-z- AbstractMissing observations are a rule rather than an exception in panel data.In this paper we discuss several tests to check for the presence of selectivity bias in regression estimates based on panel data. One approach to test for selectivity bias i n these estimates is to specify the missing data mechanism explicitly and to estimate the response mechanism and the regres-

In this paper we analyze the persistence in the performance of hedge funds taking into account look-ahead bias (multi-period sampling bias). To do so, we model liquidation of hedge funds and analyze how it depends upon historical performance. Next, we use a weighting procedure that eliminates look-ahead bias in measures for performance persistence. In contrast to earlier results for mutual funds, the impact of look-ahead bias is exacerbated for hedge funds due to their greater level of total risk. At the four quarter horizon, look-ahead bias can be as large as 3.8%, depending upon the decile of the distribution. At the quarterly level, we …nd positive persistence in hedge fund returns, also after correcting for investment style. The empirical pattern at the annual level is also consistent with positive persistence, but its statistical signi…cance is weak.

In this paper, we analyze the economic value of predicting stock index returns as well as volatility. On the basis of simple linear models, estimated recursively, we produce out-of-sample forecasts for the return on the S&P 500 index and its volatility. Using monthly data, we examine the economic value of a number of alternative trading strategies over the period 1970–2001. It appears easier to forecast returns at times when volatility is high. For a mean-variance investor, this predictability is economically profitable, even if short sales are not allowed and transaction costs are quite large. The economic value of trading strategies that employ market timing in returns and volatility exceeds that of strategies that only employ timing in returns. Most of the profitability of the dynamic strategies, however, is located in the first half of our sample period.

Copulas offer financial risk managers a powerful tool to model the dependence between the different elements of a portfolio and are preferable to the traditional, correlation-based approach. In this paper we show the importance of selecting an accurate copula for risk management. We extend standard goodness-of-fit tests to copulas. Contrary to existing, indirect tests, these tests can be applied to any copula of any dimension and are based on a direct comparison of a given copula with observed data. For a portfolio consisting of stocks, bonds and real estate, these tests provide clear evidence in favor of the Student's t copula, and reject both the correlation-based Gaussian copula and the extreme value-based Gumbel copula. In comparison with the Student's t copula, we find that the Gaussian copula underestimates the probability of joint extreme downward movements, while the Gumbel copula overestimates this risk. Similarly we establish that the Gaussian copula is too optimistic on diversification benefits, while the Gumbel copula is too pessimistic. Moreover, these differences are significant.

If repeated observations on the same individuals are not available it is not possible to capture unobserved individua: chara.cteristics in a linear model by using the fixPd effects estimator in the standard way. If large numbers ot observations are available in each period one can use cohorts of individuals with common characteristics to achieve the same goal, as shown by Deaton (19}i5]. It is tempting to analyze the observations on cohort averages as if they are observation~on iodividua!s~.vhich are observed in consecutive time periods. In this paper we analyze under whic h co~:ditiuns this is a valid approach. Moreover, we consider the impact oí the construction of t.he cohorts on the bias in the standard fixed effects estimator. Onr results show that the ef[ect.s oC ignoring tlfe Cacf thal. only a sy~~th~.'i~-panel is availah!e will be small if the cohort sizes are suíl'icicntly large and if the true rneans within each cohort exhibit sufGcient timc variation. In applications the lattcr r,ondition seems hard to fulfill, which implies th~t fa.irly lar,ge cohort sizes (1U0, 200 individuals) are needed to va,lidly ignore the cohoit nature of the data.