Predicting payments is a full-time job for every equipment finance company. Sometimes collecting payments becomes a bigger job when payment predictions do not come to fruition. Credit scores, whether purchased from various credit bureaus or defined by the business with a custom scorecard, are the primary tools that underwriters use to predict whether a payment will arrive on time.
Now there’s a new kid on the block for underwriters and business managers – AI based predictors.
While not a replacement for credit scores, predictors are related to credit scores by the inputs they both use and when credit scores are used in the machine learning models that many predictors often use. But where predictors differ from credit scores is in the scope of information they provide and the decisions they support. Scoring is a foundational aspect of everyday life. Scoring is comfortable and something with which most everyone is accustomed. From an early age, we played games and took tests in school. And as adults, scoring and especially credit scores play an important role in our personal and financial lives.
However, using predictors to help with daily decision making is something with which most operations leaders are not comfortable. This should not be the case.
Credit scores predict credit worthiness with mathematical models that use data including payment history, amounts owed, time in business, frequency of credit needs, and the mix of types of credit used. But using this method is a way to rank risk rather than measuring it. Investopedia states:
Although credit scoring ranks a borrower’s credit riskiness, it does not provide an estimate of a borrower’s default probability. It merely assesses a borrower’s riskiness from highest to lowest. As such, credit scoring suffers from its inability to determine whether Borrower A is twice as risky as Borrower B.
The key to understanding risk is understanding the probabilities of outcomes, and specifically undesirable outcomes. This is where predictors, also known as prediction machines, come to play. Where credit scores and scoreboards use a few tens of inputs, machine learning models and the predictions they facilitate can use literally hundreds of variables because of the scale of the cloud computing environments in which they run. Most finance companies do not gather hundreds of different data points during origination, but 25 to 50 different inputs are common – including credit scores for the business and all the guarantors.
The greater scale of input variables makes it possible for the machine learning models to predict multi-state outcomes. They predict the probability distribution of a set of outcomes thereby informing on the risk of not reaching the desired outcome. Consider the risk shown for the three “good” deals by a Delinquency predictor capable of predicting delinquency in ten-day states.
Deal 1 clearly has almost no risk of delinquency and as such would be priced at the lowest interest rate. Deal 2 has a prediction of 30-day delinquency – the highest probability in the distribution – but has no probability beyond 60-days. This deal looks like a customer who will be delinquent from time to time but will still pay. This deal could be priced at a higher rate and is likely to generate additional revenue from late fees. This deal would likely have a higher return than Deal 1.
Deal 3 has risk all the way out to 90-days but still no loss probability and a total of 75% probability of delinquency 60-days or less. This deal will need a higher rate to cover this risk and likely will have a higher expense due to collections efforts along the way. Even at a higher rate, Deal 3 could have a lower return than Deal 1 if these fees are not considered.
An interesting detail about the predictor curves above is that all three deals could have the same Credit Score. We have seen such cases and are able to understand them via the past outcomes of the business and the greater diversity of inputs that are put into the models. Predictors can deliver information with more precision because they can use variables like equipment class, for example. Indeed, a company with enough data could train and deploy predictors for different classes of equipment, different NAICs codes, and for different periods of time, e.g., pre and post pandemic.
Credit scores have served the industry well for almost four decades and will be a part of decision making. But scores alone have limitations when it comes to understanding and embracing risks that could lead to potential upside. Predictors extend credit scores by adding data and focusing on outcomes in ways that provide much more information with more precision. Credit Scores and predictors are related, but they represent different generations of risk management. Predictors address, quite literally, the future and provide a new way of thinking about credit decisions and risk. Giving leaders the opportunity to make better, faster decisions with greater precision than ever before.