Authorization Rates Series: Part 2
October 22, 2020 | Jim Magats, SVP Omni Payments
Authorization rates are one of the most crucial ways for merchants, small and large, to unlock revenue. Small improvements in authorization rates can make a difference of millions of dollars of volume processed for global enterprises. For all merchants, the goal is for a customer to complete their purchase successfully on the first try, and not have to worry about the payment.
As a result of a robust data set, machine learning, artificial intelligence, network tokenization, strong partnerships with networks and issuers, and multiple-funding instruments in our digital wallet, PayPal is uniquely suited to offer higher than industry average approval rates for merchants. In the last three years, we’ve been able to improve our global authorization rate by more than 300 basis points (bps) for our branded processing. For new users signing up for PayPal, we’ve been able to increase authorization rates by an average of 600 bps, and these number continues to rise.
In our first post of the Authorization Rates Series, we explored how network tokenization leads to higher authorization rates. Today, we are going to look at additional key factors that help improve authorization rates, data science, machine learning and artificial intelligence, and four reasons why they’re so powerful.
The Four Contributions of Data Science, Machine Learning and Artificial Intelligence to the Checkout Process
Optimizing the payment experience: Ensuring high probability of approval & decline prediction
We’ve all been there. We are ready to make a high-value purchase that we’ve spent weeks thinking about, maybe a new gaming console, a designer handbag or a vacation rental. And surprise, surprise, the transaction is declined, even though we are well below the card limit.
But now, by combining historical trends and transaction variables, PayPal’s machine learning models can help predict in advance if a user’s card will be declined for a transaction and prevent the purchase from being completed. If a decline is predicted, we can create a custom experience for the user that will ensure a valid purchase goes through, which has resulted in improvement of auth rates between 60-240 bps for certain merchants. For example, we can surface a different payment method within a user's PayPal wallet that has been successful in the past. We can also prompt a secondary form of authentication using 3D-secure or request a CVV to make sure the user is who they say they are.
In the event of a decline, the data model is able to determine if the payment can be recovered by user action. If so, the user receives an explanation for why the payment was declined, such as a non-sufficient balance or presumed card information stolen. We can then suggest an easy way to update card details such as recommending adding funds or even switching to another funding instrument. This process is an example of explainable AI, making AI actionable and available for users to implement a self-service solution.
“Stand Ins:” Approving a purchase for VIP’s and loyal customers
System outages are rare but costly and they can happen when a merchant least expects it. We most often see this with external party issues such as loss of connectivity, preventing any transaction from taking place until the system is back online. When this happens, being able to still approve a purchase without processing the transaction immediately is key for business continuity.
Data science models can help us identify good transactions, so that we can “stand in” for a purchase and make sure it goes through when merchants face technical issues that interrupt transaction processing. We can also take this one step further by offering VIP stand ins: for highly engaged and loyal customers, we are able to “stand in” for a purchase regardless for the reason of a decline. Both scenarios minimize losses for a merchant and gives the consumer confidence that their payment will go through when using PayPal. We are then able to recover the transaction value when the systems are back online through a smart retry recovery model.
Protecting against fraudulent transactions
Not every transaction is well intended. It is important to balance higher approval rates with keeping fraud in check. For merchants, an increase in fraudulent transactions can mean a decrease in profit: Juniper Research predicts that global merchants will lose a combined $130 billion between 2019 and 2023 via card-not-present (CNP) fraudulent transactions.
Fortunately, machine learning and real-time decisioning can help differentiate between good and bad transactions. For example, data science helps allow us to quickly identify fraudulent patterns, such as carding attacks, an increasingly used technique by fraudsters where they test a batch of stolen cards on a merchant website to identify valid cards. This activity can lead to monetary loss to merchants in the form of unauthorized chargebacks or it can impact a merchant's infrastructure availability with excessive failed authorization attempts. PayPal’s advanced machine learning algorithms help us to defend against these types of attacks at scale.
But machine learning is only as good as the data set it is learning from, an area PayPal has a strong advantage in due to our two-sided network. With over 320 million consumers accounts and 26 million merchants accounts, PayPal has a wealth of data on consumers and risk profiles. This insight into both the merchant and consumer sides of the transaction helps allow us to make the call on a true or fraudulent transaction for even the most sophisticated fraud behavior.
Retrying a declined transaction because it isn’t successful the first time is not a new strategy and is one that should be used judiciously. PayPal’s machine learning algorithms can help identify the best retry strategy based on the card used, issuer, merchant, transaction-level parameters, processor and acquirer combination and even day and time of retry. We are also able to retry a transaction with a token or card number based on success patterns identified by machine learning models. By analyzing the data, we can leverage multiple strategies to determine the best times and methods to retry a payment, and continually reassess.
As a result of implementing a smart retry strategy, PayPal has been able to harness an incremental 30 bps improvement for certain domestic U.S. token transactions, supplementing the benefits of using a token.
However, to create the optimal customer experience, a merchant should focus on getting approvals on the first try. Beyond creating a sub-par customer experience, a retry strategy can also be expensive, result in chargebacks and leave an unfavorable impression of the merchant with issuers.
Put simply, everyone from small businesses to large enterprise merchants should be prioritizing authorization rates as even a half a percentage point in lost revenue can translate into large volumes. The higher the card authorization rate, the greater likelihood for conversion, resulting in higher revenue for a merchant and a positive consumer experience. Conversely, high declines interrupt the end consumer experience, resulting in lost sales and reputational risks from a merchant’s perspective.
We want to help both our merchants and consumers leverage data to make better choices right from the beginning of the transaction journey. Our machine learning models are built upon historical purchasing behavior and issuer data, and as a result, can make intelligent recommendations to maximize the chances of a successful transaction.
But all of this would not be possible without our team of seasoned data scientists who possess deep domain knowledge and help us to leverage cutting-edge machine learning and artificial intelligence technology for payments.
Ultimately, our goal is to help merchants optimize authorization rates to the extent that processing transactions feels seamless, and it is no longer something they even need to think about.
Authorization rate improvements cited in this article were derived from global PayPal transactions data and all improvements quoted are global averages; except otherwise mentioned. Improvements may vary by region. The average duration used to establish averages varies between 3-9 months depending on the feature.