The pros and cons of “big data” lending decisions

lending.jpg

Just as borrowing options are no longer limited to the traditional bank, increasingly new types of lenders are diverging from the trusted credit score system in order to flesh out their customer profiles and assess risk in new ways. This means going beyond credit/payment relevant data and looking at additional factors that could include educational merits and certifications, employment history, which websites you visit, your location, messaging habits, and even when you go to sleep.

Undoubtedly, this is the sort of thing that strikes panic into the hearts of many of us. How much is a creditworthy amount of sleep? Which websites should I avoid? Will they hold the fact I flunked a math class against me? Nevertheless, proponents of “big data” (it’s really just data…) risk assessment claim that this approach works in favor of those who might be suffering from the effects of a low credit score.

Let’s take a look…

Pros

The fact is, credit scores don’t work for everyone and they can be difficult to improve depending upon your position. Some folks, through no fault of their own, end up getting the raw end of the deal (perhaps they’re young, a migrant, or they’ve just had a few knockbacks in life). Now given these newer models can take extra factors into account –  including how long you spend time reading contracts, considering application questions, and looking at pricing options – this additional information can add a further dimension to an application, which in turn may prompt a positive lending decision.

A recent article looked at the approach of Avant, a Chicago-based start-up lender, which uses data analytics and machine learning to “streamline borrowing for applicants whose credit scores fall below the acceptable threshold of traditional banks”. They do this by crunching an enormous 10,000+ data points to evaluate applicants. There isn’t much detail in terms of what these data points are, but doubtless they will draw upon the reams of publicly available information generated by our online and offline “emissions” – webpages browsed, where we shop, our various providers, social media profiles, friend groups, the cars we drive, our zip codes, etc etc etc. This allows the lender to spot patterns not “visible” to older systems – for example, where a potential customer has similar habits to those with high credit scores, but has a FICO score of 650 or below.

The outcome – if all goes well – is that people are judged on factors beyond their credit habits, and for some individuals this will open-up lending opportunities where they had previously received flat “nos”. Great news!

This technology is being made available to banks, or anyone who wants to lend. They may even eventually outmode credit scores, which were an attempt to model credit worthiness in a way that avoided discrimination and the unreliability of a bank manager’s intuition…

So, what are the downsides?

Cons

There are a number of valid concerns about this approach. The first of which regards what data they are taking, and what they are taking it to mean. No algorithm, however fancy, can use data points to understand all the complexities of the world. Nor can it know exactly who each applicant is as an individual. Where I went to school, where I worked, whether I’ve done time, how many children I have, what zip code I live in – they are all being used as mere proxies for certain behaviors I may or may not have. In this case they are being used as proxies for whether or not I am a credit risk.

Why is this an issue? Well, critics of this kind of e-scoring, like Cathy O’Neill, author of Weapons of Math Destruction, argue that this marks a regression back to the days of the high street bank manager. In other words, instead of being evaluated as an individual (as with a FICO score which predominantly looks your personal debt and bill paying records), you are being lumped in a bucket with “people like you”, before it is decided whether such people can be trusted to pay money back.

As O’Neill eloquently points out, the question becomes less about how you have behaved in the past, and about how people like you have behaved in the past. Though proxies can be very reliable (after all, those who live in rich areas are likely to be less of a credit risk than those who live in poor neighborhoods), the trouble with this system is that when someone is unfairly rejected based on a series of extraneous factors, there is no feedback loop to help the model self-correct. Unlike FICO, you can’t redeem yourself and improve your score. So long as the model performs to its specification and helps the lender turn a profit, it doesn’t come to know or care about the individuals who are mistakenly rejected along the way.

There is an important secondary problem with leveraging various data sources to make predictions about the future. There is no way of knowing in every case how this data was collected. By this I mean to say, there is no way of knowing whether the data itself is already infused with bias, which consequently biases the predictions of the model. Much has been made of this issue within the domain of predictive policing, whereby a neighborhood which has been over zealously policed in the past is likely to have a high number of arrest records, which tells an unthinking algorithm to over-police it in the future, and so the cycle repeats… If poor data is being used to make lending decisions, this could have the after effect of entrenching poverty, propagating discrimination, and actively work against certain populations.

