Technical Notes: Behavioral Risk Frameworks

At Huffle we’re big fans of investigating behavioural risk models. Largely because improving the understanding behaviour is easier through Big Data gathered from social media than other types of risk.

traffic-jam

Behavioural risk can be split into separate areas: internal risk and external risk.

Internal risk is created by employees, where we start to see companies monitoring employee activity to estimate a behavioural risk that may be associated with fraud or processes that lead to poor customer outcomes.

External risk in the form of customer behaviour is also important. Marketing has always investigated behaviour of customer’s purchasing. In finance we need to look at customer behaviour over the life of transactions as there is a risk of default, churn, fraud or additional leveraging.

Many parts of behaviour can be understood through statistical modelling – and procedures to understand, segment, assign probabilities are well understood and continually investigated. However, non-linear dynamics and bifurcation theory may be a better avenue*.

Why bifurcation theory?

Bifurcation theory looks at system changes as certain assumptions or inputs change. We would look for indications that would adjust one of our expected outcomes, which may include willingness or ability to service debt or likelihood of finding and switching loan provider. The advantage over stochastic modeling is that we can use historical data but also forecast what could lead to changes beyond probability likelihoods. The obvious difficulty is developing a reliable non-linear dynamic model and testing it.

We also note that there are interactions in a system. Banks will be selling agents and create internal shifts in dynamics depending on their willingness to lower margins to gain market share. Ultimately, we would also be able to use these models to manage risk as well and measure it: understanding which reactions lead to optimal outcomes.

Starting Models: How Uber may change traffic jams

One of the easiest behavioral dynamic models is traffic. We’ve all sat on the freeway, stop-starting in a traffic jam only to eventually end up moving without any accident in sight. This is a result of human inefficiencies and behavior**. Traffic flow modeling has been investigated for decades but can be summarized in a few rules:

  1. A person wants to follow the car in front at the speed limit
  2. A person doesn’t to get too close. There is a minimum headroom between cars. If they get too close, they will hit the brakes
  3. Each person has a delay in reacting to the car in front

With these rules, we can start to build interactions between cars, with the simplest example being a single-lane freeway. Cars accelerate to meet the speed limit and brake when they get too close to the car in front.

If we build this (non-linear) system we can start to observe a few things:

  1. The system can converge to where all cars move smoothly at the speed limit
  2. If the first car stops, all cars behind it will eventually stop
  3. If the first car slows down, all cars behind it will eventually meet that speed

Perturbation Theory:

The next aspect is to add a small amount of noise. Here, noise would mean a single car suddenly brakes for a reason, perhaps the driver got too close to the car in front after looking at their Facebook feed on their smartphone.

This would lead to a sudden braking. The interesting aspect to take from here is that subsequent cars, if they were travelling at the speed limit and close enough to the car suddenly braking, will be forced to brake themselves. Furthermore, if they were at the headroom, they would encroach on the safety zone and need to brake more sharply to replenish that zone.

There is a potential knock-on impact that subsequent cars need to brake substantially more if they were all travelling close to the headroom distance behind the car in front and at the speed limit.

Without modeling this, we can understand a potential amplification of the initial noise, and subsequent cars can potentially be forced to stop. At this point, we can create the traffic jam.

What factors impact the potential for a traffic jam?

  1. Car density: more cars increase the likelihood of cars travelling closer to the headroom distance as they cram onto the road in rush hour.
  2. Speed: faster motion means the encroachment into headroom can be higher (the delay in human reactions is a fixed time but the distance travelled in that time will be higher)

There are other smaller impacts. However, the above two are controllable on a system wide basis.

What are the risks?

Car crashes are a potential risk if the stopping propagates too quickly. Time delays are also possible as cars also have a delay in returning to the speed limit after they stop, and this can lead to significantly longer delays in traffic.

How do we control them?

Reducing cars is an easy one. This can be done through toll roads or pricing strategies. Ultimately, this is one aspect that makes Uber incredibly interesting as it is adopted more widely. Rush hour will become more expensive to travel.

Speed is another. We often see temporary speed limits. This diminishes the human delay and reduced the likelihood of the traffic jam propagating. The obvious cost here is that the journey times are less than optimal and may be higher on average. But dispersion and risk is reduced.

Moving back to lending

Without detailing too much, the non-linear model becomes more complex as the forces driving it are very different in the world of finance. First of all, if we are building a default model, what defines a default? We could classify it in a similar way to a car stopping in the traffic model and with a speed limit taken as repayments on a mortgage.

Multiple agents would exist: borrowers wouldn’t react with each other, instead they would react around a set of banks, gravitating towards them as better mortgage offers appear. In this system, we would have a borrower acting as a metallic ball in a large financial system that acts as a pinball machine bouncing off different lenders.

Managing risk then becomes a similar trick:

The inputs that then cause reactions in the system (similar to the car braking above) can then be changed by:

  1. Adjusting the price of loans by lowering interest rates, making debt easier to service. This can be performed by central banks. This also relates to the velocity of money.
  2. Modify the loans, which can be performed by the lenders.
  3. Restricting the new loans made available by increasing interest rates. Banks would do this if they want to reduce the risk on their books (hope for churn of “risky” customers).

