Remaining Loan Amount, State, Loan Type, NSFs, Max Days Past Due, etc.
  • The historical result: the recovery amount

What the Model Predicts:  

  • The expected recovery amount from a collection effort

Banks & Credit Unions

In addition to the lending use cases above, banks and credit unions are leveraging machine learning for a variety of banking-specific purposes, including:

1. Transaction Fraud Probabilities

Predictive models that estimate the probability a financial transaction is fraudulent. Multiple machine learning models are often leveraged for a single transaction, each of which predicts a different type of financial crime (e.g. identity theft, friendly fraud, money laundering, income fraud, etc.). The result is typically a percentage probability of fraud, which is used to approve, reject or flag a transaction for manual review.

Data Used to Train Model:

  • Predictive variables such as Recent Deposits, Recent Withdrawals, Account Balance, Transaction Frequency, Transaction Location, Receiving Bank, Transaction Amount, Third-Party Risk Scores, KBA Results, etc.
  • The historical result: was the transaction fraudulent?

What the Model Predicts:  

  • The percentage chance that a transaction is fraudulent

2. Product Recommendations

Models that recommend the best product(s) to offer a customer, based on the probability that a customer will take various offers. The result is typically a set of probabilities (the probability that product X, Y and Z will be accepted), which can be distilled into a specific recommendation for each customer.

Data Used to Train Model:

  • Predictive variables such as Income, Credit Score, Existing Products, Location, Client Tenure, and compiled bank-account transactions that indicate things such as Common Purchases, Travel Preferences, etc.
  • The historical result: when this product was recommended, was it taken?

What the Model Predicts:  

  • The percentage chance that each product will be accepted

3. Probability of Loans Closing

Models that predict the probability that an application will result in a loan (or bank account, insurance policy, etc.). This can be used to prioritize internal workflows; for example, a bank may choose to have loan officers focus their time on the highest-likelihood applications rather than work from a basic queue, increasing efficiency.

Data Used to Train Model:

  • Predictive variables such as Loan Amount, Credit Inquiries, Open Accounts, DTI, FICO, Public Records, Home Ownership, Credit History, etc.
  • The historical result: did the application result in a loan?

What the Model Predicts:  

  • The percentage chance that an application will result in a loan

4. Marketing Response Rates

Predicts the probability that a customer will respond to a paid marketing campaign, which is used to determine whether to include them in a given marketing campaign. Such predictions can also be used to assess the expected return on marketing spend.

Data Used to Train Model:

  • Predictive variables such as Income, Credit Score, Existing Products, Location, Years As Client, and compiled bank-account receipts that indicate things such as Common Purchases, Travel Frequency, etc.
  • The historical result: did the customer respond to the marketing campaign?

What the Model Predicts:  

  • The percentage chance that an application will respond to marketing (e.g. to a direct mail advertisement)

Insurance Companies

1. Loss Rate Forecasts

Predictive models that estimates the loss rate (or loss amount) on an insured, either for new issuances or existing policies. The result is typically either a percentage (e.g. 6.8% probability of a claim) or a numeric value (e.g. estimated total claims of $230).

Data Used to Train Model:

  • Predictive variables such as Age, State, and product-specific information (e.g. car type, property type, health details, etc.)
  • The historical result: did the customer experience a claim? (Or, what were the customer’s total claims?)

What the Model Predicts:  

  • The percentage chance that an insured will experience a claim, or
  • The total expected claims for an issued policy

2. Automated Claims Processing

Models that predict whether a claim would be approved or rejected by manual review, to help streamline internal operations. The result is typically a probability of approval, and thresholds are set to determine next steps (e.g. <1% chance automatically denied, 1% to 95% manually reviewed, >95% automatically approved). These models can help improve the efficiency of claims adjusters.

Data Used to Train Model:

  • Predictive variables such as Claim Amount, Claim Type, State/Location and product-specific information
  • The historical result: was the claim approved or denied?

What the Model Predicts:  

  • The percentage chance that a claim will be approved

3. Loss Estimation on Claims

Predictive models that estimate the total amount that will be paid on a claim. The result is a dollar amount that can be used for a variety of purposes, including process automation, highlighting high-dollar expenses and confirming manual work.

Data Used to Train Model:

  • Predictive variables such as Claim Amount, Claim Type, State/Location and product-specific information
  • The historical result: was the claim approved or denied?

What the Model Predicts:  

  • The expected expense that a claim will result in

4. Loss Event Forecasts

Models that predicts which existing insureds are likely to experience a high-cost claim in the next X months. The result is a probability that can be used for purposes such as ongoing case management and proactive risk mitigation.

Data Used to Train Model:

  • Predictive variables such as Claim Amount, Claim Type, State/Location, and other product-specific information
  • The historical result: did a loss event occur?

What the Model Predicts:  

  • The probability of a loss event occurring in the next X months

Conclusion

At a time when financial institutions are under pressure from a variety of sources, including changing consumer preferences, emerging digital competitors and evolving regulations, technologies that help financial services companies respond appropriately are in high demand. Given the large amount of data at the disposal of players in the space, it’s not surprising that many are embracing machine learning to help them drive better business results across a wide variety of use cases.


About DigiFi

DigiFi is a technology company that helps businesses make better automated decisions.

Our platform lets businesses easily use automated machine learning and rules management to optimize critical decisions with no coding or technical expertise required. Repetitive work that used to take hours can now be completed in minutes, letting your team focus on what matters most.