The Predictability of Returns with Regime Shifts in Consumption and Dividend Growth

Anisha Ghosh, George A. Constantinides2010 · DOI

Abstract

The predictability of the market return and dividend growth is addressed in an equilibrium model with two regimes. A state variable that drives the conditional means of the aggregate consumption and dividend growth rates follows different time-series processes in the two regimes. In linear predictive regressions over 1930-2009, the market return is predictable by the price-dividend ratio with R2 11.7% if the probability of being in the first regime exceeds 50%; and dividend growth is predictable by the price-dividend ratio with R2 28.3% if the probability of being in the second regime exceeds 50%. The model-implied state variables perform significantly better at predicting the equity, size, and value premia, the aggregate consumption and dividend growth rates, and the variance of the market return than linear regressions with the market price-dividend ratio and risk free rate as predictive variables.

Introduction

For decades, financial economists have debated whether stock returns are predictable. Early theoretical frameworks suggested that the price-dividend (P/D) ratio should serve as a reliable forecast for future returns or dividend growth. However, empirical studies in the early 2000s—most notably Welch and Goyal (2008)—challenged this notion, arguing that standard linear models performed poorly, particularly in out-of-sample tests.

This paper proposes that the perceived instability of predictive models is not due to a lack of predictability, but rather the flawed assumption of a single, stable economic relationship over time. By developing an equilibrium model that accounts for regime shifts, the authors demonstrate that the P/D ratio predicts returns in one economic state and dividend growth in another, significantly outperforming traditional benchmarks.

Theoretical Framework: A Two-Regime Equilibrium Model

The core of the study is a dynamic equilibrium asset pricing model characterized by regime-switching latent variables. A state variable ($x_t$) drives the conditional means of aggregate consumption and dividend growth rates. Crucially, this variable follows distinct time-series processes depending on which of two regimes the economy currently occupies.

To capture the complexities of investor behavior, the model utilizes Epstein-Zin recursive preferences. This allows for the separation of a consumer's risk aversion from their intertemporal elasticity of substitution (the desire to smooth consumption over time).

A significant technical contribution of this research is the "inversion" of observable financial data. The authors demonstrate how to express latent state variables—specifically the state $x_t$ and the probability of being in a specific regime $p_t$—as non-linear functions of the observable P/D ratio and the risk-free rate. This allows researchers to infer the underlying economic state from market prices.

Methodology and Data

The model parameters were estimated using the Generalized Method of Moments (GMM), based on Euler equations for various assets and unconditional moments of consumption and dividends. The study utilizes annual US data from 1930 to 2009, including:

  • Market Returns and P/D Ratios: Sourced from the Center for Research in Security Prices (CRSP) value-weighted index.
  • Risk-Free Rate: Inflation-adjusted one-month Treasury Bill returns from Ibbotson Associates.
  • Consumption Data: Per capita personal consumption expenditure on nondurable goods from the Bureau of Economic Analysis (BEA).
  • Portfolio Premia: Size (Small vs. Large) and Value (High vs. Low book-to-market) portfolios following the Fama and French (1993) methodology.

The Predictability Trade-off

The research reveals a "predictability trade-off" between returns and dividends. Predictability is essentially a zero-sum game between these two variables, dictated by the economic regime:

  1. Regime 1 (Return-Predictable): This state is more frequent, with an average duration of approximately 10 years. It is characterized by high-frequency, less persistent growth. In this regime, the P/D ratio strongly forecasts future returns but has little predictive power for dividend growth.
  2. Regime 2 (Dividend-Predictable): This state is shorter, averaging 3.3 years, and is characterized by low-frequency, highly persistent growth. Here, the roles reverse: the P/D ratio predicts dividend growth, while its ability to forecast returns diminishes.

This finding explains why standard linear regressions often "wash out" predictive signals; by averaging across both regimes, the distinct signals cancel each other out. Furthermore, the relationship between financial ratios and future returns is highly non-linear, meaning simple quadratic functions cannot capture the regime-switching dynamics.

Key Empirical Results

The empirical performance of the regime-switching model significantly outperforms traditional benchmarks across several metrics:

  • Regime-Specific $R^2$: In linear regressions over the 1930–2009 period, the P/D ratio predicts market returns with an adjusted $R^2$ of 11.7% when the probability of being in Regime 1 exceeds 50%. Conversely, it predicts dividend growth with an $R^2$ of 28.3% when the probability of being in Regime 2 exceeds 50%.
  • Out-of-Sample (OOS) Success: During the 1976–2009 period—a timeframe where standard linear models famously failed—the model-implied variables achieved an OOS $R^2$ of 5.1% for the equity premium and 10.1% for market returns. In contrast, standard linear models produced negative $R^2$ values, performing worse than a simple historical average.
  • Volatility and Premia: The model-implied state variables also provide superior forecasts for the variance of market returns and the cross-section of stocks, including the size and value premia. This suggests that regime shifts are fundamental to the entire equity market, not just aggregate indices.

Conclusion and Limitations

The study demonstrates that market predictability is robust when structural shifts are accounted for. While the two-regime model provides a significant improvement over linear benchmarks, the authors acknowledge several areas for future work. There is no inherent reason why the economy must be limited to two regimes; increasing the number of regimes could improve accuracy, though it risks overfitting. Furthermore, the current model assumes consumers observe the state variable directly, whereas econometricians must infer it. Future research could explicitly model the optimal filtering process used by consumers and expand the framework to other asset classes, such as bonds and derivatives.

Summary updated

Abstract

The predictability of the market return and dividend growth is addressed in an equilibrium model with two regimes. A state variable that drives the conditional means of the aggregate consumption and dividend growth rates follows different time-series processes in the two regimes. In linear predictive regressions over 1930-2009, the market return is predictable by the price-dividend ratio with R2 11.7% if the probability of being in the first regime exceeds 50%; and dividend growth is predictable by the price-dividend ratio with R2 28.3% if the probability of being in the second regime exceeds 50%. The model-implied state variables perform significantly better at predicting the equity, size, and value premia, the aggregate consumption and dividend growth rates, and the variance of the market return than linear regressions with the market price-dividend ratio and risk free rate as predictive variables.
BibTeX
@misc{ghosh2010predictability,
  author = {Anisha Ghosh and George A. Constantinides},
  title = {{The Predictability of Returns with Regime Shifts in Consumption and Dividend Growth}},
  year = {2010},
  doi = {10.3386/w16183},
}
Figure