Fraud Predictor · AI-powered fraud detection
Cut fraud in half without
turning away good
customers
Aryng's Fraud Predictor uses machine learning to combine your internal application, account and behavior data with external fraud scores, so you can stop more fraudsters at onboarding while approving more of the legitimate customers that rigid rules turn away.
The problem
Fraud is 2,300 years old. Today's fraudsters are just smarter and better equipped.
Circa 300 B.C.
The oldest recorded fraud can be traced back to around 300 B.C. A Greek merchant, Hegestratos, had taken out a large insurance policy called bottomry, under which merchants could borrow money and pay it back, with interest, upon delivery of their goods. Hegestratos decided to sink his boat, which was not loaded with any goods, to avoid paying back the loan. The plan failed, however, as he himself drowned while trying to sink the boat.
Fast forward 24 centuries, and we are still dealing with fraudsters. The only difference now is that the Hegestratoses of today are far more numerous, smarter, and better equipped to commit fraud, increasingly with automation and AI tools of their own. In light of this, almost every industry now has a risk team that tries to minimize fraud.
Fraud will keep happening, however, it can be minimized, and then minimized again. In that sense, building a fraud model to drive down fraud cases is a journey rather than a one-time fix, and the rule sets most risk teams start with rarely keep pace.
As part of our data science consulting practice, we have built highly sophisticated, machine learning-driven fraud models for our clients.
What is AI fraud prediction?
A fraud risk score for every applicant, built from more than your own data
Usually, very little is known about a customer before they commit fraud. A fraudster will open an online account and waste no time trying to steal money that isn't theirs, long before internal application data alone can reveal a pattern. A predictive fraud model uses machine learning to combine your internal application, account and behavior data with external fraud scores from credit reporting agencies, which may already know a fraudster's email ID, address, social media scores and cookies. The result is a fraud risk score your team can act on at onboarding and throughout the account lifecycle.
What an AI fraud model makes possible
Stop fraud at onboarding
Score applications before a new account can move any money.
Approve more good customers
Replace blanket rule-based rejections with risk-based decisions driven by AI.
Get more from external scores
Internal data typically adds 10 to 50% incremental lift on top of external fraud scores.
Tag instead of reject
Limit what risky accounts can do, then expand access as good behavior builds trust.
Cut losses across the board
Reduce losses from fraud, non-payments and chargebacks with one predictive model.
How it works
Aryng's Fraud Prediction Roadmap
Every engagement runs on BADIR, Aryng's proprietary five-step Data-to-Decisions framework, which keeps the AI and machine learning work tied to a business decision from the first stakeholder interview to the final recommendation.
-
B
Business question
An onsite assessment of the work done so far and the present scenario, plus interviews with every stakeholder, to pin down the real question and the trade-off between fraud losses and customer growth.
-
A
Analysis plan
We define what counts as fraud, non-payment and chargeback, agree on acceptance and fraud-rate goals, and pick a methodology that balances machine learning accuracy with the interpretability your team needs.
-
D
Data collection
We bring together internal application, account and behavior data with external fraud scores from credit reporting agencies, then validate it for gaps, irregularities and accuracy.
-
I
Insights
Feature engineering and predictive modeling with gradient boosting and other ensemble methods, built together with your stakeholders so the model reflects how fraud actually shows up in your business.
-
R
Recommendations
Decision thresholds tied to their impact on fraud rate and acceptance rate, plus a plan to operationalize the model in your onboarding flow.
Case study · Bravo Bank
How Aryng lowered the fraud rate by 50% for a fintech client while approving more applicants
About Bravo Bank
Bravo Bank (name changed to protect the identity of the organization) was a traditional retail bank making a foray into online banking. The bank hired Jim as its risk manager. Jim was new to online banking, however, he was well experienced in dealing with fraud in a retail banking environment.
At Bravo Bank, Jim was tasked with managing risk and reducing losses from fraud, non-payments and chargebacks. He knew fraud was going to happen here, and he was determined to live up to his reputation and make Bravo Bank fraud-free. What he didn't know yet was that he was up against smarter, better equipped fraudsters than any he had faced earlier in his career.
The real business question
How can Bravo Bank reduce losses from fraud, non-payments and chargebacks in its new online business without rejecting the good customers it needs to grow?
Online banking added a layer of complexity retail banks don't face: ID verification for customers the bank has never met in person.
The rules trap
Jim did what had always worked in retail banking: he wrote rules. Every time fraud got through, he tightened them. Every time, fraud continued and more good applicants were turned away.
-
Step 1
Retail banking rules
Jim drafted a rule set that had worked in retail banking. Leadership signed off, the rules went live and customers poured in. So did the fraud.
85% acceptance · fraud continued -
Step 2
Tighter rules
Fraudulent accounts shared certain IPs, zip codes, addresses and location mismatches. Jim added them to the rule set.
80% acceptance · fraud continued -
Step 3
Rules on rules
The process of adding rules was repeated. The stringent rules were now hurting onboarding badly.
