Credit Card Fraud Detection With Machine Learning

Product
Creativity

Credit card fraud is an ever growing, costly problem. And despite significant advances in risk management techniques, fraudsters are getting harder to spot by traditional fraud detection software.

Today’s businesses are facing an increasingly sophisticated enemy that attacks, responds, and changes tactics extremely quickly. In order to get ahead of threats, businesses need to be able to respond quickly to changing patterns and behaviors. 

This is why many businesses are turning to advanced technology like AI and machine learning. The level of insights it delivers enables businesses to be agile in and get ahead of potential threats. 

Credit Card Fraud: What is it?

Credit card fraud is an inclusive term for the unauthorized use of a payment tool, such as a credit card or debit card, to fraudulently obtain money or property.

Credit card fraud is a nightmare for both individuals and businesses alike, but particularly stressful for cardholders, as it is one of the most common types of identify theft.

Not to mention that with the evolution of the internet, and the endless supply of eCommerce sites that came with it, credit card scammers have a much easier time accessing cardholder details.

Types of credit card fraud

There are five types of credit card fraud:

  • Card-not-present (CNP) fraud 
  • Counterfeit and skimming fraud 
  • Lost and stolen card fraud 
  • Card-never arrived-fraud 
  • False application fraud 

Credit Card Fraud Stats

Identity theft and credit card fraud are two of the most common financial crimes, and each of them saw significant growth in 2020. Part of the growth was due to the pandemic, as government benefits such as stimulus checks and unemployment insurance payments were targets of fraudsters.

Another reason why we saw such a huge increase in 2020 was the unparalleled surge in online shopping during COVID-19.

  • There were 1,387,615 reports of identity theft in 2020.
  • Credit card fraud rose by 44.7% over 2019 levels to 393,207 reports.
  • New account fraud has seen significant growth in recent years. It previously increased by 24% in 2018 and 88% in 2019. 

When you think about preventing credit card fraud, the first thing you think of is protecting your card information. But as the statistics show, you’re actually far more likely to have someone open up a new account using your personal information than use your existing card information. 

This is because it’s easier for criminals to steal money through a new account, since it’s an entirely new account that the cardholder doesn’t know about. With an existing account, the cardholder, as well as the credit card company, may notice suspicious activity and lock the card.

As well, data breaches have exposed information for hundreds of millions of people. Identity thieves can use this information for new account fraud.

Why Traditional Credit Card Fraud Detection Doesn’t Work

Credit card fraud detection relies primarily on identifying fraudulent credit card transactions and stopping them before they are accepted.

This process is traditionally known as rules-based fraud detection. It’s a pre-programmed set of rules designed to identify changes in behavior to determine which transactions are fraudulent and which are normal.

While this traditional method to detect fraud worked for many years, and was the best option available to banks and other service providers, it can’t keep up with the advances in technology and our variety of modern-day financial transactions. Moreover, scammers are getting more sophisticated in their fraudulent transactions. 

As a result, the traditional rules-based system isn’t able to keep up with the constantly evolving nature of credit card fraud. 

Rule-based vs. Machine Learning 

While rules-based systems rely on pre-programmed rules to spot behavioral changes or predict outcomes, rapidly evolving machine learning systems provide a more flexible approach to identifying fraud.

Here’s a few reasons why machine learning outperforms rule-based systems:

Fraud happens fast: Rule-based systems keep companies reacting. This reactivity means that credit card fraud is only tackled after it’s happened, which costs the consumer and financial institutions big. But with machine learning, models are updated in real time. Which means any fraud detected can be shared in seconds to prevent an attack. Companies can be proactive in preventing fraud.

Fraud hides under a huge quantities of data: It can be months before someone detects fraudulent activity. The best way to detect fraud in such huge volumes of data is through machine learning. Machine learning enables real-time insights, identifies irregularities, and detects subtle changes in large chunks of dynamic data sets.

Fraud is constantly changing: Fraudsters are always changing their tactics, which makes it incredibly difficult for humans to detect—and impossible for static rules—based systems, which don’t learn. Machine learning, on the other hand, is skilled at detecting patterns, at scale, and can adapt to changing behavior.

How Machine Learning Helps Prevent Credit Card Fraud

Traditionally, financial institutions entrust this task to rule-based systems that employ rule sets written by experts. But now they increasingly turn to a machine learning approach, as it сan bring significant improvements to the process.

Financial institutions can leverage machine learning to see:

  • Greater accuracy 
  • Less manual work
  • Fewer declines of normal transactions
  • Ability to adapt and evolve 

How Obviously AI Detects Fraud

No-code machine learning tools like Obviously AI, which are significantly faster and more flexible, help organizations fight fraud at scale. All that’s needed is historical data (containing both fraudulent and non-fraudulent transactions) and users can not only accurately detect suspicious activity in real time, they can even predict whether a new transaction is likely to be fraudulent.

