E-commerce Fraud Facts(blog.siftscience.com)
blog.siftscience.com
E-commerce Fraud Facts
http://blog.siftscience.com/five-ecommerce-fraud-facts/
12 comments
Hey Aaron, I'm the OP, and CEO of Sift Science. Sorry that your experience has been subpar. I'll follow up offline.
I think there's some confusion -- we actually do have a global machine learning model, which is what customers will start with if we haven't received any labels (training examples to learn from). But, your mileage may vary with the global model. We think it's a starting point, but that fraud differs from site to site in subtle ways, and to be truly effective you need to have your own model. This is why training examples are so critical. Once we receive enough training examples, we're able to build a model specific to your site.
That said, I'll be the first to agree with you -- we haven't done a great job setting the right expectations and messaging, especially on the point above (that your mileage may vary with the global model). We've been working hard to improve this, and there are now several reminders in the console, and an introductory tour, that emphasize the importance of a thorough integration with labels to achieve best results. With machine learning especially, bad data in = bad results out. Also, our first UI/UX designer started a month ago, and improving this experience is a top priority.
With regards to making fraud facts more functional and actionable -- we actually do provide reasons we think a user is suspicious in our web console and API (see https://siftscience.com/docs/getting-scores). But, there isn't an "aggregate learnings" page like what's presented in the blog post -- that's on the 3-month roadmap.
Again, sorry for the subpar experience, and I'll follow up with you separately. We have a very technical product -- we're working hard to abstract the complexity, while making sure we set customers up for success. We've also been overwhelmed with customer demand, and have been scaling the team to keep up (7 FT employees two months ago, 15 now). That said, no excuses. Just know that we're not happy to hear about experiences like yours, and are working hard to make it right.
Thanks, Jason
(EDIT) Aaron and I just had a good phone call -- we're going to make things right.
I think there's some confusion -- we actually do have a global machine learning model, which is what customers will start with if we haven't received any labels (training examples to learn from). But, your mileage may vary with the global model. We think it's a starting point, but that fraud differs from site to site in subtle ways, and to be truly effective you need to have your own model. This is why training examples are so critical. Once we receive enough training examples, we're able to build a model specific to your site.
That said, I'll be the first to agree with you -- we haven't done a great job setting the right expectations and messaging, especially on the point above (that your mileage may vary with the global model). We've been working hard to improve this, and there are now several reminders in the console, and an introductory tour, that emphasize the importance of a thorough integration with labels to achieve best results. With machine learning especially, bad data in = bad results out. Also, our first UI/UX designer started a month ago, and improving this experience is a top priority.
With regards to making fraud facts more functional and actionable -- we actually do provide reasons we think a user is suspicious in our web console and API (see https://siftscience.com/docs/getting-scores). But, there isn't an "aggregate learnings" page like what's presented in the blog post -- that's on the 3-month roadmap.
Again, sorry for the subpar experience, and I'll follow up with you separately. We have a very technical product -- we're working hard to abstract the complexity, while making sure we set customers up for success. We've also been overwhelmed with customer demand, and have been scaling the team to keep up (7 FT employees two months ago, 15 now). That said, no excuses. Just know that we're not happy to hear about experiences like yours, and are working hard to make it right.
Thanks, Jason
(EDIT) Aaron and I just had a good phone call -- we're going to make things right.
Thanks for the call, for listening, and keep up the great work!
Yeah we work with one of their competitors and it is not simple. It takes an investment and thoughtful analysis to get use out of these services. No surprise. It's an inexact science.
But, they really shouldn't advertise it as easy to use.
But, they really shouldn't advertise it as easy to use.
I disagree: Sift can be easy to use. One of the features of Sift is its flexibility. You can make it easy to use and get some additional anti-fraud protection. Or you can make it more complicated but get better results.
Disclosure: I'm a happy current customer whose model isn't quite right yet, but is getting there.
Disclosure: I'm a happy current customer whose model isn't quite right yet, but is getting there.
>You can make it easy to use and get some additional anti-fraud protection
Well, that's kind of the point: sure, you can get it to function easily, but getting it to actually provide you with solid coverage takes some effort.
>Disclosure: I'm a happy current customer whose model isn't quite right yet, but is getting there.
That pretty much seals my point.
