We present an heterogeneous-agent asset pricing model where fundamental shocks lead to amplification cycles (bubbles), and the principle of contrarian opinion holds: in equilibrium, less-informed speculators become overly optimistic (pessimistic) when prices diverge enough from fundamentals, and the bubble is likely to burst. Informed forward-looking speculators find it optimal to ride the bubble until a time when they switch the sign of their positions. At this switching time, the bubble continues to grow as less-informed speculators become more optimistic (pessimistic). Based on the implications of the model, the aggregate speculators' net positions can be expressed as a Markov-switching model. We take these switching times as leading indicators of trend reversals and use them to construct out-of-sample directional forecasts of exchange rate changes. We then test the accuracy of our directional forecasts vis-a-vis the random-walk using the standard binomial test and its profitability by a new weighted directional test. Across the five major currencies, our forecasts outperform the random-walk over forecasting horizons from 1 months to 12 months.