When a rate of return is regressed on a lagged stochastic regressor, such as a dividend yield, the regression disturbance is correlated with the regressor's innovation. The OLS estimator's finite-sample properties, derived here, can depart substantially from the standard regression setting. Bayesian posterior distributions for the regression parameters are obtained under specifications that differ with respect to (i) prior beliefs about the autocorrelation of the regressor and (ii) whether the initial observation of the regressor is specified as fixed or stochastic. The posteriors differ across such specifications, and asset allocations in the presence of estimation risk exhibit sensitivity to those differences.

This study explores multivariate methods for investment analysis based on return histories that differ in length across assets. The longer histories provide greater information about moments of return, not only for the longer-history assets, but for the shorter-history assets as well. To account for the remaining parameter uncertainty, or ‘estimation risk’, portfolio opportunities are characterized by a Bayesian predictive distribution. Examples involving emerging markets demonstrate the value of using the combined sample of histories and accounting for estimation risk, as compared to truncating the sample to produce equal-length histories or ignoring estimation risk by using maximum-likelihood estimates.

Term-structure models from Cox, Ingersoll, and Ross (1985) imply that conditional expected discrete-period returns on discount instruments are linear functions of forward rates. Tests reject a single-latent-variable model of expected returns on U.S. Treasury bills, but two or three latent variables appear to describe expected returns on bills of all maturities. Expected returns estimated using two-latent-variables exhibit variation with business cycles similar to what Fama (1986) observes for forward rates. Inverted term structures precede recessions and upward-sloping structures precede recoveries.

The Arbitrage Pricing Theory is extended to a setting where investors possess information about future asset returns. A no-arbitrage pricing restriction is obtained with arbitrage conditioned on an investor's information. The pricing restriction contains unconditional factor loadings and either conditional or unconditional expected returns. Thus, tests of the theory can be based solely on time-series estimates of unconditional moments. Additional tests based on conditional expected returns are also appropriate.

This study investigates the sensitivity of tests of the CAPM to different sets of asset returns. Tests are conducted with market portfolios that include returns for bonds, real estate, and consumer durables in addition to common stocks. Even when stocks represent only 10% of the portfolio's value, inferences about the CAPM are virtually identical to those obtained with a stocks-only portfolio. In contrast, inferences are sensitive to the set of assets used in the tests.

During the past few decades, the fraction of the equity market owned directly by individuals declined significantly. The same period witnessed investment trends that include the growth of indexing as well as shifts by active managers toward lower fees and more index‐like investing. I develop an equilibrium model linking these investment trends to the decline in individual ownership, interpreting the latter as a reduction in noise trading. Active management corrects most noise trader–induced mispricing, and the fraction left uncorrected shrinks as noise traders' stake in the market declines. Less mispricing then dictates a smaller footprint for active management.