As support for traditional asset models of the foreign exchange market fades, there is growing interest in more general models that include flows from international trade and international investment. One advantage of flow models is that they fit naturally into the recent literature on microstructure, particularly the work on order flow. My objective here is to use intervention data to help discriminate between traditional asset models of the foreign exchange market and more general flow models. The evidence supports a flow approach.

The exchange rate literature contains two inconsistent strands. There is a large theoretical and empirical literature on overshooting. In that literature overshooting is an important explanation for exchange rate volatility. A separate literature says that exchange rates are martingales and that models do not beat a random walk. Both can not be true. I show that the evidence for overshooting is highly suspect while the evidence that flexible exchange rates are approximately martingales is rock solid. Given the strength of the evidence, models that imply overshooting probably should be rejected out of hand.

What I call the modern theories of the law of one price and purchasing power parity extend those theories in two important ways: First by recognizing the nonlinearities caused by transaction costs and other impediments to trade and second by recognizing the importance of time in commodity arbitrage. This new approach, which has developed for the last two decades, raises questions about many widely accepted ideas. These ideas include the following: (1) The relevance of PPP. (2) Tests for cointegration and unit roots. (3) Relative versus absolute PPP. (4) Border effects. (5) Excessive volatility. (6) Large half lives for deviations from PPP. (7) The large increase in the volatility of exchange rates after the collapse of Bretton Woods.

A complete solution to the forward-bias puzzle should provide an econometric solution and an economic explanation for that solution. A complete solution should also explain the closely related failure of uncovered interest parity. In addition it should explain some related anomalies. One such anomaly is that variances for changes in exchange rates are over 100 times larger than variances for interest rate differentials and forward premiums. My econometric solution is that the relevant test equations omit two variables that covered interest parity implies should be included. For my data, the missing variables explain the failure of uncovered interest parity and the forward-bias puzzle. The missing variables also explain why the variance for changes in exchange rates is over 100 times larger than the variance for both interest rate differentials and forward premiums. My economic explanation is that, in general, forward rates do not equal expected future spot rates.