However, we can also identify other ways to manage risk and measure potential changes, particularly around macro-economic noise that may lead to significantly higher defaults or churn, depending on the conditions.

Internal Risks?

The thing to note here is that behavioural operational risk comes in various forms. In an ideal world, a pricing and cost mechanism can be established. To manage the risk and control temperament if it overheats, or to retrospectively adjust if poor behaviour is not picked up in time.

However, it is hard to reclaim salaries following poor behaviour and it is also hard to pay employees fully in deferred shares. In many ways, outsourcing and deferring a payment is easier here.

Another key aspect could potentially be keeping each product to separate teams, or where we expect things to go – with a separate company (FinTech). If a particular product emerges with a problem, the relationship or business unit can be replaced or removed.

Again, our modelling aims to track what would potentially lead to a change in behavior, most notably when behavior begins to focus on commission rather than customer outcomes.

Once we have achieved this, we may become significantly better at estimating our “through the crisis” revenue streams and understanding risk driven by both borrower and competitor behaviour, enabling us to proactively manage behavioural risks.

*We prefer to move away from stochastic modeling, instead preferring to build non-linear models based more on particle theory. We think this allows us to investigate potential behavioural shifts through bifurcation theory.

**The phenomenon is a sinusoidal compression wave travelling in the opposite direction to traffic flow.

Picture from www.fatvat.co.uk

Read More

6a010534b1db25970b0147e0ae51b2970b-800wi

Musings of a FinTech – Actionable Insights from Social Media

Being able to effectively mine data produced via social media is very topical, with the emergence of companies such as Thinknum providing metrics from social media and other sources to provide new insights into company performance.

However, there is a wider question here of what data can actually be harnessed to provide genuine insight and tangible value to both companies and/or individuals – the classic problem of extracting the signal from the noise.

Thinknum provides company metrics such as Twitter/Facebook followers, employees on LinkedIn and web site traffic which arguably could be useful indicators of a company’s health for investors. A recent FT article provides a good run down of some of the current crop of investor-offerings in this space.

 

In the area of housing, a recently published piece of research from Harvard, Facebook, NYU and the Bureau for Economic Research provides one such insight using data from Facebook. Entitled “Social Networks and Housing Markets”, it looks at how social media influences an individuals perception of the attractiveness of property investment.

The key takeaway from the paper is “Individuals whose friends experienced a 5 percentage points larger house price increase over the previous 24 months (i) are 3.1 percentage points more likely to transition from renting to owning over a two-year period, (ii) buy a 1.7 percent larger house, and (iii) pay 3.3 percent more for a given house. Similarly, when homeowners’ friends experience less positive house price changes, these homeowners are more likely to become renters, and more likely to sell their property at a lower price.”

It’s interesting to see how they combined the data sources for it – the model used Facebook user data along with market research data from Acxiom at its core to build rich demographic data.

One of the key uses of the Facebook friend data was the location of where an individual’s friends reside – specifically those that are within or outside of the Los Angeles county commuting zone (they surveyed homeowners all resided in LA county). This enabled the researchers to distinguish between local friend influences and biases, versus those further afield – the assumption being that house price movements experienced by friends outside the commuting zone would have been effected via social media channels (sec 1.4 p10).

This was supplemented with the relevant housing data, and a 4 question multiple-choice survey for testing the various hypotheses:

 

  1. How informed are you about house prices in your zip code?

[x] Not at all informed [x] Somewhat informed [x] Well informed [x] Very well informed

 

  1. How informed are you about house prices where your friends live?

[x] Not at all informed [x] Somewhat informed [x] Well informed [x] Very well informed

 

  1. How often do you talk to your friends about whether buying a house is a good investment?

[x] Never [x] Rarely [x] Sometimes [x] Often

 

  1. If someone had a large sum of money that they wanted to invest, would you say that relative to other possible financial investments, buying property in your zip code today is:

[x] A very good investment [x] A somewhat good investment [x] Neither good nor bad as an investment [x] A somewhat bad investment [x] A very bad investment

 

The ordering of the questions in 35% of the surveys was changed to avoid the framing effect with people’s responses, which was an interesting point to note (although they didn’t find participants were influenced by ordering of questions in this instance).

This survey & demographic data was then utilised alongside housing transaction data, and they created a number of regression models which supported the conclusions of the paper.

 

Given all of the talk about social media & mining this data, it’s a useful paper to be aware of and illustrates not only a potential use case for harnessing the power of social media to generate insights, but also how complex a task it is to do so.

Bearing in mind that the individuals in the survey were limited to those residing in LA county, and the measured impact of social media appears to influence individuals by up to ~3% which is pretty small in the grand scheme of things*, trying to apply a similar model to something similar like how social networks influence an individuals mortgage preference would be no small task!

 

*We are very excited overall that new data can not only lead to new ways of analysing risk but potentially be a strong leading indicator, allowing more time to rebalance portfolio risk. However making a judgment call on new data presents higher modelling risk.

 

Read More