Dividend yield predictability is the finding that a stock's dividend-to-price ratio forecasts its future returns. In "Dividend Yields and Expected Stock Returns: Alternative Procedures for Inference and Measurement," Hodrick (1992) used CRSP data from 1926 to 1987. The 744-observation sample showed a one percent dividend-yield rise implied seven percent higher annual returns at one year and four percent at four years. A VAR-implied R² for four-year returns reached 37.3% in the 1952–1987 subsample, nearly triple the 12.7% full-sample estimate.
What the Study Found
The joint chi-square(3) test statistic for lagged returns, dividend yields, and T-bill rates equals 22.765 in the 1952–1987 subsample. Its confidence level exceeds .999, versus 9.867 and a .980 confidence level in the full 1927–1987 sample. A test using lagged returns alone has an 80% Type II error rate in Monte Carlo simulations. The joint test using all three variables cuts that error rate to just 2.5%. VAR-implied coefficients show a one percent dividend-yield increase raises annualized returns by seven percent at one year and four percent at four years.
"Since the vector autoregressive alternative has correct size and supplies long-horizon statistics that appear to be unbiased measurements, it emerges as the preferred technique."
Hodrick (1992), Dividend Yields and Expected Stock Returns: Alternative Procedures for Inference and Measurement, p. 29.
Methodology
The data come from CRSP monthly series on NYSE value-weighted returns, dividend yields, and one-month Treasury bill rates. The full sample has 744 monthly observations from January 1926 to December 1987. Three sub-periods are used: Sample A (730 observations, 1927:2–1987:11), Sample B (431 observations, 1952:1–1987:11), and Sample C (299 observations, 1927:2–1951:12). The VAR controls for lagged returns, lagged dividend yields, and the relative Treasury bill rate, with term and default premiums as additional controls.
Key Statistics
| Metric | Finding | Context |
|---|---|---|
| Joint χ²(3) test statistic | 22.765 | Sample B: 1952:1–1987:11, confidence level >.999 |
| Implied VAR slope coefficient, β(12)/12 | 7.147 | Sample B, one-year horizon |
| Implied VAR slope coefficient, β(48)/48 | 4.424 | Sample B, four-year horizon |
| Implied R²₂(48) | .373 | Sample B, four-year horizon |
| VAR-implied long-horizon slope | β(k) = e1'[C(1)+...+C(k)]e2 / [e2'C(0)e2] | Derived from VAR autocovariances, eq. (16) |
| Type II error rate, joint test | 2.5% | Monte Carlo, .05 critical value, alternative hypothesis |
Full Sample vs Post-1951 Subsample
| Measure | Sample A (1927:2–1987:11) | Sample B (1952:1–1987:11) |
|---|---|---|
| Joint χ²(3) test statistic | 9.867 | 22.765 |
| Confidence level of joint test | .980 | >.999 |
| Implied R²₂(48), four-year horizon | .127 | .373 |
Why This Matters
Investors timing equity allocations with dividend yields face a real methodological choice in how the predictability is tested. Standard OLS long-horizon regressions can overstate statistical significance unless standard errors account for overlapping-return bias. Hodrick's Monte Carlo evidence shows a vector autoregression avoids this bias by deriving long-horizon statistics from a well-behaved one-step-ahead model. Quantitative managers building valuation-based timing signals should note that the inference procedure chosen can change whether a signal looks statistically reliable.