ECB Working Paper Introduces Entropic Tilting of Model Forecasts to SPF Histograms

Key Takeaways

• This ECB Working Paper (No 3284) is a research publication, not a regulatory requirement.
• It presents a method for adjusting model-based forecasts so they match survey probability histograms exactly, and it reports improved accuracy, especially during the Great Recession and COVID-19.
• Compliance teams should treat it as supervisory-adjacent context for model risk management and stress-testing practices, not as a new obligation.

Deep Dive

This ECB Working Paper (No 3284) is a research publication, not a regulatory requirement. It presents a method for adjusting model-based forecasts so they match survey probability histograms exactly, and it reports improved accuracy, especially during the Great Recession and COVID-19. Compliance teams should treat it as supervisory-adjacent context for model risk management and stress-testing practices, not as a new obligation.

The paper, by Todd E. Clark and Elmar Mertens, carries an ECB disclaimer stating that it does not represent ECB views. It creates no rules, reporting obligations, deadlines, or enforcement expectations. Its relevance to compliance is indirect. It signals how central banks and sophisticated forecasters may combine statistical models with external judgment, which can shape supervisory expectations over time for forecasting, scenario design, and risk assessment.

The key methodological change is a direct approach to incorporating survey density forecasts. The authors reweight simulated model scenarios by entropic tilting so they reproduce the U.S. Survey of Professional Forecasters (SPF) histogram probabilities exactly. They use Iterative Proportional Fitting to handle multiple histograms. Because whole scenarios are reweighted rather than altered, relationships between variables are preserved. The tested application is a Bayesian VAR with time-varying volatility on U.S. data from 1996 through 2025. Tilting improved point and density forecasts, including for variables the survey does not target, such as employment, investment, and long-term interest rates. Accuracy gains were concentrated around the Great Recession and COVID-19. The method performed about as well as moment-matching alternatives while being computationally lighter. The authors note it can be applied to the ECB's own SPF, and possibly to survey-based recession probabilities in the future.

For compliance and model risk teams, the main implications concern forecasting models used in stress testing, IFRS 9 expected credit loss (ECL) scenarios, ICAAP/ILAAP, and capital planning. The findings suggest that purely statistical baselines may lag expert judgment in turbulent periods, which is a model limitation worth documenting. Any adoption of tilting would constitute a model change under model risk governance (for example SR 11-7 or ECB model-risk expectations). It would require validation, documentation, and governance of the external survey inputs. Caveats include reliance on aggregate "consensus" histograms, auxiliary conventions for scenario combinations with no simulated draws (which matter mainly at COVID-19's onset), and results that cover only U.S. data.

Recommended actions: (1) Log the paper for awareness with model development and validation teams. (2) Assess whether current macroeconomic forecasting and scenario-weighting processes could benefit from structured judgment overlays, and ensure existing overlays are documented. (3) If the approach is piloted, run a validation covering out-of-sample performance, sensitivity to the survey inputs, and the treatment of empty scenario combinations. (4) Monitor for any ECB or supervisory guidance building on this work. No immediate compliance action or deadline applies, so the overall impact is low.

Source: European Central Bank