Despite high volatility, Bitcoin is known to offer diversification benefits through its relatively low correlation with stock markets. Unlike traditional safe-haven assets, Bitcoin prices strongly respond to time-varying correlations and diversification benefits. We find that a decrease (an increase) in correlation between Bitcoin and S&P500 index returns strongly predicts higher (lower) Bitcoin returns the next day. Under the classical mean–variance framework, we develop a stylized model of Bitcoin prices utilizing extreme disagreement among heterogeneous Bitcoin investors. When our model is calibrated to the observed predictability of Bitcoin returns, the model simultaneously explains the lack of predictability in traditional safe-haven assets such as gold and long-term treasuries.

We put forward a model in which analysts are uncertain about a firm’s earnings process. Faced with the possibility of using a misspecified model, analysts issue forecasts that are robust to model misspecification. We estimate that this mechanism explains approximately 60% of the autocorrelation in analysts’ forecast errors. The remainder stems from the cross-sectional variation in mean forecast errors and in analysts’ estimation errors of the persistence of earnings growth shocks. Consistent with our model, we find that analysts learn about some features of the earnings process but not others, and this learning reduces, but does not eliminate, the autocorrelation of forecast errors as firms age. Other potential explanations for the autocorrelation of analyst forecast errors are rejected. Our model of robust forecasting applies not only to analysts’ forecasts but also to all model-based forecasts.