Allan Timmermann

Research Interests

  • Asset Pricing and Learning Effects

    This research is concerned with explicitly modeling how agents form expectations of future payoffs when the parameters of the payoff distribution, as well as its functional form, are unknown to investors. The research takes off from the standard assumption of rational expectations and then considers asset prices under adaptive and Bayesian learning rules. My most recent research in this area focuses on solving for asset prices in the presence of structural breaks in the dividend process and studies learning effects in asset prices immediately after such structural breaks. This work also characterizes analytically the dynamic properties of asset prices in the presence of learning effects.

    Collaborators on this work include Massimo Guidolin from UCSD.

  • Cyclical Variations in Stock Returns

    A large literature in finance shows that stock returns is, to some extent, predictable by means of forecasting variables such as dividend yields, interest rates and macroeconomic variables with a clear business cycle component. However, standard forecasting models essentially assume that the same forecasting specification stays in effect throughout the entire sample.

    If stock prices are predictable, simple notions of market efficiency would suggest that they should not be predictable for very long periods of time. The challenge is to identify the 'pockets in time' where predictability prevails. Predictability is likely to be linked to the state of the economy. In periods with abundant liquidity investors will exploit any predictable patterns, while in periods with low liquidity, such as during economic recessions, predictable patterns may still be present in stock prices.

    Part of this research focuses on forecasting models that separate the data on stock prices into two states, namely a recession state and an expansion state. Both first-moment predictability ass well as predictability of volatility, skewness, kurtosis and the full density of returns are considered in the context of time-varying mixtures of normal distributions. I find that the best forecasting model is very different during recessions and expansions. Forecasting variables such as interest rate shocks and default premia have a large impact on stock prices during recessions and a much smaller impact during expansions. This research also finds that stock returns can be predicted only during the very short periods around turning points of the economic cycle. Lastly, evidence on how to exploit predictability in a trading strategy is reported and used to discuss the consequences of the evidence for market timing strategies.

    The research also considers whether stock returns are predictable even after accounting for model specification uncertainty and possible shifts in the underlying forecasting equation. Finally we test the predictions of recent imperfect capital market theories with regard to cyclical asymmetries in the stock returns of small and large firms.

    Collaborators on this research include Hashem Pesaran, University of Cambridge, and Gabriel Perez-Quiros, New York Federal Reserve Bank.

  • Data-snooping

    High frequency asset returns appear to be predictable by means of technical trading rules and simple deterministic calendar effects. However, both findings are the outcome of the financial research community's extensive search across hundreds of possible forecasting models. Since none of the technical trading rules or calendar effects were predicted ex ante by theory, it is quite possible that the findings reported in the finance literature are simply the result of extensive data-snooping, i.e., the fact that the same data set was used to formulate the hypothesis (that a given forecasting rule is successful) and test it. This research applies a new bootstrap procedure to assess the performance of the best forecasting model in the context of the full set of forecasting models under consideration. Our findings suggest that data-snooping effects can be very significant. For example, it can be strongly rejected that the best calendar rule, when viewed in isolation, does not possess predictive power over daily stock returns. Once data-snooping effects are accounted for, however, the best calendar rule no longer produces returns that are statistically significantly different from those of a simple passive investment strategy. Studying the predictive power of a large set of technical trading rules on daily Dow Jones data after 1986, we similarly find that, after accounting for data-snooping effects, there is no evidence that the best trading rule outperforms a simple benchmark.

    Collaborators on this work are Ryan Sullivan and Halbert White, both from UCSD.

  • Mutual Fund and Pension Fund Performance

    This research investigates various aspects of the performance of mutual funds in the UK. Questions addressed include the risk-adjusted performance of funds, the persistence of their performance, evidence of performance clustering across mutual fund managers, returns to strategic asset allocation, market timing and security selection, and mutual fund managers' performance incentives.

    Collaborators on this work are David Blake, Birkbeck College, and Bruce Lehmann, UCSD.

  • Duration Analysis of Financial Data sets

    One part of this research models the hazard of mutual fund managers. It computes the conditional probability of fund closure given a fund's age and the performance of the sector in which the fund operates as well as the risk-adjusted performance of the fund itself. These are important to understanding the incentives under which mutual fund managers operate. For example, if a small under-performance strongly increases a young fund's chances of being closed down, then this would lead managers to follow very conservative investment strategies. In fact we find that the effect of under-performance on fund hazard rates are very large; we also find that very young and very old funds are less likely to be closed down.

    Future research also plans to study duration dependence in bull and bear markets.

    Collaborators include Asger Lunde, University of Aarhus and David Blake, Birkbeck College.

  • Forecasting with non-linear (and possibly non-stationary) models

    My interest in this area originates from the observation in financial markets, that many time-series not only are dominated by highly non-linear effects but also are non-stationary. For example, a simple notion of market efficiency suggests that if returns on a given asset is predictable at a given point in time, then such predictability should rapidly dissipate as a result of investors allocation of capital into exploiting the resulting investment opportunity. Hence conditional forecasting relations cannot be expected to be stable in the financial markets, unless of course they simply reflect time-varying risk premia. Some of my current work addresses the issue of modeling predictability of asset returns in 'pockets in time'.

    Collaborators include Hashem Pesaran, University of Cambridge, and Gabriel Perez-Quiros of Federal Reserve Bank of New York.