Lastly (and I’m not pretending these lists of pros and cons are exhaustive), there is a problem when it comes to the so-called “chilling effect”. If I do not know how I am being surveyed and graded, this might lead me to behave in unusual and overcautious ways. You can interrogate your FICO report if you want to, but these newer scoring systems use a multitude of other unknown sources to understand you. If you continue to get rejected, this might result in you changing certain aspects of your lifestyle to win favor. Might this culminate in people moving to different zip codes? Avoiding certain – perfectly benign – websites? Etcetera, etcetera. This could lead to the unhealthy manipulation of people desperate for funds…

So, is this new way of calculating lending risk a step forward or a relapse into the bad practices of the past? Well having worked for the banking sector in years gone by, one thing still sticks in my mind when discussions turn to lending obstructions: lenders want to lend. It’s a fairly important part of their business model when it comes to making a profit (!). At face value, these newer disrupters are trying to use big data analytics to do exactly that. In a market dominated by the banks, they’re using new and dynamic ways to seek out fresh prospects who have been overlooked by the traditional model. It makes sense for everyone.

However, there is clearly the need for a cautionary note though. Although this method undoubtedly praiseworthy (and canny!) we should also remember that such tactic can breed discrimination regardless of intentions. This means that there needs to be some kind of built-in corrective feedback loop which detects mistakes and poorly reasoned rejections. Otherwise, we still have a system that continually lends to the “same type of people”, even if it broadens out who those people might be. The bank manager returns.

Having a fair and corrigible process also means that lenders need to be more open about the data metrics they are using. The world – and particularly this sector – has been on a steady track towards more transparency, not less. This is difficult for multiple reasons (which warrant another discussion entirely!) but as important as it is to protect commercial sensitivity and prevent tactics like system gaming, it is also critical that applicants can have some idea with regards to what reasonable steps they can take to improve their creditworthiness if there are factors at play beyond their credit activity.

What if Twitter could help predict a death?

I want to use this blog to look at how data and emerging technologies affect us – or more precisely YOU. As a tech ethics researcher, I’m perpetually reading articles and reports that detail the multitude of ways in which data can be used to anticipate bad societal outcomes: criminality, abuse, corruption, disease, mental health, etc etc. Some of these get oxygen, some of them don’t. Some of them have integrity, some don’t. Often these tests, analyses, and studies identify problems that gesture toward ethically “interesting” solutions.

Just today this article caught my attention. It details a Canadian study that tries to get to grips with an endemic problem: suicide in young people. Just north of the border, suicide causes no fewer than 24% of deaths amongst those aged between 15 and 24 (Canadian Mental Health Association). Clearly, this is not a trivial issue.

In response, a group of researchers have tried to determine the signs of self-harm and suicide by studying the social media posts of those in the most vulnerable age bracket. The team – from SAS Canada – have even speculated that, “these new sources could provide early indication of possible trends to guide more formal surveillance activities.” So, with the prospect of officialdom being dangled before us, it’s important to ask how this social media analysis works. In short, might any one of us land-up being surveilled as a suicide risk if we happen to make a trigger comment or two on Twitter?

Well the answer seems to be “possibly”. This work harvested 2.3 million tweets, of which 1.1 million were identified as “likely to have been authored by 13 to 17-year-olds in Canada”. This determination was made by a machine learning model that has been trained to predict age by relying on the way young people use language. So, if the algorithm thinks you tweet like a teenager, you’re potentially on the hook. From there, the team looked for where these tweets related to depression and suicide, and “picked some specific buzzwords and created topics around them, and our software mined those tweets to collect the people.”social media

Putting aside the undoubtedly harrowing idea of people collection, it’s important to highlight the usefulness of this survey. The data scientists involved insist that the data they’ve collected can help them narrow down the Canadian regions which have a problem (although one might contest that the suicide statistics themselves should reveal this), and/or identify a particular school or a time of year in which the tell-tale signs are more widespread or stronger. This in turn can help better target campaigns and resources, which – of course – is laudable, particularly if it is an improvement on existing suicide statistics. It only starts to get ethically icky once we consider what further steps might be taken.

