Model instability and long-term investors

We analyze the effect of model instability on long-term investors using a time-varying VAR(1) model that we estimate using Bayesian Markov Chain Monte Carlo techniques for state-space specifications. Our model is able to handle time-varying intercepts, time-varying slopes, time-varying volatility, time-varying correlation, the leverage effect and fat tails. We calculate the optimal portfolios of long-term investors using numerical Monte Carlo techniques. Time-variation in intercepts and slope coefficients is not persistent enough to be important for long-term investors, while time-variation in the error covariance matrix (especially error volatility) is persistent and therefore very important for long-term investors. Fat tails disappear once time-varying volatility is incorporated. Random walk specifications (persistence equal to 1) or regime-switching models (same persistence for all parameters) lead to a large overestimation of perceived stock market risk and an underinvestment in the stock market. Results are robust to changes in the specification.