I mean, who wants semi-effective fraud protection? That's where the extra work comes in. And what you will find is that fraudsters change their game and you have to evolve your model, even once you get it customized. So, if you are simply using the "easy-to-use" out of the box signals, you are going to find that fraudsters have already been trained by other Sift customers. Sure, it might take them a moment to realize that you have implemented these measures, but they will adapt quickly when they do. If not, then you are dealing with amateurs and you really don't have much of a fraud problem.
And, that's one of the reasons it surprises me when these services publish fraud signals (as in the subject post), even if it's only a few simple ones. You will find that one of your most effective weapons against fraudsters is secrecy and misdirection. Never "train" them by letting on how you are trying to detect them and never give immediate feedback when you bust them. For instance (and I will give only a couple of simple examples), make them believe there is a problem with the site or that their order is on the way (just don't send it). The longer it takes for them to realize that their game isn't working, the better. And the more confused they are about whether it actually is working, the better. You want them confused, frustrated, and ready to move on to an easier target.
Again, my point is that for anyone who has a real fraud problem and needs a service like this, devising an *effective" strategy for using it is not simple. Instead of simplicity, Sift should advertise that fraud is tough, but they are tougher. They should talk up their models, machine learning, etc., but also their experience in and commitment to understanding customers and leading them to an effective use of their tools and mitigation strategies. This stuff isn't turnkey. It is very much a customer service and consulting business, whether they would prefer it to be pure self-serve SaaS or not.
In any event, good luck.
Well, that's kind of the point: sure, you can get it to function easily, but getting it to actually provide you with solid coverage takes some effort.
>Disclosure: I'm a happy current customer whose model isn't quite right yet, but is getting there.
That pretty much seals my point.
I mean, who wants semi-effective fraud protection? That's where the extra work comes in. And what you will find is that fraudsters change their game and you have to evolve your model, even once you get it customized. So, if you are simply using the "easy-to-use" out of the box signals, you are going to find that fraudsters have already been trained by other Sift customers. Sure, it might take them a moment to realize that you have implemented these measures, but they will adapt quickly when they do. If not, then you are dealing with amateurs and you really don't have much of a fraud problem.
And, that's one of the reasons it surprises me when these services publish fraud signals (as in the subject post), even if it's only a few simple ones. You will find that one of your most effective weapons against fraudsters is secrecy and misdirection. Never "train" them by letting on how you are trying to detect them and never give immediate feedback when you bust them. For instance (and I will give only a couple of simple examples), make them believe there is a problem with the site or that their order is on the way (just don't send it). The longer it takes for them to realize that their game isn't working, the better. And the more confused they are about whether it actually is working, the better. You want them confused, frustrated, and ready to move on to an easier target.
Again, my point is that for anyone who has a real fraud problem and needs a service like this, devising an *effective" strategy for using it is not simple. Instead of simplicity, Sift should advertise that fraud is tough, but they are tougher. They should talk up their models, machine learning, etc., but also their experience in and commitment to understanding customers and leading them to an effective use of their tools and mitigation strategies. This stuff isn't turnkey. It is very much a customer service and consulting business, whether they would prefer it to be pure self-serve SaaS or not.
In any event, good luck.
> what you will find is that fraudsters change their game and you have to evolve your model, even once you get it customized.
That is exactly why I want a machine learning model, so that it can constantly be modifying and improving itself. Even better that it has a network of many clients, so that it can see lots of fraud examples and IP addresses where fraud is happening in real time.
> you are going to find that fraudsters have already been trained by other Sift customers.
The opposite should be true: fraud attempts on other Sift customers have helped improve your fraud protection.
> who wants semi-effective fraud protection?
If I can get immediate marginal improvement inexpensively with a probability that the fraud protection becomes very effective, then yeah...sounds good. Sign me up.
That is exactly why I want a machine learning model, so that it can constantly be modifying and improving itself. Even better that it has a network of many clients, so that it can see lots of fraud examples and IP addresses where fraud is happening in real time.
> you are going to find that fraudsters have already been trained by other Sift customers.
The opposite should be true: fraud attempts on other Sift customers have helped improve your fraud protection.
> who wants semi-effective fraud protection?
If I can get immediate marginal improvement inexpensively with a probability that the fraud protection becomes very effective, then yeah...sounds good. Sign me up.
With domain name registration the factors that we have noticed that are almost certainly fraud orders are (in various combinations):
1) credit card payment is all lower case and/or obvious non understanding of how US addresses are formatted
2) domain name has "hack" or some foreign sounding word. Or is anything related to vietnam (get plenty from vietnam)
3) IP location doesn't match customers location
4) Multiple attempts in a row with different credit cards
5) Registrant name doesn't match the name on the credit card and/or address
6) Customer name doesn't relate to email address used in any way.
