Fraud is the scourge of banking, eCommerce, and financial services in the 21st century. Every online merchant faces fraud and losses from it sometimes rack up to 10% of total revenues. Obviously, finding a reliable countermeasure that ensures fraud prevention is imperative for businesses that want to secure their bottom lines from unneeded expenses. Employing a machine learning algorithm to fight fraud can be a lifesaver here. As an end-to-end anti-fraud system, Covery explains why it is so, based on our experience with deploying ML models for fraud detection and prevention.
As defined in Cambridge Dictionary, “fraud is the crime of getting money through deception”. It can be aimed at any party, as scammers can defraud either merchants, customers, or both. However, fraudulent transactions always have certain flags or features that can help distinguish them from legitimate ones. The problem is, even with hundreds of transactions daily, checking each one manually is hard, and when we talk about thousands — it’s downright impossible. This is where using a Machine Learning algorithm is a better option.
How is the Machine Learning algorithm used to fight fraud?
Machine Learning is a subset of AI research focused on teaching mathematical models to identify patterns in data. This way, a machine learning algorithm can analyze the historical data, identify the patterns in it and learn to do it with real-time data flow, identifying abnormal (fraudulent) activity on the fly and making decisions that help keep fraud at bay.
What does it actually mean?
Every transaction has at least 12 unique identifiers — IP address, email address, geolocation, email domain, IBAN or BIC number, etc. Trustchain is a global reputational record database developed by Covery and updated by all members of the Covery community. It currently has more than 500 million records, and whenever any online merchant identifies a customer or transaction as a fraudulent one, this information is automatically propagated to all the other community members.
Thus said, not all such identifiers are necessarily fraudulent ones. For example, a fraudster once used a public IP address (of a coffee-house, a library, or an airport, etc.). Naturally, this does not mean that all other traffic from this IP address is fraudulent too. However, this makes all transactions from this IP address slightly riskier for some period of time. Let’s now say the fraudster used the *@gmail.com email address to register their account. Does this make Gmail fraudulent? Of course not. But it adds several more points to the overall risk score for this transaction.
Thus said, the next time this customer tries to use any of these credentials with your website or with any other member of the Covery community, the profile will be marked as a potential fraudster at once, and all transactions from it will have to undergo tighter scrutiny.
Yet another use for a machine learning algorithm is evaluating the risk score of any transaction in real-time. Due to assessing all the available risk factors, using ML allows coming up with precise risk scores at scale. This helps merchants make informed decisions and decline potentially fraudulent payments.
Which machine learning algorithm is used in fraud detection?
There are several possible approaches to training a machine learning algorithm:
- Unsupervised — when the model has to learn and discern the needed patterns based on an unlabeled data set, which allows training to discern a variety of patterns
- Supervised — when the model has to learn based on a labeled data set, which makes learning faster and results more precise, but is more costly, as it requires human experts to label the training data set
- Reinforced — when the model is able to determine optimal behavior based on a reinforcement feedback signal.
Due to the specifics of Fintech, eCommerce, and other services involving online transactions, supervised learning works best. Data scientists training the model know exactly the result they want to achieve. Thus, training takes less time and money while providing accurate analysis based on a wide variety of factors. Most importantly, Covery has already trained its machine learning algorithm on cases similar to yours, so one of the biggest expenses is discarded at once.
How does a machine learning algorithm detect credit card fraud?
As we explained above, a machine learning algorithm trained in fraud detection scans every transaction in real-time for signs of fraud. Whether these signs include identifiers flagged as fraudulent, or simply differ much from normal patterns for this account — an online merchant gets a detailed report.
What’s more important, with Covery you can use an in-depth risk logic rule engine to configure scenarios for any risk score. For example, a customer of yours who usually logs to your eCommerce platform in the evening from Dallas, Texas and orders some minor grocery, suddenly logs in close to noon from Miami, Florida and orders several items of jewelry.
Several conditions there increase the risk score, so a machine learning algorithm automatically launches a specific scenario to deal with a potential fraud case. For example, a 3D Secure check procedure can be invoked. A legitimate customer will have no trouble entering a one-time password or a CVV code, while a fraudster will be cut off.
Conclusions
Deploying a machine learning algorithm for fraud detection is one of the best business decisions you can make. The ML model works equally well with tasks of any scope, evaluates all the available risk factors in under a second, can automatically invoke appropriate business logic scenarios and only gets better and more precise over time.
Covery is a comprehensive anti-fraud solution available in the US, the EU, the UK and worldwide. We deploy a supervised machine learning algorithm to provide our customers with the best fraud detection and prevention capabilities.
Should you want to know more about the value Covery can provide for your business — contact us for a free demo, we are always glad to assist!