Overview
With the uncertainty and possible economic harm surrounding the recently-announced tariffs—whether they stay, disappear, reappear, or mutate—bankers face the challenge of integrating their effects into Current Expected Credit Loss (CECL) and stress test estimates.
To generate credible estimates, they might consider the following actions:
- Adjusting scenarios to effect modeled outcomes
- Creating overlays of modeled results
- Expanding sensitivity and outcomes analyses, including leveraging other models
The Challenge of Current Scenarios
Take Capital Planning and Stress Testing (CPST) as an example: the Federal Reserve published CCAR scenarios in February 2025. However, those scenarios don’t account for the effect of tariffs—whether they be broad tariffs, extreme-but-targeted ones or any type for that matter.
Small banks, unlike the largest ones that are required to incorporate global market shocks into their CPST forecasts, often do not include the effects of foreign trade in their loss forecasting models, even on segments like Commercial and Industrial (C&I) loans, where such effects might be direct (but also likely to be unmeasurable over a portfolio’s sample history).
With first-quarter earnings announcements approaching, we anticipate various responses from banks. Some may shift weight from business-as-usual (BAU) scenario outcomes to adverse scenario outcomes, while others might delay adjustments until the second quarter.¹
Smaller banks that rely on external scenario providers will surely hope that those vendors provide insights along with reasonable forecasts, but that’s a tall order!
Even if the tariffs are immediately revoked, they are still relevant while their threat remains. Indeed, the uncertainty over tariffs will influence business investment, consumer confidence and spending, etc., for the foreseeable future; so, it’s crucial to acknowledge that this month’s “business-as-usual” forecast may be quite different from last month’s. (Not surprisingly, exogenous shocks have shock value; there is economic and psychological hysteresis.)
Unprecedented Times
Recent developments lack clear historical parallels. For forecasting, that prevents the use of available and reliable correlations among economic variables derived or calculated under similar conditions. Additionally, other historical relationships may falter in novel settings, i.e., all correlations go to one in a crisis.
So, any scenario incorporating tariffs or their threats—whether internally generated or externally sourced—will require subjective, qualitative adjustments. The lessons from Smoot-Hawley Tariffs may offer qualitative insights into potential economic downturns but fail to provide quantitative relevance due to societal and technological changes since the Depression era.²
While such adjustments aren’t inherently negative, transparency is key. Banks must be able to ask and answer:
Knowing how providers built their forecasts—and the evidence to support those choices—determines the scenario’s efficacy; otherwise, it becomes “made-up,” rather than hypothetical.
Large banks often customize generically-generated scenarios via economic forecasting committees that adjust rough forecasts to reflect market intelligence and other portfolio-specific information. Smaller banks might benefit from forming ad hoc committees to evaluate and suggest necessary adjustments over the coming months.³
Recommended Actions
Below, we provide several ‘do’s and a ‘don’t’ that should help the reader solve his or her highly constrained, loss forecasting problem in the best, most defensible, way possible. The best solution may not be good, and might not bring much comfort, but it will be defendable, which will make any loss forecasts as credible as they can be.
✅ DO: Respond to Uncertainty with Multiple Scenarios
When dealing with the binary or ternary nature of tariffs—on, off, maybe—and their uncertain effects, banks should develop and use a variety of scenarios, e.g.:
- Threats/no taxes
- High tariffs
- Extreme tariffs causing a global trade war
Note that substituting a generic downturn scenario without evidence that it fully captures the harm caused by a global trade war would likely lead to misleading results. For example, it would be a challenge to show that shocks experienced in 2008’s Global Financial Crisis, by themselves, are proper proxies of The Great Depression’s effects. A starting point, possibly, but with work to do, and when that work is done, the result would be usable but likely unsatisfying. Such is life.
The queasiness that one feels using subjectively determined, hypothetical scenarios can be somewhat mitigated with more work.
✅ DO: Use All That You Have or Can Build Cheaply
One of Spero Risk’s development principles is: never throw anything away. During development, banks select a model of record at the cost of discarding other reasonable specifications—ones with different parameters, transformations, regression methods, etc. Our steady advice is to incorporate the results of reasonable alternative specifications as part of quarterly outcomes analysis. Not surprisingly, results tend to be similar, which provide corroborating evidence and a degree of comfort.
Conversely, when results vary, managers have the option of generating overlays by weighing the alternative specifications, especially if those specifications contain regressors or transformations with wider domains, i.e., relatively higher maxima or lower minima in the sample set. Wider domains may provide greater sensitivity in certain situations by allowing for bigger shocks.
