In Part I of this post (HERE), we discussed the Gambler’s Fallacy and how the loan process itself can inject bias into a bank’s decisioning. In particular, we looked at how the order of how loans are reviewed for credit makes a difference. This sequence bias comes from an inherent cognitive belief in humans that want to assume the world is less random than it is. Flip a coin enough times, and every time the result is heads in a row it is natural to assume that tails are due. Despite each coin flip is an independent event and that a coin has no memory of the past, flip five heads in a row, and the feeling is intense that tails will be next. Flip ten heads in a row and that intensity for tails is even more palpable. This syndrome is also alive for loan processing as many underwriters succumb to the process as we pointed out in Part I. The question is how do we limit this bias in banking so we can allocate risk and capital more efficiently?
One would think that the first stop to eliminating this cognitive bias would be to educate lenders as to the impact that sequencing credits has on loan decisioning. Surely this helps and so we suggest you forward this blog to others in your lending department to educate them for preventative measure. Unfortunately, we can’t represent that this will help. Just like that fateful night in Monte Carlo in 1913 where management took a pause after the 17th time that black showed up to explain to the crowd that the game was not rigged and that the odds that the ball drops into a red square were the same as it was during the first spin, the crowd chose to ignore logic. This cognitive bias was too strong, and gamblers loaded up on red until it finally showed up on spin 27.
A study by Beach and Swensson back in 1967 showed participants a shuffled deck of cards with various shapes on them. The non-control group was educated as to the Gambler’s Fallacy and were explicitly instructed not to let consecutive shapes influence their decision. At the end of the experiment, both groups had about the same outcome which helps support the notion that education alone is not enough to limit bias.
While education about the bias hasn’t proven that effective, the general academic level of education does. From the study that we referenced yesterday, “Decision-Making Under the Gambler's Fallacy: Evidence from Asylum Judges, Loan Officers, and Baseball Umpires,” Messrs. Chen, Markowitz, and Mme. Shue found that having a graduate level degree reduced the bias from being 3.2% less likely to approve if the previous loan was approved to 2.1% less likely to approve.
As we pointed out in Part I, the Gambler’s Fallacy is somewhat correlated to experience. The study referenced above by Harvard Law School and the University of Chicago updated in 2016 showed that loan officers with less than ten years of experience were 23% less likely to approve a loan if they approved a previous loan compared to a bias of 8% for the general population of lenders. If you adjust for all other factors and pull out those loan officers with more than ten years’ of experience, the bias drops from being 3.2% less likely to approve a credit if the previous credit was approved to just 1.3% less likely to approve. Thus, it appears that the higher number of credits that are reviewed the more a seasoned lender can apply that experience in an unbiased fashion. This argues that banks can offset this bias by having greater peer review.
On a related basis, being younger also had a large impact on the bias, as a younger loan officer was 6% less likely to approve a loan if the previous loan was approved.
Many banks build this in by having a “challenge rule” or the ability for the lender to ask for a different underwriter to review the credit with management oversight to get a difference of opinion. This structure has proven very effective at not only limiting the Gambler’s Fallacy but also eliminating other biases and moderating extreme opinions as well.
Per the graphic above, the more time a loan officer spends on a credit the less bias they exhibit. Conversely, if the average loan officer is 3.2% less likely to approve a loan if they just approved the previously loan, reducing the minutes spent on reviewing that loan caused an increase in bias and the loan officer became almost 6% less likely to approve a loan if the previous loan was approved.
Another mitigating factor comes in the form of restructuring the process. We have found that the longer the time between the sequence and the decision, the less sequence bias is present. Thus, having lenders structure reviews so that they complete the review and then move to another task or lunch will serve to dampen the bias.
To the extent the credits can be re-sequenced and reviewed by additional underwriters, this also helps to mitigate much of the Gambler’s Fallacy bias. The same goes if you can frame the second credit review as the start of the next sequence instead of connected to the previous credit series. In experiments, this re-framing of the sequential order helped participants reduce the bias error.
Finally, the rise of credit scoring helps eliminate this bias altogether. To the extent the credits in question can be scored, this serves to anchor the loan officer decision around the scored outcome thus partially mitigating the bias.
One capitalistic way to control for bias is through incentive compensation. The referenced study tested the impact of incentives. They created a payment unit equal to 1.5 times the hourly wage of each underwriter. They then created the following compensation schemes:
Flat: Here the loan officer received two units for approving a loan that ultimately performs, two units for approving a loan that ultimately does not perform and they received nothing for a rejected loan regardless of loan’s ultimate outcome.
Moderate: Here, loan officers received two units for approving a loan that ultimately performs, nothing if the approved loan does not perform and 1 unit if they reject a loan.
Strong: In this scheme, loan officers were compensated five units for approving a loan that ultimately performs, they had their total compensation reduced by ten units for approving a loan that ultimately does not perform and received nothing for rejecting a loan.
Below is the experiment’s outcome:
The differences across incentive schemes were material. Under a flat incentive structure, loan officers were 9.6% less likely to approve a loan if the previous loan was approved. For the Moderate Incentive structure, that bias was reduced so that loan officers were only 1.2% less likely to approve a loan if the previous loan was approved. For the Strong Incentive structure, the bias was negligible at 0.3% less likely.
Interestingly, the impact of incentives had a greater impact on loan officers with less than ten years’ experience. For the moderate and strong incentive scenarios, less experienced loan officers were approximately 5% and 6% less likely to approve a loan if the previous loan was approved, respectively.
Putting This Into Action
While none of these loan officers we studied had imposed quotas, many acted as if they did. These self-imposed quotas manifested themselves in part because we can surmise that the loan officers felt emotionally tied to their decision at some level. Our evidence for this comes from one study we researched where the loan officer recorded higher levels of guilt when they approved two loans and then turned down two loans compared to when they alternated so that they approved a loan, then turned down a loan, approved a loan and then turned down the last loan. This also explains why a loan officer is more likely to be biased after a streak of consecutive positive or negative loan decisions.
Acknowledging and understanding the Gambler’s Fallacy bias is the first step to building a lending team and a process around a framework that tries to limit this bias whenever possible. Creating a compensation structure that rewards positive outcomes while disincentivizing adverse results was the best for limiting this bias. In addition, making sure your most experienced lenders are handling credits with potentially higher expected losses that are approved sequentially will also help limit suboptimal outcomes.
Since the Gambler’s Fallacy is subtle and not as overt as say a stated bias against hospitality lending or loans without recourse, it is what makes this bias even harder to deal with since most lenders that succumb to the Gambler’s Fallacy have the best intentions. While true, since capital is precious, this shouldn’t stop banks from creating a system where banks can be protected against even the best well-intentioned thoughts.
Submitted by Chris Nichols on November 30, 2016