外汇市场中的羊群效应

Herding in the foreign exchange forward premium, the real exchange rate, and the depreciation rate. a variety of reasons unrelated to how they perceive or respond to the consensus.For this reason, the paper considers a range of complementary reduced-form tests and extends the analysis to a panelsetting that explicitly allows for cross-sectional dependence across individual forecasters.We also We employ a dataset that spans a maximum period from March 1995 to December 2024, with anaverage of 40–50 forecasters per currency. This rich dataset provides a wider range of currencies, alonger time period, and a more diverse cross section of forecasters than is typically employed in theexisting literature on FX herding, offering a robust foundation for analysis. Our findings do not point Finally, despite the common view that uncertainty may encourage market participants to look moreECB Working Paper Series No 3243 the factors that may lie behind these behaviours, is useful for policymakers, market participants, and al. (2006), which complements one of the regression-based tests. The approach uses both individual forecaster-level regressions and panel mean group estimation, allowing for cross-sectional dependence. We take advantage of an extensive dataset collected from the Eikon Refinitiv platform, consistingof individual market participants’ nominal exchange rate forecasts for 1-month, 3-month, 6-month, of expectations formation, we gain some understanding of the various influences on expectationsformation and, in turn, on herding-type behaviour.We examine a range of factors largely basedon uncertainty, volatility, and well-known FX predictor variables.The motivation for doing so istwofold.First, theories of herding emphasise that when the environment is uncertain, agents mayplace greater weight on the behaviour or beliefs of others, either because they believe others possessECB Working Paper Series No 3243 ature on herding in exchange rate forecasts. The applications in Gallo, Granger and Jeon (2002) andClements (2018) are to macroeconomic prediction datasets, whereas our application is to a large panelof individual exchange rate forecasts.As in Clements (2018), and in a number of papers analysingherding in FX markets, such as Fritsche et al. (2015) and Frenkel et al. (2020), we also compute theBernhardt et al. (2006) statistic. Our analysis extends the literature in three ways. First, it exam-ines the regression-based herding tests in a panel context using mean group estimation that allows rejected, the evidence more often points to anti-herding. This finding is reinforced by the Bernhardt etECB Working Paper Series No 3243 of the tests using both individual forecaster regressions and mean group-based estimation, togetherwith an interpretation of the findings. Section 6 investigates the effects of alternative explanations, exceptions of NOK and SEK, which average 25.The variability overtime is reasonably high, withstandard deviations of the number of participants mostly in the range 6-11, but is up to twice as large for the major currencies, GBP, EUR and JPY. For the six currency pairs, AUD, CAD, CHF, EUR,GBP, and JPY, the maximum number of forecasters contributing lies within the range 58-70 . For the three remaining currency pairs, the number of forecasters available ranges from 38 to 54.In Panel B we provide summary statistics regarding the number of forecast observations.For section at timet. We then compute the average over time and forecast horizons, of the four summarystatistics given in the columns. Individual forecasters for the main traded currencies, the GBP, EUR,and JPY, have an average number of observations of 133, 129, and 135, respectively.Hence, the over two-thirds. For the AUD, CAD, NZD and CHF we observe an average of 108, 110, 84 and 107observations, respectively. For the NOK and SEK currencies, the average number of observations ismuch lower at 44, since the sample begins in 2013.The variability among individual forecasters is1The sample period for each currency, noting that the number of forecasters differs across periods, is as follows: consultancies (e.g., Oxford Economics, EIU) and research institutes (such as WIIW, NIESR). Contributors to FX Pollfrom Refinitiv/Reuters include economists, buy/sell-side research analysts, strategists, and research think-tanks.ECB Working Paper Series No 3243 but also can be very low with some forecasters only appearing once in the sample (the subsequent ECB Working Paper Series No 3243 properties of the difference between aggregate and individual forecasts, over currency and forecast square error (RMSE) ratios, relative to a random walk, examining forecast accuracy. We define consensus and individual forecast errors, respectively, as follows:et+h=st+h−¯st+h|t,(1a)ej,t+h=st+h−sjt+h|t,(1b)wherest+h=log(St+h)andSt+his the nominal spot exchange rate for each currency in periodt+h(where for ease of notation we suppressi,i= 1,2, ...., N,