Once again no one factor is definitive usually but a combination of several together almost always indicate a fraud order.
Those are off the top there are more. Bottom line is when you simply look visually at the orders you can tell with near 100% certainty that an order is fraudulent.
Otoh, here is a fictional example of an order that wouldn't appear fraudulent at all:
domain: bobspartycity.com
Registrant: Bob Wagner Address: 76 Walnut St., Williamette IL [email protected] And IP is in that vicinity etc.
...etc. It could be of course but we've never had a case where a fraudster puts much effort into faking an order using knowledge of what we look for.
1) credit card payment is all lower case and/or obvious non understanding of how US addresses are formatted
2) domain name has "hack" or some foreign sounding word. Or is anything related to vietnam (get plenty from vietnam)
3) IP location doesn't match customers location
4) Multiple attempts in a row with different credit cards
5) Registrant name doesn't match the name on the credit card and/or address
6) Customer name doesn't relate to email address used in any way.
Once again no one factor is definitive usually but a combination of several together almost always indicate a fraud order.
Those are off the top there are more. Bottom line is when you simply look visually at the orders you can tell with near 100% certainty that an order is fraudulent.
Otoh, here is a fictional example of an order that wouldn't appear fraudulent at all:
domain: bobspartycity.com
Registrant: Bob Wagner Address: 76 Walnut St., Williamette IL [email protected] And IP is in that vicinity etc.
...etc. It could be of course but we've never had a case where a fraudster puts much effort into faking an order using knowledge of what we look for.
Agree with Vietnam. As a % of fraud I'd say 70% is Vietnam. And few if any legitimate orders from Vietnam. In our experience.
Do you have this kind of ruleset codified or automated in any way? It looks like a great example of knowing your own market.
No it's not automated. Although it would be possible to mechanical turk if you train in the rules.
One thing I forgot to mention is that in edge cases what you do is send off an auto email to the registrant with a challenge question or requesting additional information from them. Or you say that the bank has declined the charge.
So for example if the person says they are "Bob Wagner" a business owner living in the US but replies to the email using broken english or doesn't understand the question you know something is wrong. Of course you try and do it in a way that doesn't tip them off that you suspect anything (other than in the case where you tell them that the credit card has been declined (when it hasn't) to see how they respond). Someone who has committed fraud will of course be paranoid. In fact a kid using his mothers credit card will be paranoid and respond in a unique way that says "shit" in some way shape or form.
Edit: Auto email must be formatted and appear to have been personally written by a person just to the recipient. Otherwise many cases won't be responded to at all.
One thing I forgot to mention is that in edge cases what you do is send off an auto email to the registrant with a challenge question or requesting additional information from them. Or you say that the bank has declined the charge.
So for example if the person says they are "Bob Wagner" a business owner living in the US but replies to the email using broken english or doesn't understand the question you know something is wrong. Of course you try and do it in a way that doesn't tip them off that you suspect anything (other than in the case where you tell them that the credit card has been declined (when it hasn't) to see how they respond). Someone who has committed fraud will of course be paranoid. In fact a kid using his mothers credit card will be paranoid and respond in a unique way that says "shit" in some way shape or form.
Edit: Auto email must be formatted and appear to have been personally written by a person just to the recipient. Otherwise many cases won't be responded to at all.
Hushmail: In my experience, if the email address is [email protected], it's a fraudster 100% of the time.
10minutemail.com and yopmail.com too! :-)
The patterns that emerge for fraudulent orders are amazing. And, as the article notes, often specific to a particular merchant. Fraudulent orders often come in waves lasting up to several months, and pattern recognition can be particularly helpful in identifying parts of those longer waves.
I'm also working on a project in this space - http://www.merchantprotector.net
I'm also working on a project in this space - http://www.merchantprotector.net
(I upvoted you but I might have clicked the downarrow instead).
On your site this phrase:
"We used to have a problem with fraudulent orders."
My suggestion is that the following is much clearer:
"We no longer have a problem with fraudulent orders."
Also the other info I would present in a less negative way. People tend to respond better (imho and experience) to a message presented in a positive way rather than negatively.
So for example rather than:
"Fraud can kill your business" I would say:
"Reduce Fraud and Make more money".
Instead of "Stop the cycle of worry" I would say:
"Sleep at night and make more money"
etc.
Obv. there are many twists to this that's just two examples.
