Survey histogram tilting improves model forecast accuracy
Entropically tilting Bayesian VAR forecasts directly to Survey of Professional Forecasters histograms improves predictive accuracy across US macroeconomic variables. Todd Clark and Elmar Mertens develop an exact multi-histogram framework evaluated on real-time data from 1996 to 2025.
Exact matching via iterative fitting
The authors develop an analytical entropic tilting framework solved by Iterative Proportional Fitting to match survey histogram probabilities from the US Survey of Professional Forecasters without parametric approximations.
Applied to a Bayesian VAR with stochastic volatility and fat tails across real-time data from 1996 to 2025, the method targets annual GDP growth, unemployment, and core PCE inflation up to three years ahead.
Nowcast GDP growth root mean squared error falls by nearly half, while unemployment forecast errors decline by 25 to 30 percent at medium horizons.
Accuracy gains also propagate to non-targeted variables, reducing 5-year Treasury yield errors by 6 to 8 percent.
Turning points expose model lag
Existing literature extracts only mean or variance moments by fitting continuous distributions to survey histograms, discarding tail information and risking numerical instability.
Direct histogram tilting achieves comparable accuracy to mean-variance matching while eliminating density-fitting steps and runtime spikes.
Accuracy gains concentrate heavily around turning points like the 2008 financial crisis and the 2020 pandemic, when professional forecaster judgment adjusted faster than statistical models.
Clean math with empirical punch
The framework provides central bank modelers a clean alternative to arbitrary parametric moment-fitting.
Yet its reliance on uniform mixture padding during severe shocks like 2020 underscores ongoing Monte Carlo limits.
For routine policy work, direct survey density integration is a major practical improvement.