briefly, first by focusing on what we call as “traditional/text-book” reasons that undermines
the UIP condition. Thereafter, we present studies benefiting from recent methodological
advances that may shed light on the highly stylized forward premium bias.
(i) Rational expectations assumption does not hold, i.e. st+k = E[st+k|It] + ηt+k where
the forecast error ηt+k depends on the information available at time t, which yields excess
returns even when agents are risk-neutral. In particular, under such non-rational expecta-
tions, we have the following UIP regression:
∆kst+k = β0 + β1 (ftk - st) + ηt+k + ut+k
where ut+k is a white-noise disturbance term. Both the experimental (see Marey, 2004a-b,
and references therein) and survey evidence (see MacDonald, 2000; and Pesaran and Weale,
2006; for a review of this literature) on the formation of market expectations of the exchange
rate reveals that market participants form their expectations in a non-rational way.
(4)
(ii) Risk neutrality assumption does not hold and risk-averse investors demand a pre-
mium for holding assets that are perceived to be risky. Accordingly, defining the risk
premium as ρt = ftk - ste+k within the context of equation (3b), the risk-premium adjusted
UIP condition can be stated as
∆ek st+k = β0 + β1 ftk - st - ρt + ut+k
where ut+k is a white-noise error term. Since tests for the UIP condition involve a joint
hypothesis that comprises of rational expectations and risk-neutrality, one can use survey-
based expected depreciation data to isolate the effect of risk neutrality. In this vein, for
instance, MacDonald (2000), and Chinn and Frankel (2002) rely on survey-based data on
the bilateral US dollar exchange rate for different forecast horizons, and document that
the UIP slope coefficient is significantly different from one, justifying the presence of a
time-varying risk premium.
(5)
Next, we consider the literature on the “not-so-traditional” reasons for the observed
deviations from the UIP condition. As aforementioned, a more complete picture can be