- 10 -
must be more volatile than expected depreciation. Both types of risk premium models fail
on the second count. In a static CAPM, Lewis (1995) showed that the risk premium’s
variability must come from one or more of the following three: (a) shares of home and
foreign wealth held in different assets (b) shares of world wealth held by home and
foreign countries or (c) conditional variance of exchange rate and co-variances between
exchange rate and domestic and foreign inflations. None of the three are sufficiently
variable in the data to account for the high variability in risk premium. Other portfolio
balance models, for example those discussed in Giovannini and Jorion (1989), Flood and
Rose (1996) and McCallum (1994) postulate the risk premium as a function of volatility
of exchange rate; and of considerations of liquidity, size and depth of financial markets5
[Davanne, 1990]. Lucas (1982) model can in theory explain the higher variability of risk
premium than of exchange rate change, but the degree of risk aversion required to obtain
this result is very large [Bekaert and Hodrick, 1992]. Other versions of CAPM do not fare
much better. Allowing for habit persistence as in Backus, Gregory and Telmar (1993) or
allowing for first order risk aversion as in Bekaert, Hodrick and Marshall (1997)
increases variability of risk premium, but not by enough. Moreover, Froot and Frankel
(1989) used survey data to decompose excess return and found that the component due to
risk premium was not large. A bigger challenge that risk premium explanations face is to
explain the dependence of betas on sample period used, for example, the pattern shown in
Figure 4 below.
Two sources of forecast errors have been identified in the literature: irrational
expectations and rational systematic errors. Irrational expectations may arise because of
presence of heterogeneous traders in the market [Carlson and Osler, 1999; Mark and Wu,
5 This explanation is used to justify the negative risk premia on US$.