The technicians on the project speculate as to how this data might be used in the future. Remember, we are not dealing with anonymized surveys here, but real teen voices “out in the wild”: “He (data expert Jos Polfliet) envisions the solution being used to find not only at-risk teens, but others too, like first responders and veterans who may be considering suicide.”

Eh? Find them? Does that mean it might be used to actually locate real people based on what they’ve tweeted on their personal time? As with many well-meaning data projects, everything suddenly begins to feel a little Minority Report at this point. Although this study is quite obviously well-intentioned, we are fooling ourselves if we don’t acknowledge the levels of imprecision we’re dealing with here.

Firstly, without revealing the actual identities of every account holder picked-out by the machine learning, we have no way of knowing the levels of accuracy these researchers have hit upon when it comes to monitoring 13-17 year-olds. Although the use of certain language and terminologies might be a good proxy for the age of the user, it certainly isn’t an infallible one in the wacky world of the internet.

Secondly, the same is true of suicide and depression-related buzzwords. Using a word or phrase typically associated with teen suicide is not a sufficient condition for a propensity towards suicide (indeed, it is unlikely to even be a necessary condition). As Seth Stephens-Davidowitz discussed in his new book Everybody Lies: Big Data, New Data, And What the Internet Can Tell Us About Who We Really Are, in 2014 research found that there were 6,000 Google searches for the exact “how to kill your girlfriend” and yet there were “only” 400 murders of girlfriends. In other words, not everyone who vents on the internet is in earnest, and many who are earnest in their intentions may not surface on the internet at all. So, in short, we don’t know exactly what we’ve got when we look at these tweets.

Lastly, without having read the full methodology, it appears that these suicide buzzwords were hand-picked by the team. In other words, they were selected by human beings, presumably based on what sorts of things they deemed suicidal teens might tweet. Fair enough, but not particularly scientific. In fact, this sort of process can be riddled with guesswork and human bias. How could you possibly know with any certainty, even if instructed by a physician or psychiatrist, exactly which kinds of words of phrases denote true intention and which denote teenage angst?

Hang on a second – you might protest – these buzzwords may have been chosen by a very clever, objective algorithm? Yet, even if a clever algorithm could somehow ascertain the difference between a “I hate my life” tweeted by a genuinely suicidal teen and a “I hate my life” tweeted by a tired and hormonal teenager (perhaps based on whatever language it was couched in), to make this call it would have to have been trained on data which used the tweets of teens who have either a) committed suicide or b) have been diagnosed/treated for depression. To harvest such tweets, the data would have to rely upon more than Twitter alone… all information would have to be cross-referenced with other databases (like medical records) in ways that would undoubtedly de-anonymize.

So, with no guarantees of accuracy, the prospect of physical intervention by social services or similar feels like a scary one – as is the idea of ending up on a watchlist because of a bad day at school. Particularly when we don’t know how this data would be propagated forward…

Critically, I am not trying to say that the project isn’t useful, and SAS Canada are forthcoming in their acknowledgment that ethical conversations that need to take place. Nevertheless, this feels like the usual ethical caveat which acts as a disclaimer on work that has already taken place and – one might reasonably assume – is already informing actions, policies, and future projects.

Some of the correlations this work has unveiled clearly have some value, for example, there is a 39% overlap between conversations about suicide and conversations about bullying. This is a broad trend and a helpful addition to an important narrative. Where it becomes unhelpful, however, is when it enables and/or legitimizes the external surveillance of all bullying-related conversations on social media and – to carry that thought forward – some kind of ominous, state sanctioned “follow-up” for selected individuals…