On your site this phrase:
"We used to have a problem with fraudulent orders."
My suggestion is that the following is much clearer:
"We no longer have a problem with fraudulent orders."
Also the other info I would present in a less negative way. People tend to respond better (imho and experience) to a message presented in a positive way rather than negatively.
So for example rather than:
"Fraud can kill your business" I would say:
"Reduce Fraud and Make more money".
Instead of "Stop the cycle of worry" I would say:
"Sleep at night and make more money"
etc.
Obv. there are many twists to this that's just two examples.
Thanks for the feedback! Those are great ideas, I will test them out!
We've found that good indicators include: a large distance between billing and shipping addresses, a large distance between estimated IP location and billing address, large order size, using a free email like gmail/yahoo/hotmail (that's the smallest of the factors, but virtually all of our fraud orders use them). Even combining these and others with a threshold, it's still hard to reliably detect without too many false positives.
At a previous gig we found the same basic factors. I wrote a quick script to iterate through all the available Weka[1] classifiers using our manually flagged data as a training set. Then I took the top 20 performing ones and used them on incoming orders in production. If more then half the classifiers agreed a transaction was fraud, we denied it. Though this seems a very blunt hammer (I'm not a machine learning expert by any stretch) it worked remarkably well.
[1] http://www.cs.waikato.ac.nz/ml/weka/
[1] http://www.cs.waikato.ac.nz/ml/weka/
What happens to the false positives - ie the ones you denied that were real? Do people get a way to prove they really are legit?
(Matt Cutt's blog claimed I was a spammer when I made a comment about two factor authentication - I have no idea why. It told me to email the administrator to get it accepted but of course provided no clues on what address to use etc so I didn't bother. I'm betting the anti-spam software is claiming a victory when it was actually a failure.)
(Matt Cutt's blog claimed I was a spammer when I made a comment about two factor authentication - I have no idea why. It told me to email the administrator to get it accepted but of course provided no clues on what address to use etc so I didn't bother. I'm betting the anti-spam software is claiming a victory when it was actually a failure.)
For any denial we popped a "technical issues, please call customer support" message. We found that fraudsters were far less likely to call then real customers.
Fraud "facts" like that applied in a blanket fashion would frequently flag international customers, 3 of the 5 rules listed apply to me.
A US company selling internationally will have to be very careful, especially at first, applying any "fraud best practices."
One of the (apparent) advantage to the OP's service is that there is a built-in learning component, presumably tailored to your particular store. That should help quite a lot with recognizing patterns unique to an individual situation.
One of the (apparent) advantage to the OP's service is that there is a built-in learning component, presumably tailored to your particular store. That should help quite a lot with recognizing patterns unique to an individual situation.
Are you familiar with ensemble methods and boosting algorithms?
How does Sift Science combine multiple signals like these (which individually are pretty weak) into one fraud detection system with a high level of predictive accuracy?
How does Sift Science combine multiple signals like these (which individually are pretty weak) into one fraud detection system with a high level of predictive accuracy?
We really wanted to like Sift, as we suffer from a substantial amount of fraud attempts (as most business who sell digital products). However, their model is not a good fit for eCommerce sites, and that's a shame - it seems to be built specifically for services marketplaces like AirBnb, where there is typically a time delay between payment and service provision.
I exchanged a few Email with their support previously, and there is no way to get real time fraud scoring. I would expect to receive in the response to a transaction event the risk score associated with it - something similar to what the Minfraud service does (which we use), but taking into account more factors since they collect more data via their Javascript API.
One can only hope they'll offer this capability in the future, and we'd be glad to try it out again.
I exchanged a few Email with their support previously, and there is no way to get real time fraud scoring. I would expect to receive in the response to a transaction event the risk score associated with it - something similar to what the Minfraud service does (which we use), but taking into account more factors since they collect more data via their Javascript API.
One can only hope they'll offer this capability in the future, and we'd be glad to try it out again.
Hey Eran! We actually help quite a few e-commerce sites today. To be clear, 99% of our scores are updated and ready within a second of receiving the transaction. The fact that the response to a transaction does not include this score is a design flaw, and you're not the first to bring it up. We're working on a fully revamped API that should address this, and the revamped API will be launched in the next 2-3 months.
I do remember your emails from back in the day -- will followup offline.
I do remember your emails from back in the day -- will followup offline.
