Predictive regressions are subject to two small sample biases: the coefficient estimate is biased if the predictor is endogenous, and asymptotic standard errors in the case of overlapping periods are biased downward. Both biases work in the direction of making t ‐ratios too large so that standard inference may indicate predictability even if none is present. Using annual returns since 1872 and monthly returns since 1927 we estimate empirical distributions by randomizing residuals in the VAR representation of the variables. The estimated biases are large enough to affect inference in practice, and should be accounted for when studying predictability.

The flattening of yield curves at long-term maturities is proven to be approximately proportional to the reciprocal of the time to maturity under general conditions. This is a consequence of the persistence of earlier forward rates in the averaging process, which produces yields from forward rates. This relationship suggests the use of a “reciprocal maturity yield curve, ” which significantly facilitates the interpretation of the behavior of long-term yields by linearizing them for display over a shorter interval. This is illustrated using a yield curve for U.S. Treasury bills.

New results on the exact small sample distribution of the instrumental variable estimator are presented by studying an important special case. The exact closed forms for the probability density and cumulative distribution functions are given. There are a number of surprising findings. The small sample distribution is bimodal. with a point of zero probability mass. As the asymptotic variance grows large, the true distribution becomes concentrated around this point of zero mass. The central tendency of the estimator may be closer to the biased least squares estimator than it is to the true parameter value. The first and second moments of the IV estimator are both infinite. In the case in which least squares is biased upwards, and most of the mass of the IV estimator lies to the right of the true parameter, the mean of the IV estimator is infinitely negative. The difference between the true distribution and the normal asymptotic approximation depends on the ratio of the asymptotic variance to a parameter related to the correlation between the regressor and the regression, error. In particular, when the instrument is poorly correlated with the regressor, the asymptotic approximation to the distribution of the instrumental variable estimator will not be very accurate.

Econometric analysis of time series data is frequently preceded by regression on time to remove a trent component in the date. The resulting residuals are then treated as a stationary series to which procedures requiring stationarity, such as spectral analysis, can be applied. The objective is often to investigate the dynamics of transitory movements in the systems, for example, in econometric models of the business cycle. When the data does consist of a deterministic function of time plus a stationary error then regression residuals will clearly be unbiased estimates of the stationary component. However, if the data is generated by (possibly repeated) summation of a satisfactory and inevitable process then the series cannot be expressed as a deterministic function of time plus a stationary deviation, even though a least squares trend line and the associated residuals can always be calculated for any given finite sample. In a recent paper, Chan, Hayya, and Ord (1977) hereafter CHO) were able to show that a residuals from linear regression of a realization of a random walk (the summation of a purely random series) on time have autocovariances which for given lag are a function of time and thereafter that the residuals are not stationary. Further, CHO established that the expected sample autocovariance function (the expected autocovariances for given lag averaged over the time interval of the sample) is a function of sample size as well as lag and therefore an artifact of the detrending procedure. This function is characterized by CHO in their figure 1 as being effectively linear in lag (although the exact function is a fifth degree polynomial) with the rate of decay from unity at the origin depending inversely on sample size.

We hope to answer three questions: Has there been a structural break in postwar U.S. real GDP growth towards stabilization? If so, when? What is the nature of this structural break?We employ a Bayesian approach to identify a structural break at an unknown changepoint in a Markov-switching model of the business cycle. Empirical results suggest a break in GDP growth toward stabilization, with the posterior mode of the break date at 1984:1. Furthermore, we find a narrowing gap between growth rates during recessions and booms that is at least as important as any decline in the volatility of shocks.

The synthesis of the dynamic factor model of Stock and Watson (1989) and the regime-switching model of Hamilton (1989) proposed by Diebold and Rudebusch (1996) potentially encompasses both features of the business cycle identified by Burns and Mitchell (1946): (1) comovement among economic variables through the cycle and (2) nonlinearity in its evolution. However, maximum-likelihood estimation has required approximation. Recent advances in multimove Gibbs sampling methodology open the way to approximation-free inference in such non-Gaussian, nonlinear models. This paper estimates the model for U.S. data and attempts to address three questions: Are both features of the business cycle empirically relevant? Might the implied new index of coincident indicators be a useful one in practice? Do the resulting estimates of regime switches show evidence of duration dependence? The answers to all three would appear to be yes.

This paper reexamines the empirical evidence for mean-reverting behavior in stock prices. Comparison of data before and after World War II shows that mean reversion is entirely a prewar phenomenon. Using randomization methods to calculate significance levels, the authors find that the full sample evidence for mean reversion is weaker than previously indicated by Monte Carlo methods under a normal assumption. Further, the switch to mean-averting behavior after the war is about to be too strong to be compatible with sampling variation. The authors interpret these findings as evidence of a fundamental change in the stock returns process.

We examine a variety of models in which the variance of a portfolio's excess return depends on a state variable generated by a first-order Markov process. A model in which the state is known to economic agents is estimated. It suggests that the mean excess return moves inversely with the level of risk. We then estimate a model in which agents are uncertain of the state. The estimates indicate that agents are consistently surprised by high-variance periods, so there is a negative correlation between movements in volatility and in excess returns.