Fintech is not a natural fit for a traditional generalist approach to venture capital investments. Clearly there are many funds and investors doing quite well in fintech, but I would argue that specialisation matters. Fintech is a different beast, and it requires specialised managers and educated patient investors.
Why? Time to market is longer and more costly than consumer internet. Building a digital product that touches peoples’ financial lives is hard enough. But obtaining licensing, drafting policy and procedures, and developing regulatory infrastructure and operations, takes time and substantial investment. Solving for engineering and product variables is just one part of the equation. In fintech, demonstrating market traction, or delivering an initial minimum viable product takes longer, generally costs more and may not fit the traditional funding approach and timeframe.
Customer acquisition is costly. The big exception might be in payments – for example, digitial wallet company Venmo is the definition of viral. For the rest of fintech, products are generally targeted at unseating an incumbent, and those customer relationships are very sticky. For a consumer, redirecting automatic bill pay, or payroll deposits, adds just enough friction in the process that conversion gets tight. For small businesses the stickiness is even higher. The customer base is heterogeneous, and acquisition channels are less mature than in the consumer arena. Driving trial is a lot harder, and requires a well structured strategy, experienced team and deeper pockets.
Exits are less certain, more complex. The obvious buyers for a fintech startup should be banks or other large financial institutions. However, that is only practical and attractive if the startup has invested in building its regulatory infrastructure correctly. Regulatory asymmetry – which may be why it is easier to innovate outside of banks – can also impact exit prospects. When acquired by a bank, a fintech startup’s regulatory oversight changes overnight. If the startup has not built robust bank-level regulatory infrastructure, then it can become unacquirable or require substantial retrofitting under its new bank parent. This impacts valuation, and potential buyers are forced to consider additional investments and liabilities required after the close.
Banks may shy away from providing exits to platform and aggregation businesses. Any fintech business that aggregates financial products, or relies heavily on banks as a customer or supplier can become a tricky acquisition for another bank. Bankrate’s acquisition by Red Ventures is a perfect example. A bank acquiring Bankrate would not have made any sense. In this hypothetical, the new owner is likely to have destroyed value by driving away many of Bankrate’s affiliates – other banks – which would no longer want to do business with a company owned by a direct competitor.
This same dynamic could play out for businesses like credit bureaux, payment service providers, loan marketplaces and distributed ledger and blockchain applications. A lot of the value in these platform business models resides in their ubiquity and neutrality. There is a reason that when Visa grew and scaled it was owned by a consortium of banks, not just one.
Similar arguments could be made for venture investing in other regulated industries, but for now I will stick to what I know. The good news is that fintech is huge, and specialisation among investors has already started. This bodes well for the industry as a whole, and most importantly it bodes well for the customer.
I now want to go a bit deeper, and show how I think this relates to different verticals within financial services. Here is how I see this applying to lending, for instance.
The promise of having startups and technology players “hack” lending is immense. The premise is simple. The availability of new technology and data should transform lending by:
- Cost-effectively creating relevant and tailored customer experiences and products.
- Enhancing lenders’ ability to identify and reach the right customers through new channels.
- Delivering superior pricing – and by proxy enhancing financial inclusion – through the combination of huge datasets and increased processing power and machine learning.
But when I look at the online alternative consumer lending market in the US, this promise – long in the making – is incomplete. Building relevant,vertical specific and tailored customer experiences has worked, has driven down acquisition costs and has improved customer satisfaction. However, when it comes to risk modelling and pricing, the online lending revolution has come up short.
I believe there is a structural reason for this, and I truly believe than banks – incumbents – are best positioned to deliver this promise, and to find ways to bring millions into the financial mainstream.
About data
Grabbing a bunch of new data and slamming it through your analytical tools of choice to determine if something has predictive value is not trivial. But this assumes you have come by the data in the first place, and amassing large data sets is also not trivial.
So if you have lots good data, and good analytical chops, you should be good to go, right? Well, not all data can be used, no matter how predictive it may be. This is a good thing from a societal standpoint to prevent discriminatory practices. However, from a purely intellectual standpoint it reduces the predictive potential of future models by excluding certain variables from the start. We do have a lot of new data, but some of it cannot – and should not – be used.
You could imagine a world where someone’s cell phone GPS data could be used for underwriting. For instance, a consumer displays a steady pattern of getting up in the morning and going to work several miles away without ever missing a day. However, if that location data is for a poor largely minority neighbourhood this could be the basis for the wrong kind of discrimination and would justifiably require additional scrutiny as the basis for decision-making
The impact this has on financial inclusion is uncertain. On one end, it protects consumers from abusive products and practices. However, it may also limit the reach of new lending models to serve the underbanked. Traditionally, risk models do not know how to price subprime loans properly, and the outcome is that for many, consumer choice is constrained. This creates a void, drives up prices and can create other types of abusive behaviour on the servicing and collections side. This is to say that data by itself is not enough.
Running live experiments
The second structural impediment has to do with how expensive it is to determine effectively the predictive value of new data. Like any experiment, you hypothesise about what variables may have predictive power, you run tests and then you analyse whether or not the variables taught you anything new about the outcomes.
In lending what does this mean? It means that you have to lend a lot of money to a lot of people, and hope a good number of these loans default. If no one defaults on the loans, then you get your money back but you have learned nothing. Without a lot of defaults you cannot possibly estimate whether or not one of you new variables might have predicted it.
If you already have a large amount of loan data you could mitigate your losses by backtesting. You could append your new variables to the existing loans and backtest to see if a new risk model would have had more predictive power. This, however, presupposes that you have a large dataset of loans already issued and repaid or not. It also presupposes that you can somehow append the new data – advantage incumbents.
To be successful at either approach requires:
1 High levels of conviction – a strong hypothesis that the data you laboured to collect may actually be legal and predictive.
2 Real analytical chops – you have the capabilities, expertise and technology to find a signal in the data.
3 Existing loan data – you have something to backtest against.
4 Deep pockets – you have enough dry powder to run live experiments and sustain losses.
5 Patience – you have enough time, and money, to watch your loans default, and sufficient wisdom to learn from it.
Items 1 and 2 may explain why we are not seeing traditional incumbents tackling this. Items 3, 4 and 5 may explain why we are not seeing VC funded startups winning in this space.
What is inside the box?
The third structural impediment has to do with transparency. Machine learning technology has the promise to deliver enhanced predictive value to lending models that should easily surpass that of traditional logistical regression models. However, these new models and the underlying techniques must be explainable to regulators. If machine learning delivers highly predictive insights but regulators cannot understand what is happening inside the “black box”, we will be at a standstill.
This is a solvable problem but it will take time. Technologists and data scientists must strive to make models that are easily understood and accessible. In parallel, regulators must evolve and build their own teams of data scientists who can scrutinise and manage the risk stemming from machine learning-based underwriting.
Here incumbents may have the slightest edge, only in that they have built an expertise out of explaining things to regulators.
I am convinced that, at its core, finance has unlimited potential to be a transformative force for good. On a daily basis, this is in large part what gets me out of bed. When it comes to lending, banks have a unique opportunity and responsibility to transform the way risk is assessed for the underbanked. The impact of new risk models for consumer lending alone could bring millions into the financial mainstream, and have a huge impact on social mobility globally. Let’s get to work.