And the great thing is: all of this can be done in seconds with Obviously AI.

For instance, on our platform you can:

  1. Use the Banking Fraud sample dataset (in our data store) to later structure your dataset; or 
  2. Upload your own machine learning-ready dataset and start making predictions right away 

We'll use our sample dataset as an example here. 

All we need to do is select the column “is_Fraud” to predict, and hit “Start Predicting.”

And just like that, we’ve generated a machine learning report that can show us the chances of fraudulent transactions. All in a matter of seconds. 


Behind the scenes here, Obviously AI is running the dataset through standard preprocessing stages before moving onto training multiple models. Once training begins, the backend runs various hyperparameter combinations of those models. This means that each model/algorithm is being tried with different settings, mixing and matching them, resulting in 10,000+ combinations of algorithms being ran in parallel.

Finally, the top two best combinations of each algorithm are selected by the platform. 


Related Reading: How No-code Machine Learning Algorithms Work

This all happens in seconds. The report generated displays the best combination of the top performing algorithm/model and you can now use this trained model to make predictions on new data! 

Let’s take a look at the report itself.



Here, we can see the top drivers related to the banking fraud dataset, such as the transaction amount, the sender’s old balance, the receiver's new balance, the type of transaction, and more, ranked in order of their impact on predicting outcome. 

You can start to explore the top drivers by toggling each one to see their impact. 



For instance, a close look at the distribution graph of the top driver, ‘transaction_amount,’ shows that there is a 36% chance that a fraudulent transaction is occurring if the transaction amount is $250,000.

However, the chance of it being a fraudulent transaction increases dramatically when that number increases: there’s an 80% chance that a $750,000 transaction is fraudulent.


Diving into the deeper details in the Overview tab, you can see how payment methods compare to each other. There are four types of transactions:

  • Cash-in (deposit)
  • Cash-out (withdrawal of cash)
  • Payment
  • Transfer 

It looks like deposits and payments indicate 0% chance of fraudulent transactions.

On the other hand, “cash_out” transactions have over 50% chance of being fraudulent, and transfers have the highest rate of being fraudulent at 80%.


With this kind of knowledge, it’s now incredibly easy and fast to not only spot fraud, but prevent it from happening with new customers as well. For instance, in the Predictions tab, we can input different values for each of the drivers to see whether a transaction would be fraudulent or not. So, we can see that a transfer with an amount of $325,000 has a 78% chance of being a fraudulent transaction.

Moreover,Obviously AI’s export feature means that you can share your trained models so that anyone, even if they don’t have an Obviously AI account, can use the trained model to make their own predictions. 

Which is fantastic—when you make predictive analytics available to everyone, no matter their technical background, you empower your team to become data-driven and ensure that results are consistent. 

Case Study: No-code AI and Fraud Detection

Online retailers currently deal with around 206,000 attacks on their stores each month. As online shopping grows, so does the opportunity for cybercriminals to scam online businesses. And one the types of fraud they encounter is affiliate fraud.

In affiliate marketing, online merchants pay affiliates a commission for sales that affiliates refer. When a shopper clicks on one of the links that an affiliate provides and makes a purchase, the merchant rewards the affiliate for the referral by giving the affiliate a commission. This is usually a percentage of the sale price.

In affiliate fraud, criminals game the system and defraud the online merchant using fake activity to either generate commissions or to increase the amount of the commissions.

A well known mobile app company in Palo Alto wanted to fight affiliate fraud to save up on piling platform usage costs. They decided to turn to Obviously AI to build their fraud detection engine.


The Solution

With Obviously AI, the company instantly connected their historical data of customer demographics and transaction history such as date, type, amount, etc.

In just seconds, they had fully trained an AI model that accurately predicted the costs an individual is likely to incur and integrated these predictions directly into their internal dashboard using Obviously AI’s API.

The Results

Obviously AI was able to identify 73% of fraudulent transactions, which enabled the team to take preemptive action prior to them clearing out. 

As a result, they say a 10x increase in internal efficiency, allowing them to save significant transaction costs.

Summary

Machine learning has the potential to help save financial institutions billions of dollars in fraud losses. And with no-code machine learning, it's never been easier or faster to accurately detect, predict, and prevent those losses.

To learn more about how Obviously AI can help you with your use case, reach out to our team to book a demo.


Become the Data Scientist your team always needed.

Get Started

Get Started Now

See how no code machine learning can transform your business and change how you make decisions.