The Science of Predicting Consumer Bankruptcy
By David R. Kelly and Gregg A. Weldon
INTRODUCTION
Worldwide, consumer bankruptcies have risen at an alarming rate in recent years. Reasons for this include a troubled economy, widespread availability of credit, lower standards for credit due to competition among lenders, and government regulations that favor the bankrupt party. Particularly disturbing for lenders is the fact that traditional approaches to quantifying credit risk have not proven to be particularly effective in ranking bankruptcy risk.
One of the major challenges in forecasting bankruptcy is that consumers that declare bankruptcy look quite different than the typical ‘bad’ consumer:
| Typical Bad | Bankrupt | |
| Number of Trades | 10.42 | 14.34 |
| Number of Credit Inquiries | 1.41 | 1.78 |
| Historical 30+ Trades | 3.69 | 3.93 |
| Number of Collection Items | 1.02 | 0.76 |
| Average Age of Trades | 40.61 | 42.14 |
| Revolving Credit Utilization% | 52.71% | 65.99% |
| New Trades (Last 24 Months) | 2.78 | 3.77 |
From the above chart, we can infer some ways in which a candidate for bankruptcy varies from a typical ‘bad’ account. For one thing, the bankrupt individual is much more credit active. Also, the overall credit quality does not appear to be worse than the traditional ‘bad’ (see mean number of 30+ trades and collection items). These differences are reflected in many of the models that are developed. Bankruptcy models tend to place more emphasis on utilization of credit, recent delinquencies, recent inquiries, etc.
Clearly, there are some significant differences between the typical consumer that becomes charged-off or seriously delinquent, versus the consumer that declares bankruptcy. Does this mean that it is impossible to predict whether or not a consumer will go bankrupt? While it is clear that traditional risk tools are not optimal in predicting bankruptcy, it is also clear that a divergent approach can be successful.
BANKRUPTCY MODELING
Given that ‘Bankruptcy’ is included in the BAD definition of most general risk models, many lenders feel that they are already taking an optimal approach to predicting bankruptcy. Is this a fact?
To test this hypothesis, we tested two scores:
- A general risk score with a 90+, charge-off, and bankruptcy BAD definition;
- A bankruptcy specific model.
Each consumer in the study had at least one trade on their credit bureau and was not bankrupt at the observation point. Both models have been validated on multiple data sets over time to insure model stability. It was determined that the scores had a correlation of 55.0%. This indicates a directional similarity, but not anything approaching overlap.
To further test the scores overlap, Kolmogorov-Smirnov statistics were produced for each score on each outcome (i.e. 90+ and Bankruptcy), as detailed in the following chart:
| Score | Outcome1 (90+) | Outcome2 (BK) |
| General Risk | 59.6 (max K-S) | 42.1 |
| Bankruptcy | 50.1 | 56.5 |
It is immediately obvious that neither score is optimal in predicting the opposite outcome. For example, the General Risk Score gets a K-S of 59.6 when predicting 90+ but only a 42.1 when applied to predicting bankruptcy. Likewise, the bankruptcy model gets a KS of 56.5 when predicting bankruptcy but only a 50.1 when predicting 90+.
A matrix of the two scores reveals the true value of a two-tiered risk management approach. For example, an overlapping group consisting of 4.5% of the population scored in the lowest 10% of the bankruptcy score AND the lowest 10% of the General Risk score. These people made up 30.6% of all bankruptcies. Many lenders avoid these lowest scoring applicants anyway. However, further analysis was done, assuming a General Risk score cut-off that would give a cumulative Bad Rate of 2.4%. By using the bankruptcy score to remove an additional 1.3% of this very clean population, 22% of the future bankrupts were eliminated. As an added bonus, the Bad Rate dropped to 2.2% as well. Because these two scores are predicting different results, the use of them in tandem makes for a much more powerful tool.