Hi Jason,
Yes - that was the same claim I got from your support back then (score update within a second mostly) - but that breaks up a linear process into an asynchronous one (since you can't complete the purchase process immediately). Hope that makes sense - we will be looking forward to your API update.
Yes - that was the same claim I got from your support back then (score update within a second mostly) - but that breaks up a linear process into an asynchronous one (since you can't complete the purchase process immediately). Hope that makes sense - we will be looking forward to your API update.
E-commerce that deals with physical goods does have a time delay, so I think you're really trying to say that it's not a good fit for digital goods sales rather than that it's not a good fit for e-commerce...
This doesn't seem really smart since some of those apply to me. I am not from America and I sometimes order from there meaning it would seem like I'm ordering at 4am when its 2pm my time. Another thing is the email, I know a lot of people with their birth year inside their email. What about them?
Sift intern here.
We use the time zone of the customer rather than of the website for scoring riskiness. Sorry if that wasn't clear.
As for customers including birth years in their emails, you're correct that this would be counter to the general trend. However, a fraud detection system like Sift has many data points on which to score a customer. Hopefully, the person would otherwise look benign and thus not have a very high overall score.
We use the time zone of the customer rather than of the website for scoring riskiness. Sorry if that wasn't clear.
As for customers including birth years in their emails, you're correct that this would be counter to the general trend. However, a fraud detection system like Sift has many data points on which to score a customer. Hopefully, the person would otherwise look benign and thus not have a very high overall score.
Thanks for clarifying that. I though it used the server time.
"It might turn out that size 10 shoes are more fraudulent than size 15 shoes."
There should be a pretty limited list of mailing addresses that would order size 15 shoes (Shaq's house?) and the black-market for reselling them must be a lot tighter.
There should be a pretty limited list of mailing addresses that would order size 15 shoes (Shaq's house?) and the black-market for reselling them must be a lot tighter.
So for something like "fraudsters don't use capital letters," does your system discover a fact like this automatically, or do I have to think up these indicators myself and hope they are relevant?
This was discovered automatically. That's one benefit of a machine learning-based system: you feed it a lot of data and tell it when fraud actually occurred and it adapts its rules predict fraud accordingly.
There is no relevancy here to be had, these are simply trivially interesting trends.
Did anyone else find it odd that the 2am and 4am times weren't qualified? (US? East Coast/West Coast?)
I think they're actually quite sophisticated about it - they look at the time zone according to the shopper's browser. So they're looking at local-to-the-shopper.
Yep! We try to localize the time.
So the fraudsters in Vietnam are trying to time it for 2am to 4am in the local market of the site they're defrauding?
Not quite...an order is more likely fraudulent when it was placed at 2-4am local time. Local time for the fraudster (in Vietnam or wherever else), not local time for the site they're defrauding.
What exactly is a "fraudulent order" anyway? Someone has a credit card, pays you and you send the product. Where is the fraud? Isn't it external to the company or service? If someone steals a credit card and makes an online purchase, isn't that the responsibility of the card company in securing its account proxy more fully?
E-commerce fraud takes many forms. Three main types of fraud impact merchants: payment fraud, new account fraud and account takeover. We described all three recently at Sift in a blog post: http://ow.ly/oPrS2.
In the stolen card scenario you describe (a type of payment fraud), the e-commerce merchant is actually liable. In other words, if that card is reported stolen by the cardholder after the goods are shipped out, the merchant loses the revenue. This is b/c for online credit card transactions, they are categorized as "card not present" transactions, since the merchant can't be as certain the actual cardholder made the purchase. If the transaction had occurred in a physical store, the card company would be liable.
In the stolen card scenario you describe (a type of payment fraud), the e-commerce merchant is actually liable. In other words, if that card is reported stolen by the cardholder after the goods are shipped out, the merchant loses the revenue. This is b/c for online credit card transactions, they are categorized as "card not present" transactions, since the merchant can't be as certain the actual cardholder made the purchase. If the transaction had occurred in a physical store, the card company would be liable.
Their home page says, "Get going in minutes: Integrate in just three steps: paste a Javascript snippet onto your site, log transactions from your servers to our REST API, and send examples of banned users." But it isn't so. They require a long-term in-depth model of your site/usage including multiple fraudulent examples (what if you've mostly solved fraud?) before returning meaningful results.
According to this post and previous posts, they should be able to return meaningful results with very basic things: time of transaction, email address, etc. They should start off with: "Here's our recommendation, but it's based on limited information so we feel X strongly about it." Instead they say: "Give us more information, we can't help you yet."