For those who haven’t followed this expansive approach, no worries… as long as development processes were SR 11 – 7 compliant. If so, search the developmental documentation and evidence for reasonable, but discarded, specifications that can be used justify overlays. What’s better than a qualitative adjustment that’s based on actual empirical evidence? Not much.
Additionally, simple models using call report or other publicly available data can be used to further justify overlays and can be built surprisingly fast. (While there are obvious differences, we frequently apply optimal relationships for a portfolio to industry-wide data, and that gives evidence that our custom findings hold more broadly and through longer periods of time than the internal sample history.) While the approach isn’t perfect, similar methods can help justify overlays.
Lastly, this might be the time to re-estimate models with stale data. If models haven’t been updated since their original CECL implementation, it’s probably time to see if the old relationships still hold. Or it might be time to build versions that capture the stability of good times and heightened sensitivity in downturns, which is Spero Risk’s regime-based approach.
✅ DO: Ask and Answer ‘What If?’ until Exhaustion
Ask your most curious, creative, and obsessive-compulsive managers and analysts to analyze sensitivities and outcomes every way possible. Go beyond internal data. Don’t be limited by statistical tests. In unprecedented times, any hint of possible outcomes is useful.
The task isn’t equivalent to a criminal case—beyond a reasonable doubt—or even a civil case—where a preponderance of evidence is the standard. It is much simpler: is there any evidence, anywhere, that could help answer: how much could we lose?
❌ DON’T: Extrapolate
At Spero Risk, we love simple solutions (as the ultimate expression of sophistication) but never simplistic ones. Don’t fall victim to the Politician’s Fallacy: “We must do something. This is something. Therefore, we must do this.”
Avoid extrapolation!
The fact that you can do something is quite different than that you should do it. Extrapolations are projections without evidence. Unfortunately, extrapolations—being calculations—may seem logical or scientific, but they’re neither. That makes their use both misleading and dangerous.
While this isn’t an example of extrapolation, it gets to the same point: multiplying five apples by six oranges doesn’t result in 30 apple-oranges but the same 11 pieces of fruit. The calculation is meaningless—just like extrapolation. By violating the assumptions of your model, you’ll have a number, but it won’t represent anything.
Note: this is also why data miners shouldn’t expediently drop sample outliers in hopes of a better fit. There might be a time or place or scenario, when the model owner really wished that the sample domain and range were wider.
If forecasted variables (or transformations of forecasted variables) are outside of the historical domains, bind them at their maximum or minimum values. Ideally, this should be controlled automatically within your loss-forecasting platform/implementation. Supplemental models built on call report data that have longer sample histories and, therefore, the possibility of more extreme values that may get closer to extreme forecasts.⁴
Projecting or extending relationships beyond their relevant domain may be a last resort, but don’t hide action within an abused model. Be transparent and explicit, and make the adjustment, which is likely a first-order linear projection explicitly as an overlay.⁵
✅ DO: Build a Comprehensive Overlay Framework
Use a well-defended Qualitative Adjustment Framework (QAF), including—but not limited to—adjustments justified by other, related models and/or outcomes analyses that use other variables (that don’t violate their regression method’s assumptions).
This recommendation goes beyond the current problem of modeling tariffs’ effects. While it can’t cover every contingency, a QAF should be built during development to compensate for known weaknesses immediately upon implementation.
Think of a framework as a composite function or mapping. If the model is a function, f(x), with inputs x, then the reported forecast value is g(f(x), x, y) where y are other, non-modeled factors to consider that could affect the final estimate.
Beforehand, one can’t include every unprecedented event among the y-factors, but for tariffs, applicable y-factors include other variables, other models, outcomes analyses, reviews of related historical settings, etc.
QAFs formalize overlay procedures and reduce their ad hoc nature to the extent that certain overlays can be automatically calculated and proposed. Given the novelty of the current situation, automation for tariff-specific adjustments is unlikely. However, a relatively comprehensive QAF will incorporate recent trends that are likely to move in negative directions, and those indicators can be used as evidence for overlays—one variable at a time.
Conclusion
Following our recommendations won’t necessarily bring comfort. That’s hard to find in unprecedented, shocking, and uncertain times, especially when one’s job involves forecasting.
For the forecaster, the best outcome is to hear “nothing,” when he or she asks an internal or external party scrutinizing their work: “what should we have done differently?”
Please contact Andy at 205.423.5668 or Alex Zaboli, our Bank Development Advisor, at 703.909.4528 or [email protected] or use our contact form.