Tobias Berg and others have an abstract that says,
We analyze the information content of the digital footprint – information that people leave online simply by accessing or registering on a website – for predicting consumer default. Using more than 250,000 observations, we show that even simple, easily accessible variables from the digital footprint equal or exceed the information content of credit bureau (FICO) scores. Furthermore, the discriminatory power for unscorable customers is very similar to that of scorable customers. Our results have potentially wide implications for financial intermediaries’ business models, for access to credit for the unbanked, and for the behavior of consumers, firms, and regulators in the digital sphere.
This is interesting for many reasons.
Sounds gameable.
Good game. People spend time and money gaming good budgets. The defaults happen, bad game players. But we catch a stream of those defaults and estimate costs, in a good game.
Please write so people can understand what you are trying to say.
Unless you have nothing useful to say, in which case you can write as you have been writing and be ignored. (Is that really what you want?)
Yep. As soon as people find out about which trivial behaviors hurt their scores, they’ll change (or hide) them. Which is why the use of costly, hard-to-fake signals (income, existing debts, assets, history of reliable payments, etc) are necessary. They cannot really be gamed.
costly, hard-to-fake signals (income, existing debts, assets, history of reliable payments, etc) are necessary
That is what we track in the new system, they all have bearer asset versions backed by signed legal contracts. The ‘interesting’ part is that the research poses a game, easy to cheat. But you allow for technology advance in deployment, where can give our personal guarantee, digitally. We still fail, but the probability drops wau below good completions, cost easily born with mild default insurance.
Good news.
We have a path to pre-qualify traders. A little organization and we get a credit score ladder. Each trader can find a trader venue that matches risk, and knows how to move up the ladder.
All this proves is how stupendously bad a mechanism the credit bureau system is.
How is it possible that anyone can think this is a good idea? If someone told you that how often you ate at Chipotle was an even better indicator of predicting consumer default, would you be interested in that?
Edgar is correct.
But this is still interesting for many reasons.
If it determines credit dependability, it will almost certainly also affect insurance risk profiles. A LOT. The personality of the insured is one of the most important factors in property and casualty insurance rating and underwriting. (I was one of the pioneers of the use of credit reports in P&C insurance in the early 90s.) I am long since retired, but I would be amazed if actuaries and statisticians weren’t digging into this already.
Yes, Edgar’s correct. Here are some possible consequences:
(a) As Edgar suggests, people change their online habits to improve their credit.
(b) Advisory services crop up to help people better their credit via online browsing advice.
(c) Businesses appear that will browse the web for you using an avatar clone of you, thereby improving your credit worthiness with no new habits needed by you.
(d) Lenders develop counter-measures to catch those gaming the systems.
Yet another arms race…
Don’t be surprised if many of the most predictive attributes in this alternative model are deemed socially (and thus soon legally) unacceptable due to disparate impact.
Right.
If improving the accuracy of the heuristics behind any actuarial forecast yields results which go the politically incorrect way, the state will intervene (one way or another) to either prohibit the use of such information and/or impose the equivalent of “community rating” obscured cross-subsidies on the matter, which undermines the incentive to collect and analyze new sources of information or to invest in superior techniques.
Under such constraints, better predictions could only be applied to further differentiate the riskiness of individuals strictly within some identity group. But that’s doesn’t have a lot of bang for the buck.
This is one of those cases in which making little toy mathematical models really helps illustrate the point better and faster than can be done verbally. I recommend you play around with it: what is the difference in potential gain to better statistical prediction between the no-constraint and politically-constrained cases? You’ll see that under reasonable assumptions, the political-constraint takes out most of the upside.