More and more of us are using PayPal and Venmo to make transactions rather than using credit card companies like Visa, Mastercard, and others. While there are multiple benefits to this approach, the main crux is ease of use. The only way this model works is if everyone conducts themselves honestly and responsibly. While that’s a wide umbrella to cast, the work of machine learning scientist Alban Zammit is focused on this very premise. He and his team at PayPal are celebrities for their work on the ACH Risk Model (ARM) which is estimated to have saved the company hundreds of millions of dollars in losses every year. For those of us who want to truly comprehend how our money is handled and why the process works so effectively in this modern manner of transaction, we want to provide an easy-to-understand observation of Mr. Zammit’s work and how profoundly impactful it is in the modern financial world.
We’re used to things happening instantly but that’s not actually the reality of it all. When you make a purchase through a company like PayPal, your bank is the funding source that PayPal contacts with a withdrawal request to the ACH network, which then tries to take the funds from your bank account. In the US, this process takes anywhere from three to five days to validate but some international transactions can take up to two or three months. Regardless of the timetable, there are numerous opportunities for a “hiccup” or bad actor to take advantage of the situation, leaving you (or PayPal) frustrated. To negate this, PayPal’s system makes the funds available right away even though the company hasn’t actually received the money from the bank. The small fee that PayPal charges for these ACH transactions are a substantial part of the company’s total transactions. ARM is a major safeguard for the company in dealing with potential problems. Historical data from billions of transactions are used to train ARM in spotting fraudulent and suspicious trends. Feature engineering is included to assess factors with the sender, receiver, and transactions type, allowing more risky situations to become highlighted. When you consider that PayPal handles forty percent of all e-commerce transactions on the planet, their data base of this type of information is massive. Written in Python, the binary classifier model training for ARM was a huge undertaking that was cloud based with multiple computers working in tandem due to the mammoth amount of information. It’s essential to understand that the creation of ARM is much more than simply coding, the design and engineering process honed the initial concept into a tool which safeguards PayPal and its users in a revolutionary way.
One of the most challenging aspects of ARM was the company’s insistence that it work not only effectively but quickly. Their requirement of a 100-millisecond limit to ensure a smooth user experience for users was achieved through “adaptive header architecture” which establishes the maintenance of high accuracy while minimizing latency. Alban explains, “As a Machine Learning Scientist, I balance the need for speed against the complexity of the model. Typically, a more complex model provides better accuracy, but it also takes longer to process. This trade-off is like choosing between different versions of ChatGPT: smaller models give quicker responses, while larger models offer more detail but take more time. The “size” of a model, in technical terms, usually refers to its number of trainable parameters. For example, a simple linear regression y= a.x+b has just 2 parameters: a and b. In contrast, ARM has millions of trainable parameters and models like ChatGPT which can have billions. The larger the model, the more time it takes to produce results. Given the 100ms limit, ARM can’t rely on overly large or complex models. Instead, we use streamlined neural networks with just a few layers or gradient-boosted trees with controlled depths to meet this time constraint without compromising on performance.” He continues, “The adaptive header architecture solved our problem because instead of calculating every specialized score and then combining them, ARM starts with a shared model structure for all transaction segments, followed by a specialized ‘header’ tailored to the specific type of transaction. This means that ARM only processes the relevant segment, minimizing unnecessary computation and meeting the strict latency limits. This adaptive header approach lets ARM retain—and even exceed—the performance of the previous system, with far faster processing times, making it possible to stay well within the 100ms limit.”
Consider that somewhere between one and two trillion dollars will transition through PayPal this year and you can glimpse how a fraction of percent in the model performance of ARM can mean in saving the company in regard to loss prevention; literally multi-millions. Innovations like the use of adaptive header architecture in ARM has resonated throughout the entire company, providing mutually increased satisfactions and security. When you make your next purchase, the skill of Alban Zammit and his team at PayPal should serve as a reminder that good tech is on your side in ways that you may never consciously comprehend.
Writer : Basil Thomson