As a FinTech founder and mathematician, I should probably favour algorithms over human assessment. I often do. However I am getting tired of hearing about yet another algorithm that is able to predict, forecast and act on data to make financial decisions. Ultimately we’ve seen it before. Will the algorithms and robots win or does human analysis shine through for another business cycle?
This story starts with a cornerstone of basic AI and automated credit decisioning. Algorithms are built that take a lot of consumer data and perform analysis called logistic regression to develop a “good-fit” model. The observed data is then used to produce an equation that prescribes an expected behaviour. For lending, this is a probability of default forecast.
My first skepticism is due to my mathematics background. I loved engineering mathematics and fluid dynamics, when pure mathematics is used to derive equations that are fundamental to the development of aircraft, cars and all kinds of other high technology equipment we have today.
This mathematics saved time: a single equation developed through algebra meant we had ways to describe what would happen. The human mind equipped with paper was able to forecast without calculation.
As time went on, computational ability improved. In non-linear dynamics, we became interested into when and where systems went from calm to chaos. In effect, we wanted to know when our derived equations became unstable. Computational methods appeared that were able to map out these cases and build up smart eco-systems which then formed a more accurate forecast of reality than the human generated equations.
In the 1970s we started applying the initial human derived models without computational power to start predicting really annoying things like traffic jams. Over time, computers were introduced and we learnt that the computational part was much more influential that the initial models we derived: the human equations mattered less than the rapid ability for a computer to estimate all the potential ways the system would break down into a traffic jam – and then alleviate those risks. My last piece of research on this was in 2005 and computer power has move on exponentially.
Then in 2006 I did what all reasonably good engineers did: I went to work for Wall Street. Same equations and a real application of the mathematics (I failed to mention that the traffic jam research was mostly focused on driverless cars, something that will take another decade to become mainstream). Correlations, causations and logistic regressions where everywhere. In fact there was an attractive mix of human-built equations and computer generated ones. But the data source was bad: a long stretch without a recession and notably no significant house price corrections meant the instability was never tested. We know what happened next.
Moving forward to today, we see far less human developed equations. In any case, they have become far too complicated for many and I believe the variation of data in the world is significantly more important than the fine-tuning of the human equations, just like the traffic jam models.
But my concern now is about asking if these models are correct this time round? Can we trust the goodness of fit? Are sufficiently good stressed data sets used? Are people caring about the what-ifs? In my mind, the opportunity to beat the algorithms and robots is just as relevant as ever:
- Australian banks view investor loans as less risk than owner-occupier loan. The data they use shows this. International evidence suggests this is not the case.
- Data shows that SME lending is not worthwhile (and many banks have pulled back from this). But have they missed out on how new relationships can be formed or more suitable products can be developed.
- Banks continue to adore their credit card businesses yet we all believe this rewards-based value capture will end at some point.
At the same time, we must also be aware that there will be human interference. Investor loans benefit from a tax break, SME lending will always be politically important and credit card use has been sticky.
The key to all this is not only capturing vast amounts of data but then finding ways that it can be effectively used to test your human or computer-generated equations but also how an artificially intelligent machine can adapt its equations when there will be known human interference, such as political and legislative changes.
This is how we can beat the existing robots.