What it means
Traditional software follows instructions: if a payment is over a certain amount, flag it. Machine learning works the other way round.
You give the system many past examples, such as transactions labelled as genuine or fraudulent, and it works out for itself which patterns separate the two. There are three broad styles.
In supervised learning, the system learns from labelled examples, which suits fraud detection and credit scoring. In unsupervised learning, it looks for natural groupings in unlabelled data, such as customer segments, and in reinforcement learning it learns by trial and error against a goal.
For a finance team, the appeal is speed and scale. A model can review millions of transactions a day, update itself as behaviour changes and pick up weak signals that humans would miss.
Typical uses include forecasting cash flow, matching invoices to payments, flagging unusual journal entries and estimating the chance a borrower will default. The risks are just as important.
A model trained on biased or incomplete data will produce biased or incomplete answers, which is why people say Gold in, Gold out. Models can also be hard to explain, which is a problem when a regulator, auditor or customer asks why a loan was refused.
Good practice therefore involves testing a model on data it has not seen, tracking its accuracy over time and keeping a person responsible for its decisions. Models decay as the world changes, so a fraud model trained on last year's tactics may miss this year's.
A final nuance is that accuracy alone can mislead. If only 1% of transactions are fraudulent, a model that always says genuine is 99% accurate and completely useless, so measures such as precision and recall are used instead.
In practice
Real-world examples.
Example
A credit card issuer trains a model on millions of past transactions to spot suspicious spending. When a card used in London is suddenly used for a large purchase in another country minutes later, the system blocks it and sends the customer a text.
Example
A retailer uses a forecasting model to predict weekly demand for each product in each store. Because the model learns from sales, weather and promotions, stock levels are better matched to demand and write-offs for unsold food fall.
Example
An accounts payable team uses a model to match incoming invoices to purchase orders. It learns the layout of each supplier's invoices and clears most items automatically, leaving staff to check only the exceptions.
Formula
Calculation
Precision = True positives / Everything the model flagged, and Recall = True positives / All actual cases
A bank tests a fraud model on 10,000 transactions, of which 160 are truly fraudulent. The model flags 100 transactions, and 80 of them turn out to be fraud. Precision = 80 / 100 = 80%, meaning four out of five alerts are real. Recall = 80 / 160 = 50%, meaning the model catches half of all the fraud. If each missed fraud costs $500, the 80 missed cases cost 80 x 500 = $40,000.Case study
Seen in the real world.
Calloway Credit is an illustrative, fictional lender that decided to replace a simple scorecard with a machine learning model for small business loans. The data team trained the model on five years of loan history, including repayment behaviour and bank account activity. On test data it ranked borrowers by risk far better than the old method.
Before launch, the risk officer asked for an explanation of why particular applicants were declined. The team found the model leaned heavily on postcode, which acted as a stand-in for factors the lender was not allowed to use. They removed the variable, retrained the model and accepted a slightly lower accuracy.
After a year, the illustrative lender found that default losses fell by about $1,200,000 on a book of $60,000,000. The finance director also set a quarterly review because the model's accuracy would drift as economic conditions changed.
Watch out
Common mistakes.
- Believing a model is objective because it is mathematical, when it learns whatever patterns and biases exist in its training data.
- Judging a model only on overall accuracy, which can hide poor performance on rare but costly events like fraud.
- Building a model and leaving it unattended, when its performance can decay as customer behaviour and the economy change.
Questions
People also ask.
Is machine learning the same as artificial intelligence?
Machine learning is a subset of artificial intelligence, which is the wider goal of making computers perform tasks that normally need human intelligence.
Do you need a data scientist to use it?
Many tools now offer ready-made models, but someone with data skills is still needed to check the data, test the results and explain the output to decision makers.
How much data does a model need?
It depends on the problem, but more varied and accurate data generally helps, and quality matters more than sheer volume.
From the founder's library
Take it further with the book.
Build your financial confidence beyond this definition. Shihan's full-length guide, Accounting Fundamentals, takes the same plain-English approach and turns it into a complete, practical playbook for non-finance managers, business owners and students - with chapter-end quiz answers and presentation slides included.
25% off with code MMHQ25, applied at checkout. Priced in USD - checkout may show the equivalent in your local currency.
View the book and save 25%