Kim-Miller-Ozanne (2004) Estimating and Forecasting Capital Gains with Quarterly Models

Summary

Kim, Miller, and Ozanne (2004) address the Congressional Budget Office's (CBO) need to estimate current-year and project 10-year capital gains realizations for revenue forecasting. They construct a tax-adjusted gains series — removing the mechanical effects of changing capital gains tax rates via an OLS pre-filter — and then propose two Bayesian vector autoregression (BVAR)-based improvements over CBO's historical mean-reversion benchmark (1-year root mean squared error (RMSE) 18.57 percentage points (pp) after tax adjustment). The two-step approach forecasts macro explanatory variables with a BVAR(5) and then regresses annual gains growth on forecast variable changes, achieving a 1-year RMSE of 14.80 pp (−20%); the integrated quarterly approach interpolates annual gains to quarterly frequency, includes them in a unified BVAR, and uses Kalman filtering to revise current-year estimates as quarterly data arrive, achieving 11.92 pp (−36%). BVAR methods dominate only at the 1–3 year horizon; mean reversion or random-walk with drift performs comparably or better beyond that.

Key Claims

CBO historical mean-reversion benchmark (gains/GDP → historical average) produced 1-year RMSE of 22.3 pp on raw gains and 18.57 pp after tax adjustment.
Tax adjustment: Miller-Ozanne OLS equation regresses $\Delta\log$ (gains/potential GDP) on $\Delta$ permanent tax rate, $\Delta$ transitory tax rate, $\Delta\log$ (output gap), and $\Delta\log$ (S&P 500 Q4); constructs counterfactual series as if 1998 tax rates applied throughout 1948–2000.
Two-step method (best variant E2): 1-year RMSE 14.80 pp, 20% improvement over the tax-adjusted CBO benchmark.
Integrated quarterly BVAR with linear interpolation (best overall): 1-year RMSE 11.92 pp, 36% improvement.
Linear interpolation wins over economic and in-model interpolation because annual residuals are positively serially correlated (Durbin-Watson = 0.78); the quarterly model amplifies revisions $1.6\varepsilon$ vs. $1.0\varepsilon$ in the annual model, providing richer within-year information content.
BVAR prior follows Sims-Zha Prior / Robertson-Tallman (1999) style with 6 hyperparameters ( $\mu_1$ – $\mu_6$ ): overall tightness ( $\mu_1$ ), first-lag diagonal weight ( $\mu_2$ ), intercept variance ( $\mu_3$ ), higher-lag decay ( $\mu_4$ ), unit-root dummy weight ( $\mu_5$ ), cointegration dummy weight ( $\mu_6$ ).
Best explanatory variables for two-step: real GDP, labor productivity, interest-rate spread, real compensation per hour, real fixed private domestic investment.
2001 out-of-sample: actual gains fell −46%; no model came close (best model −6.9%), suggesting a structural break around the 1990s boom.
Multi-year horizon: BVAR advantage fades; mean reversion or random-walk with drift comparable at 4+ years.
Three error types defined: E1 = estimation error (current year only), E2 = forecast|estimation error (using BVAR-estimated current year as starting point), E3 = forecast|actual (2-year combined).

Concepts Introduced or Extended

Entities Mentioned

Quotes

"The integrated quarterly model allows revisions to the current-year estimate as quarterly data arrive during the year, amplifying the information content of high-frequency macro releases."

"Linear interpolation dominates the more sophisticated interpolation schemes because the annual gains equation residuals are positively serially correlated (DW = 0.78)."

My Take

The paper's central contribution is demonstrating that mixed-frequency Kalman filtering can convert an inherently annual forecasting problem into a quarterly updating exercise with material RMSE gains. The tax-adjustment pre-filter is essential but relies on a single linear OLS equation; structural breaks in the tax-gains relationship (evident in the 1990s boom) are not modeled. The 2001 out-of-sample failure is a sobering reminder that capital gains are driven by asset-price dynamics that BVAR macro models do not capture in tail events. The BVAR approach adds most value at the 1-year horizon where quarterly revisions matter most; longer horizons are better served by simpler mean-reversion models, consistent with the broader evidence on BVAR forecasting performance.