Refinancing: AI and behavioural science help banks defend their home loans portfolios
In home loan refinancing, it is better to attack than to defend.
Banks traditionally focus on issuing new home loans rather than retaining existing borrowers. This is due to information asymmetry, where refinancing banks have more data about the borrower than the current bank, giving them the power of comparison and the upper hand in the refinancing battle. As a result, borrowers hop on and off, boosting brokers’ commissions while cutting banks’ profitability.
By enhancing the comparison bias with AI, current lenders now have a weapon to defend their home loan portfolios.
Comparison bias
Let’s run a simple experiment.
Imagine, you have three bowls of water: one hot, one cold, and one with room temperature water. You put your left hand into the hot water and your right hand into the cold water at the same time, then pull hands both out and put them together into the room temperature water. What happens is remarkable. The left hand will feel cold while the right one will feel warm. Both hands are in the same water, but they feel differently.
As this experiment demonstrates, previous experience matters.
Our mind is not designed to understand absolute values. Giordano Bruno’s On the Infinite Universe and Worlds famously starts with “sense is no source of certainty, but can attain thereto only through comparison”. This is how our brain is wired. It understands things through comparisons.
Sales people know this phenomenon very well. It is why real estate agents show a prospective buyer some inferior house first, before showing the house they actually want to sell. For the same reason, the last time you bought a new car, you probably ended up buying extras — an extended warranty, insurance, roof racks — which initially you had no interest in. Compared to the price you just paid for the car, the extras look nearly free, even if they are hugely overpriced and outright unnecessary.
In home loan refinancing, comparison bias, enhanced by AI, becomes a powerful weapon, too.
Paradox: Data asymmetry
Home loans are a commodity business, with customers caring about only one thing: the cost of capital, meaning the interest rate.
Banks use risk-based pricing (RBP) models to calculate interest rates for each individual borrower, but RBP models need data — most importantly, income, credit score, and assets. And here comes the interesting part. Although this data is available for the lender at the moment of the loan application (or refinancing), it is typically NOT available through the lifetime of the loan, especially when the borrower is doing day-to-day banking elsewhere.
This creates a peculiar paradox as the current lender knows less about its borrower than its competitors do. This advantage is critical because, by having an accurate estimate of interest rate, competitors can trigger comparison bias in the borrower.
Why can’t banks retain borrowers?
Some banks don’t even have a retention department. This may sound strange, as typically any business with a high customer acquisition cost (CAC) and back-loaded customer life-time value (LTV) would benefit tremendously from retention. A famous Bain report estimated that, by increasing retention by as little as 5 per cent, profits can be boosted by as much as 95 per cent.
So, why don’t banks retain their borrowers?
With no income or credit score information, current banks can’t accurately estimate the refinance rate. Low detection accuracy means many false positives churn errors. Translating from the data science language — current bank is not sure which borrowers are a “flight risk”.
Giving refinancing discounts to borrowers who are not considering leaving is a costly mistake as it means less interest income, which means less profit. Running home loan retention with low accuracy churn prediction is a prohibitively expensive venture.
This is one reason why banks simply observe how competitors, like sharks, bite off pieces of their mortgage portfolio and swim away.
Solution: Proxy data
The digitalisation of all areas of life creates data that was not previously available. For example, LinkedIn and social media activities, as well as third-party payment platforms, can provide a good substitute for information on income and credit scores. With some luck, this proxy data can help achieve the necessary accuracy in interest rate estimation. This data can be a game changer in the battle to retain home loan customers.
The process can be simplified into four main steps:
- Obtain proxy data
Obtain data from LinkedIn, social media, third-party payment systems, and any other sources that have predictive power for default risk. The process of gathering this data should be continuous.
2. Determine the refinance rate by training r-RBP model
Use proxy data to train a refinancing risk-based pricing (r-RBP) model that replicates, as closely as possible, the results of the actual RBP model. The outcome of the r-RBP model is a refinancing rate for each individual borrower in each moment in the past, not just at the time of the application.
3. Estimate the comparison gap
The difference between the r-RBP refinancing interest rate and the borrower’s current rate is the comparison gap, which is, again, the main driver behind churn.
4. Train a new churn model
Once accurate estimates of the comparison gap are obtained for a sufficient period of time in the past, a new churn model can be trained on historical data. When tested against the old churn model, the new model should be more accurate at predicting borrowers who are likely to refinance, making it, hopefully, financially viable for banks to retain their home loan portfolios.
De-risking
As always with high-stakes ML systems, it is better to follow the stepping stone de-risking strategy to minimise the costs of inevitable mistakes.
For example, it may be worthwhile to start with loans that are close to maturity. In these cases, if the churn is predicted incorrectly, the potential loss would be limited to only a few years of interest payments.
In addition, continuous experimentation is essential. A/B testing should be run fanatically. This, of course, will require a proper MLOps design and agile product management processes.
Results
Banks have been using machine learning automation for decades, yet it is surprising how many still do not use a basic behavioral bias such as comparison bias to defend their home loan portfolios. This bias can unlock a significant amount of value. For example, the author was involved in a project where a similar system generated additional annual revenue totaling more than $50 million.
Post script
One final note. Changes in systems often create unexpected side effects. For example, they can trigger undesired behaviour at intermediary levels, such as by motivating sales managers or brokers to fail to act in line with the responsible lending code. To identify and prevent such behavioural deviations early, it is important to have a feedback loop and anomaly detection systems in place.
About the Author
60% of this article was written by ChatGPT. The remaining 40% was written by Vladimir Yuzhakov, an ML product leader who specialises in solutions that change human behavior using ML and AI. He has a particular interest in group behavioral science, unconventional ML/AI applications, and general management.
Vladimir has industry expertise in banking and financial services (EY, KPMG, PwC, Toyota Financial Services, The National Bank of Canada, HSBC, Nomura).
Founded award-winning ML-driven behaviour management start-up (PosiSense).
Forbes columnist who has authored 25 articles covering economics and behavioural science.
Academic background: Columbia University (MBA), Humboldt University zu Berlin (MA in Economics).