~50% acceptance · fraud continued -
Step 4
Rules plus fraud scores
Jim combined external fraud scores from credit reporting agencies, including Experian, with his rule set.
Still too low too many rejections · fraud continued -
Step 5 · Aryng
Machine learning model
A gradient boosting model on internal data plus the external fraud score.
70% acceptance · fraud down 50%, to 2%
Why the old playbook failed
Jim had come to understand that his old methods were not working in the online environment, where ID verification added a layer of complexity that retail banks don't have. He needed fresh, innovative ways to tackle the problem, so he decided to use data from credit reporting agencies, including Experian.
Since credit reporting companies may already know a fraudster's email ID, address, social media scores and cookies, their fraud score is a good indicator. Combined with the rule set, however, it still wasn't enough: the bank kept rejecting too many customers and fraud continued. That's when Jim brought in the professionals at Aryng.
The Aryng approach
We carried out an onsite assessment, understood the work done in the past and the present scenario, interviewed all the stakeholders, and built a solution together with them.
Model inputs
Model
A gradient boosting tree, a machine learning ensemble method, predicting the likelihood of fraud for each application
The results
application acceptance rate, up from roughly 50% under the rules
fraud rate with the predictive model in place
reduction in fraud rate, while approving more legitimate customers
Why machine learning beats rules and traditional statistics for fraud
Fraud is a needle-in-the-haystack phenomenon, often occurring with little known history. There is less data to work with and even fewer continuous variables, so typical statistical methods produce limited results. For fraud models, machine learning methods such as neural networks, support vector machines, gradient boosting and other ensembles (combinations of many models working together) produce much better results.
The trade-off is that fraud models are often black boxes, where you can't fully understand the effect of any given predictor variable on the entire model. If your organization is ready to live without completely understanding the model, or to spend significant effort gaining partial interpretability through explainable AI techniques, machine learning and deep learning algorithms are a perfect choice for building fraud models.
Three tips on fraud modeling
What our data scientists tell every risk team
Combine internal data with external fraud scores
Build a refined fraud model by combining the two, since internal data typically adds 10 to 50% incremental lift to most fraud models.
Choose machine learning with eyes open
Machine learning algorithms are likely to produce better results than statistical processes, though they are harder to operationalize and interpret.
Tag accounts instead of rejecting them
Use the fraud score to tag potentially fraudulent accounts and give them limited functionality, then unlock more functionality based on customers' good activity over time.
Outcomes
What to expect from the Fraud Predictor
A risk score for every applicant
A validated, AI-driven fraud score built from your own data and external fraud signals.
Higher approval rates
Fewer false rejections, so more legitimate customers make it through onboarding.
Lower losses
Reduced losses from fraud, non-payments and chargebacks, measured against your current baseline.
Smarter use of external data
A clear view of how much lift your internal data adds on top of third-party fraud scores.
Tiered account access
A tagging strategy that limits risky accounts without losing the good customers among them.
Pair it with LTV
Combine fraud risk with the LTV Predictor to see which applicants are both safe to approve and worth acquiring.
From Aryng's data science team
"Fraud will keep on happening, but we can minimize it and keep on minimizing it. In a sense, having a fraud model to play down the cases of fraud is similar to a journey."
Aryng
Fraud Modeling 101
Frequently asked questions
What is a fraud prediction model?
A fraud prediction model uses machine learning to estimate how likely an application or account is to be fraudulent, based on internal application, account and behavior data combined with external fraud scores. It gives your risk team a score to act on instead of an ever-growing list of static rules.
Why not just add more rules?
Rules only catch the patterns you already know about, and every new rule also rejects some good customers. At Bravo Bank, stacking rules pushed rejections to almost 50% of applicants while fraud continued.
We already buy external fraud scores. Why build our own model?
External fraud scores are a good indicator, however, they don't see how customers behave on your platform. Combining them with your internal data typically adds 10 to 50% incremental lift to a fraud model.
Isn't a machine learning fraud model a black box?
Often it is, which is why we discuss the trade-off with your stakeholders up front. If your team needs to explain decisions, we can invest in partial interpretability using explainable AI techniques, at the cost of extra effort.
How does this work with the LTV Predictor?
The Fraud Predictor tells you which applicants are risky, and the LTV Predictor tells you which customers will be most valuable. Together, they help you decide whom to approve, whom to prioritize and where to invest.
Stop more fraud without stopping growth
Talk to Aryng's data science team about building an AI-powered fraud model on your own data.
About Aryng
Aryng is a data science and AI consulting, training and advising company. Aryng's SWAT data science team solves complex business problems, builds each company's Data DNA through data literacy programs, and delivers rapid ROI using machine learning, deep learning and AI. Clients include Google, Box, HERE, Applied Materials, Abbott Labs and GE.
For Corporations
For Indiviuals
Aryng
Policies
Copyright © 2011 - 2026 Aryng LLC. All rights reserved
