Matching Earnings Statistics Can Leave Saving Predictions Wide Open

A good match to what?

Quantitative economists often choose model parameters by calibration: adjusting them until the model reproduces selected statistics, such as earnings variability or the frequency of large losses. The appeal is transparency. Researchers can focus directly on features they consider economically important.

In Does Calibration Pick the Right Model?, Jiuyi Cheng and Alexis Akira Toda ask whether agreement with those targets also disciplines the model’s economic predictions. Their answer, in an application to earnings risk and household saving, is that it can leave substantial room for disagreement.

They compare calibration with likelihood estimation, which evaluates the joint distribution of each person’s complete observed earnings history. Instead of asking only whether selected summaries match, likelihood asks how well a model accounts for the observed sequences themselves.

Making whole-history comparisons practical

Earnings models commonly separate permanent differences between people, persistent shocks that fade gradually, and temporary fluctuations. The persistent component is unobserved. With nonlinear dynamics or non-Gaussian shocks, calculating likelihood can require difficult integration over its possible histories.

The authors approximate this continuous hidden component with a finite set of states, creating a hidden Markov model. Earnings observations remain continuous. A recursive calculation then combines possible hidden histories through matrix operations, while numerical integration handles permanent individual differences.

The likelihood calculation requires no simulation, although the paper uses simulations for diagnostics and calibration. The method also retains individuals’ entry ages and actual gaps between observations. Its practical scope is models with low-dimensional hidden states, and numerical approximation accuracy requires checking.

What the earnings histories favor

The application uses 28,422 observations for 4,683 men aged 30–60 from the Panel Study of Income Dynamics. Interviews span 1997–2023 and report earnings for alternating income years from 1996 through 2022. The researchers remove predictable age, education, and year effects before estimating earnings dynamics.

They compare four persistent-risk models: a Gaussian mean-reverting benchmark, called OU; a model with Gaussian-mixture innovations, GKOS; a model combining losses with lasting earnings damage, GMR; and their piecewise linear variance model, PLV, whose volatility depends on the persistent state. Each is estimated with either Gaussian or Gaussian-mixture temporary shocks.

Allowing mixture temporary shocks improves every model’s log likelihood by more than 1,500 points. Among those specifications, PLV ranks first, followed by GKOS, GMR, and OU. However, the statistical comparison does not clearly separate PLV from GKOS. A numerical winner is not an unambiguous statistical winner.

The authors also show that stationary versions of all four processes can share the entire earnings covariance function. Thus even agreement about variability and correlations across time need not identify the underlying risk process.

Better targets, worse histories

The main calibration exercise targets 70 statistics of earnings changes and levels, including dispersion, skewness, extreme changes, percentiles, and age-specific patterns. The authors fixed targets and weights before computing results.

Calibration reduces the target-mismatch objective by 78–87 percent relative to likelihood estimation. Yet every calibrated model fits complete histories worse than every likelihood-estimated model with mixture temporary shocks—including the benchmark with Gaussian persistent dynamics.

This is not a case of calibration matching everything perfectly. No calibration reproduces all targets within sampling error. Rather, substantial improvements on the chosen score do not translate into better whole-history fit.

Why the distinction matters for saving

The researchers feed the earnings processes into the same household saving problem, holding preferences, prices, mean earnings, and transfers fixed. Households cannot borrow. Mean raw earnings are normalized to one, so reported asset values are model quantities, not dollar estimates.

For PLV, stationary mean assets fall from 10.51 under likelihood estimation to 3.76 under the reported calibration. More importantly, evaluated PLV calibrations with similar target fit imply assets ranging from 3.1 to 9.3. GKOS produces a range of 2.2–8.2. These are neither confidence intervals nor exhaustive bounds.

The authors do not claim that likelihood discovers the true model. Every candidate is misspecified, and even preferred specifications miss extreme earnings changes. Survey reporting error also matters: treating all estimated temporary fluctuations as actual income risk overstates that risk in the saving exercise.

Their recommendation is a division of labor: use likelihood to compare candidate models, then use targeted statistics to diagnose their failures. Matching selected facts is useful evidence—but, in this application, it does not settle the economic answer.