The rise of the machines v the human touch 31/01/2018
Amid current fascination with driverless cars, it’s worth considering how far we are from a driverless credit process, as it were – what do algorithms and AI have to offer, and what still requires human consideration?
Algorithmic and manual lending processes have distinctive and often complementary characteristics. At ThinCats, we believe both are necessary to correctly evaluate a business loan.
Algorithmic lending has become a game changer. It allows a vast amount of data to be analysed, ranked and rated far more quickly than humans are capable of. The richness of such data is increasing all the time, as is its immediacy. For instance, we use (among many other inputs) ‘days beyond terms’ or DBT data, which measures the number of days beyond the contractual due date that a business pays its bills based on tradelines that have been updated in the previous three months.
We also believe we offer our investors an extra layer of support in their lending decisions, by providing quantitative analysis of not only a business’ credit worthiness, but also the strength of its security.
Algorithmic lending works particularly well if one is able to aggregate more data on individual and similar loans. This suits those making smaller loans to many borrowers. However, when it comes to larger loans, there is less granularity – fewer data points from which to extrapolate.
For example, in September we made a £6.7m loan to Chelsea Yacht & Boat Company. There are much fewer loans of this magnitude and complexity made, and doubtless a vanishingly small number with comparable characteristics to this business. This is where manual underwriting capability is indispensable. Algorithms can remove cognitive biases from the lending process; nevertheless, sometimes you need that human slant.
Algorithms have the ability to process vast amounts of data - but they can only learn from the data that they are provided with and in the manner in which they are programmed. In the absence of data, only a human can (so far) nose around the office or factory floor and get a feel for a business. Only a human can distinguish between the qualitative factors that make businesses distinct. Likewise, putting the right covenants in place for specific businesses remains down to human experience and judgement, which is why quality secured lending still needs to draw on human underwriting skills not required by unsecured lenders.
Take two businesses working in construction, for example. Assume they have indistinguishable financial metrics and similarly experienced boards of directors – except they do qualitatively different things. Business A digs the foundations, while business B installs the windows. To an algorithm, they may look the same, but to a credit analyst, there is at least one important distinction: the firm digging the holes gets paid first. That should make it lower risk in the event of, say, a downturn leading to a liquidity crisis. Likewise, the question “what’s the money for?” currently remains difficult to capture in data terms, but remains crucial in determining many loans.
We recently agreed a loan to south coast jewellers W. Bruford. In terms of judging the value of security, it is normal to estimate the value of stock to be about 30% of cost in a stressed situation. However, much of W. Bruford’s stock comes from either Rolex or Pandora, and both of these suppliers operate a buy-back at cost policy, which is difficult to capture in terms of data. Knowing that made the loan a more attractive, lower risk proposition for our credit team.
The sophistication of algorithmic lending is increasing all the time and will be further improved by the advent of open banking, which will significantly expand the data available on which to make decisions. For the foreseeable future, however, the combination of manual and algorithmic analysis, especially with larger and more sophisticated deals, provides the level of service both our lenders and borrowers require.