Explainable AI: Building Trust In AI For Financial Services
Artificial Intelligence and Machine Learning are certainly hot topics on everyone’s lips lately. With AI seemingly present in every facet of modern life, from checking in with Siri or Alexa about tomorrow’s weather to the latest tools like ChatGPT redefining how we interact with AI, the possibilities are seemingly endless.
Many people know the surface-level basics of what AI is and can do, but may not know just how much AI affects the financial sector. A lack of understanding often leads to mistrust of the technology, especially when money is involved.
Understanding Explainable AI and what it could mean for you will open your eyes to a whole new world of possibilities on how AI can transform your Credit Union, Bank or Insurance Provider.
Explainable AI
Unlike “black box” machine learning, in which even its designers are unable to explain why an AI model made a specific decision, Explainable AI includes the results of the solution readily understandable to humans. This understanding is important to have so that both engineers and management benefit from identifying areas of weakness and bias, and users can more actively engage with it
How to Measure Accuracy
Part of what is appealing about AI is the accuracy associated with it. To understand this accuracy, in a confusion matrix analysis, the model prediction is compared to the real outcome known from the history of the data, and the number of true positives, negatives and false positives is reported.
- True Positive Rate
Proportion of applications that were correctly predicted as approved. - False Positive Rate
Proportion of applications that were incorrectly predicted as approved. - True Negative Rate
Proportion of applications that were correctly predicted as declined. - False Negative Rate
Proportion of applications that were incorrectly predicted as declined.
These rates can be shown as follows:
- True Positive Rate
True Positive / True Positive + False Negative - False Positive Rate
False Positive / False Positive + True Negative - True Negative Rate
True Negative / True Negative + False Positive - False Negative Rate
False Negative / False Negative + True Positive
For simplicity, we can shorten these terms to be:
- TP: True Positive
- FP: False Positive
- TN: True Negative
- FN: False Negative
A model’s predictive accuracy is defined by how many accurately predicted cases are included among all the cases. The formula for accuracy is defined as follows:
TP + TN / TP + FP + TN + FN
When using predictive models to make credit decisions, a low false-positive rate is much more important than a low false-negative rate. To understand this way of thinking, a real-world, practical example can help:
A person would like to borrow money from a lending company. If the lending company makes the false prediction that this person will be a good credit risk, the lending company can suffer. Giving credit to the “good” credit risk, which turns out to actually be bad, is going to cost the lending company money if the person fails to make payments.
In another scenario, a person is classified as a bad credit risk and denied a loan. If this prediction turns out to be false and the person was actually a good credit risk, the worst that has happened is that the lender denied the person a loan.
If we compare the outcomes of having a false positive and a false negative, it’s easy to see that predicting a false positive will be far more detrimental to the business than a false negative.
The Importance of Context
When assessing the accuracy of predictive models, context is crucial. Medical diagnoses, for instance, can also result in false positives and false negatives. Again, a practical example can demonstrate this:
In this scenario, a patient comes in complaining of an illness. The doctor examines this patient, does some testing, and, upon the results of the test, comes up with a diagnosis and treatment plan.
In one case, we have the test come back with a false positive. Because the test came back positive, the doctor makes their diagnosis and prescribes a course of treatment. In this case, any treatment the patient receives is unnecessary.
In the second case, the tests come back with a false negative. Because the tests came back negative, the doctor does not make a diagnosis or make the wrong diagnosis. No treatment or the wrong treatment is prescribed for the patient. It is easily deduced that a false negative can lead to a missed diagnosis of serious diseases and missing the opportunity for treatment of said disease.
By looking at these examples along with the previous scenario with our borrowers, we can see the difference in consequences of false-positive and negative results. In the medical scenario, a false negative is much worse than a false positive. In a financial scenario, it’s the opposite – a false positive is worse than a false negative.
There Are More Ways To Measure Accuracy
Receiver Operating Characteristic Area Under Curve (ROC- AUC) score
This method helps one to visualise how well a classifier is performing. The Receiver Operating Characteristic (ROC) curve is a graph that plots the true positive rate against the false-positive rate at various thresholds. The Area Under this Curve (AUC) measures how likely a classifier is to rank a positive instance higher than a negative instance chosen randomly.
Cohen-Kappa Coefficient (κ:)
As a statistical measurement, Cohen’s Kappa identifies how often two raters are reliable, as well as how often the two raters agree when they are rating the same quantity.
Specificity
Specificity indicates how accurate a diagnostic test is at identifying normal (negative) conditions. It refers to the proportion of the true negatives correctly identified by a diagnostic test. Accuracy is the proportion of true results, either true positive or true negative, in a population.
Cross-Validation Mean Score:
Cross-validation refers to the evaluation of a model by resampling based on a small sample size. In k-fold cross-validation, a specific value for k is chosen. The sample is split at random into k equal-sized sub-samples. Following this, the mean accuracy score is calculated.
Test Error Rate
The test error rate refers to the frequency of errors that occurred. The formula to determine test error rate is (1 – Accuracy)
Prevalence
The accuracy of the model if the majority class was always predicted (baseline metric).
Where do I start?
AI, automation and other forms of digitisation may seem overwhelming at first glance. The complexity lies in how to make digitisation a reality in a traditional organisation without any external standards to measure how well you are performing. At Luna Connect we have developed a free tool to help you plan your automation journey, check out our lending automation assessment to find out how you can start your own journey BENCHMARK YOUR LENDING OPERATION USING THE 6 LEVELS OF LENDING AUTOMATION
Contact us at hello@lunaconnect.io
