Terminating costly fraud: the rise of machine learning

Like all businesses, online retailers need to keep revenue high and costs low in order to stay profitable. Chargebacks – and the inadequate prevention of them – bleed profits in two ways. First, transactions which are approved and allowed to go through but later turn out to be fraudulent incur losses due to costly chargebacks (and the associated fees). Secondly, legitimate transactions which are erroneously declined decrease revenue, both for each single transaction, and possibly much longer as each customer impacted is then driven to seek a more frictionless shopping experience elsewhere.

With stakes this high, advances in artificial intelligence can provide a competitive edge, but in order to deliver the best results, machine learning algorithms need to be fed real-world data – and lots of it.

Feeding the Machine

In order for machine learning to work, the machine first needs data to learn from. Since both criminal and friendly fraud purchases are rare events, it’s a real challenge gathering enough examples of sketchy transactions for the system to develop, test and refine its model.

Riskified, a provider of ecommerce fraud prevention software was able to hook early adopters by their enticing chargeback guarantee. In exchange for data on all the transactions which they had already decided to decline, the firm offered to strain out any offers its system deemed good but was rejected by the merchant. Since only Riskified would be financially on the hook for the transactions it approved, the merchants had nothing to lose and some margin to gain. With a few of these merchants feeding Riskified’s system data, the supervised machine learning algorithm (based on deep learning) had enough legit and fraudulent transaction data to build its own filter.

This illustrates the key benefit of a SaaS/cloud provider for services like this: you benefit not only from your data, but also from the data of your provider’s other customers. This is especially true of machine learning based offerings, because more data means better decisions. It’s ultimately those decisions: the split-second judgment calls on every incoming transaction which determine either revenue or loss, satisfied customers or lost customers, beating back online criminals or siphoning money to them.

Machine learning is a necessity

The old fraud detection standbys of software decision trees and risk scorecards aren’t quick enough to keep up with the blistering pace of today’s ecommerce, especially at times of high transaction volumes. Newer approaches to fraud detection, ones based on machine learning, solve the main problems of modern fraud detection:

Scalability: Online retailers, especially startups and successful mainstays, have many users from multiple countries, sell multiple products, and gather long transaction histories on each customer. There are even more features of each potential transaction and each one can be a clue to whether that order is fraudulent or legit. Human analysts, who can be very good at spotting suspicious transactions, aren’t able to keep up with the deluge of orders which stream in during peak business seasons. As your product line increases and expands, complicated decision trees need to be modified and new weights assigned, thus semi-automated techniques fall short due to the near-constant manual tuning and maintenance required.

Speed: Humans can take several seconds to make a decision, whereas computers need only milliseconds. During periods of rapid growth or high demand, those extra seconds can become an insurmountable bottleneck.

Adaptability: Accuracy is easy when you’re aiming at a stationary target. Nothing about fraud detection, however is stationary. Fraudsters are constantly learning from their failed attempts (declined transactions). Also, legit transaction patterns fundamentally and abruptly change. Want an army of unhappy customers due to false positives, or to hemorrhage money due to chargeback-related losses? Accuracy must be matched by intelligent adaptability, which isn’t possible with software decision trees and manually set weights.

Cost: People are far more expensive than servers. Computers run 24/7/365 and don’t need coffee breaks, or benefits. Cloud-based SaaS solutions are even more economical as capacity can scale along with demand.

The costs of fraud is a major concern for today’s online merchants. Machine learning, Big Data and the cloud have converged to spawn a new generation of solutions which can do a fine job in protecting your business.

Image by geralt pixabay