BEST PRACTICES IN BANKRUPTCY SCORE DEVELOPMENT
One of the fundamental pitfalls encountered when validating Client-developed bankruptcy models is the inclusion of individuals that are bankrupt at the time of observation. Incorrectly classifying these cases will result in a sub-optimal model going forward. Why should these individuals be excluded? If they are at bankruptcy at the observation point, it’s not reasonable to call them ‘Bad’ (bankrupt) at performance. In essence, you would be predicting behavior that has already happened. The resulting model may be very good in predicting these past occurrences, but proves much weaker in predicting those accounts that WILL go bankrupt. However, classifying these as ‘Good’ accounts also makes no sense. They were already bankrupt and unable to “become bankrupt again” during the performance time frame due to regulations in place in most countries. By excluding the already bankrupt accounts, we are left with a sample of customers whose behavior during the performance time frame will be entirely up to them.
A second potential pitfall relates to the performance timeframe. While standard timeframes range from 12 to 24 months, our recommendation does vary by client. For example, a performance timeframe of 24 months would be sub-optimal for a portfolio of loans with a maturity of 12 months. On the other hand, we do not recommend performance outcomes exceeding 24 months regardless of the length of term. The tradeoff between the relevance of aged data outweighs the benefits of a closely matched performance timeframe at that point, in our opinion.
Swap-set Analysis
The most realistic way to integrate a bankruptcy tool into a business environment is via the swap-set analysis. Essentially, the goal of this analysis is to swap the worst individuals above the prior (general risk) cut-off with the best individuals falling below said cut-off. With the assumption that the strategic goal is to minimize bankruptcy exposure while controlling for other types of risk, the following approach should be taken:
Swap-out Set (Bottom 10%)
| BK Scores | Count | Bads | Bankrupts | Bad Rate | BK Rate |
| 190-206 | 1,995 | 122 | 58 | 6.1% | 2.9% |
| 161-189 | 1,972 | 128 | 65 | 6.5% | 3.3% |
| 124-160 | 2,002 | 150 | 82 | 7.5% | 4.1% |
| 88-123 | 1,993 | 161 | 94 | 8.1% | 4.7% |
| 7-87 | 1,989 | 171 | 105 | 8.6% | 5.3% |
Swap-in Set (Top 10%)
| BK Scores | Count | Bads | Bankrupts | Bad Rate | BK Rate |
| 961-988 | 2,007 | 92 | 8 | 4.6% | 0.4% |
| 919-960 | 1,955 | 102 | 18 | 5.2% | 0.9% |
| 870-918 | 1,988 | 117 | 32 | 5.9% | 1.6% |
| 834-869 | 2,011 | 135 | 44 | 6.7% | 2.2% |
| 802-833 | 1,963 | 143 | 55 | 7.3% | 2.8% |
| Count | Bads | Bankrupts | Bad Rate | BK Rate | |
| Current Strategy | 100,578 | 3,923 | 1,106 | 3.9% | 1.1% |
| Less: Swap-out | 5,984 | 483 | 281 | 8.1% | 4.7% |
| Add: Swap-in | 5,950 | 311 | 57 | 5.2% | 1.0% |
| New Strategy | 100,544 | 3,751 | 883 | 3.7% | 0.9% |
| Net Gain | -34 | -172 | -223 | -0.2% | -0.2% |
In the above example, this lender could implement a bankruptcy score into an existing credit policy framework and reduce credit losses by a substantial margin. A side benefit of this approach is that the ‘swap-in’ population is likely to be underserved by traditional credit channels, resulting in less interest rate sensitivity and higher response rates.
CONCLUSION
As lenders continue to battle increasing bankruptcies while fending off decreasing market share, the addition of a bankruptcy model to the risk management toolbox becomes imperative.
The development of an appropriate bankruptcy tool requires a focus on the unique drivers of bankruptcy prediction. Our experience has been that there are several pitfalls that must be avoided in this development, chief of which is the treatment of individuals with a prior bankruptcy.
The above study, using data from a lender utilizing sophisticated custom tools (but no bankruptcy score), shows that it is possible to minimize bankruptcies without sacrificing loan volume or profits.
Dave Kelly is the founder and President of AnalyticsIQ, Inc.
Gregg Weldon is the Chief Analytical Officer of AnalyticsIQ, Inc.