Jack M. Guttentag, a professor in the University of Pennsylvania's Wharton School, has proposed a new approach to underwriting, making use of modern computer technology. Under his proposed system, micro and macro projections and the lender's policy toward risk are specified. Each party responsible for the different aspects of underwriting can be held accountable. The following excerpts are taken, with permission, from an article in Volume 3, Issue 1 of Housing Policy Debate, a quarterly publication of the Federal National Mortgage Association.
This section sets out an ... approach toward automated underwriting that ... is ... easily implemented with current technology. But it requires lenders and investors to "bite the bullet" in terms of projecting the future.
... [T]he proposed system involves explicit macro and micro projections, with the underwriter responsible only for the latter. Furthermore, in both cases the characteristics of the mortgage that bear on default are an integral part of the projections, and the accept-reject decision is automatic once all the relevant information has been entered in the system.
There is one major difference, however. In the idealized version, it was assumed that the future could be defined in probabilistic terms, so that projections were strictly a technical exercise. In the proposed system, we must come to terms with the fact that we cannot define the future in this way, and therefore the projections must be part of the lender/investor's policy toward risk. Specifically, the lender/investor defines risk policy in terms of a set of maximum (or minimum) values of selected risk variables plus a set of macro projections. These factors will be considered in turn. ...
The limiting values of risk variables defined in the proposed approach resembles qualification requirements under the existing system, but there is an important difference. Under existing arrangements, the limiting values of risk variables are compared with actual values in month zero, whereas in the proposed approach they are also compared with projected values in every month thereafter until term. ...
Structuring the limiting value of risk variables in this way captures the potential interaction of payment risk and property risk.
Payment risk is the risk that the borrower at some point will be unable to meet the required payment. Property risk is the risk that at some point the value of the property will fall below the loan balance. The borrower's inability to meet the payment may not cause significant loss to the lender or insurer if the loan balance is amply covered by property value. Similarly, borrowers who can maintain their payments without strain may well continue to do so despite negative equity in their property. ... The greatest default risk is associated with an overlap of high property risk and high payment risk, and the method of specification shown above provides complete flexibility in accounting for this interaction. Where would the actual limiting values come from? From the same places that qualification requirements come from: experience, intuition and "gut feelings' in the short run, and, it is hoped, from careful research on default experience in the long run.
The crux of the underwriting system is a set of projections of future values of all risk variables for which limiting values are set. The underlying premise is that what really matters is what happens to risk variables in the future.
The standard objection to this view is that current values can be established with some degree of certainty whereas projections are strictly guesswork. ... This statement is true, but the inference that guesses should be avoided does not follow. Since it is the future that is relevant, an informed guess about the future is better than certainty about the present. To argue otherwise is akin to arguing that the hunter should shoot at the tree rather than the deer because the tree is easier to hit.
Assumptions about the future cannot really be avoided, because that is where it is happening; loans rarely default in the first month. The refusal to make explicit projections is tantamount to making implicit projections, which causes the underwriter to make macro forecasts of house prices and decide how much risk the lender should be taking. Since underwriters for the most part have tended to assume that house prices would rise in the future as they had in the past, the system worked well enough during the decades when the market was validating this assumption. It is not workable today.
Depending on the risk variables employed, projections probably will be needed for the following: (The loan balance is not included because it is derivable from the characteristics of the loan.)
1. Property value.
2. Borrower income.
3. Nonhousing debt.
4. Taxes and insurance.
5. Interest rates (for ARMs only).
Each of these projections has three components: definitions of alternative scenarios, quantification of each scenario, and selection of the particular scenario that will be used in the individual case. The lender/ investor is responsible for the first two and the underwriter for the third.
... In quantifying any of these scenarios, the lender might use different numbers for different parts of the country. It would be plausible, for example, to incorporate greater variability in property value scenarios in regions that have enjoyed exceptionally high rates of appreciation.
In the case of ARMs, interest rate scenarios would be specified by the lender, reflecting a policy decision regarding the rate environment with which the lender believes borrowers should be able to cope. Only one scenario is needed, and the underwriter has no role.
The lender might choose any of the following types of scenarios:
1. No change. The index rate to which each ARM is tied stays at the level prevailing at the time the loan commitment was made.
2. Worst case. The index rate rises to 100 percent in the second month, with the actual rate on the ARM determined by the rate caps [applied to] the particular instrument.
3. Bad case. The index rate rises according to some lender-defined scenario.
4. Historical. The index rate to which each ARM is tied changes from its start level according to the actual historical pattern of that particular index over a specified past period.
Within this underwriting system the individual underwriter does not accept or reject loans. (An exception may be categorical rejections based on adverse credit reports, as noted below.) Neither does the underwriter make macro projections or decide how much risk the lender should assume. Rather, the underwriter operates within his or her expertise to select the appropriate category for each aspect of the transaction, where the categories have been defined.
Once these categorizations have been made, the computer projects all the risk variables, compares them every month with the limiting values specified by the lender, and if none of the limits has been breached, automatically approves the loan, otherwise the loan flunks.
If the loan flunks, the underwriter can ask the system to provide the data showing which risk variables were violated and when. With this information, the underwriter might decide to return the file to the loan officer, with suggestions as to how the loan might be recast so that it will pass.
Recasting might involve, for example, buying down the rate, increasing the downpayment, or shifting to another type of mortgage.
Nothing prevents the lender from giving the underwriter discretion to override a rejection decision made by the system. There is much to be said for allowing underwriters to raise maximum allowable expense-income ratios, for example, much as they can do now where circumstances warrant it. The logic of the automated system militates against such discretion, however. (The underwriter would retain discretion to overrule a favorable decision made by the system on the basis of an adverse credit report.)
This approach to underwriting would enormously facilitate both the training of underwriters and the assessment of their performance. Each transaction would leave an objective trail of the underwriter's judgment on each of a number of features of the transaction that affect its risk. ...
In the long run, the categorizations for property value, borrower income, and the like that constitute the underwriting function could be automated as research revealed how different categorizations were related to different types of raw information in the file. Artificial intelligence could be applied to explain a spectrum of choices, separately for each of several aspects of the transaction. This application of AI has far better prospects for success than attempting to explain a single yes or no decision.