LEO. Svensson/ European Economic Review 46(2002)771-780 variables that enter the loss function. The corresponding"target levels"are I'and zero. The zero target level for the output gap corresponds to an output target equal to potential output. There is general agreement that inflation-targeting central banks do normally not have overambitious output targets, that is, exceeding potential output. Thus, discretionary optimization does not result in average inflation bias, counter to the case in the standard Kydland-Prescott-Barro-Gordon setup Since the inflation target is subject to choice but not the output target, there is an asymmetry between the inflation and output target, consistent with the inflation target being the "primary objective". 3 Regarding the two parameters, 8 and 4, the discount factor is for all practical pur- poses likely to be very close to one, especially whe model is use estingly, when the discount factor approaches one, the limit of the intertemporal loss function is the weighted sum of the unconditional variances of inflation and the output lim t= varn]+4. Var[x,] (when the unconditional mean of infation and the output gap equal the inflation target and zero, respectively, E[]= I' and ELx]=0).As mentioned, flexible inflation targeting corresponds to i>0.Strict"inflation targeting would be the unrealistic case of i=0 2.1. Commitment to a simple instrument rule How do and should inflation-targeting central banks achieve the inflation target and minimize the loss function? most of the literature discusses this in terms of a com- mitment to a simple reaction function, a simple"instrument rule", where the central bank mechanically sets its instrument rate (usually a short interest rate like a one- or two-week repurchase rate), i,, as a given simple function of a small subset of the nformation available to the central bank. Although several different simple instrument rules have been discussed since the 1970s the best known and most discussed is the Taylor(1993)rule, a frequent variant of which can be written i=(1-f1)[F+元1+fn(π1-π*)+fxx]+f-1, where the response coefficients fx, fx and fi fulfill fx>0,fr>0 and0<fi<1 (although cases with fi l have also been discussed) and r is the average real interest rate. A large volume of research, for instance in Taylor (1999), has examined the papers have also estimated empirical reaction functions of this type terminacy properties of (2.3)and its variants in different models, with respect to determinacy of equilibria, performance measured by (2.2 ), robustness to different models, etc. Several The advantages of a simple instrument rule like(2.3)are:(1) The rule can easily be verified by outside observers and a commitment to the rule would therefore be techni- cally feasible.(2)Variants of the Taylor rule have been found to be relatively robust 3 An interesting and important research area, discussed in my presentation at the EEA 2001 Annual Congress, concerns to what extent inflation targeting as represented by (2.1)corresponds to maximizing the welfare of the representative consumer. Another research area is to what extent a quadratic loss function is sufficient, or if higher-order terms corresponding to asymmetric preferences are needed.774 L.E.O. Svensson / European Economic Review 46 (2002) 771 – 780 variables that enter the loss function. The corresponding “target levels” are ∗ and zero. The zero target level for the output gap corresponds to an output target equal to potential output. There is general agreement that in ation-targeting central banks do normally not have overambitious output targets, that is, exceeding potential output. Thus, discretionary optimization does not result in average in ation bias, counter to the case in the standard Kydland–Prescott–Barro–Gordon setup. Since the in ation target is subject to choice but not the output target, there is an asymmetry between the in ation and output target, consistent with the in ation target being the “primary objective”. 3 Regarding the two parameters,  and , the discount factor is for all practical pur￾poses likely to be very close to one, especially when a quarterly model is used. Inter￾estingly, when the discount factor approaches one, the limit of the intertemporal loss function is the weighted sum of the unconditional variances of in ation and the output gap: lim →1 Lt = Var[t] +  Var[xt] (2.2) (when the unconditional mean of in ation and the output gap equal the in ation target and zero, respectively; E[t] = ∗ and E[xt] = 0). As mentioned, exible in ation targeting corresponds to  ¿0. “Strict” in ation targeting would be the unrealistic case of  = 0. 2.1. Commitment to a simple instrument rule How do and should in ation-targeting central banks achieve the in ation target and minimize the loss function? Most of the literature discusses this in terms of a com￾mitment to a simple reaction function, a simple “instrument rule”, where the central bank mechanically sets its instrument rate (usually a short interest rate like a one￾or two-week repurchase rate), it, as a given simple function of a small subset of the information available to the central bank. Although several diNerent simple instrument rules have been discussed since the 1970s, the best known and most discussed is the Taylor (1993) rule, a frequent variant of which can be written it = (1 − fi)[ Rr + t + f(t − ∗) + fxxt] + fiit−1; (2.3) where the response coeJcients f; fx and fi ful@ll f ¿ 0; fx ¿ 0 and 0 6 fi 6 1 (although cases with fi ¿ 1 have also been discussed) and Rr is the average real interest rate. A large volume of research, for instance in Taylor (1999), has examined the properties of (2.3) and its variants in diNerent models, with respect to determinacy of equilibria, performance measured by (2.2), robustness to diNerent models, etc. Several papers have also estimated empirical reaction functions of this type. The advantages of a simple instrument rule like (2.3) are: (1) The rule can easily be veri@ed by outside observers and a commitment to the rule would therefore be techni￾cally feasible. (2) Variants of the Taylor rule have been found to be relatively robust 3 An interesting and important research area, discussed in my presentation at the EEA 2001 Annual Congress, concerns to what extent in ation targeting as represented by (2.1) corresponds to maximizing the welfare of the representative consumer. Another research area is to what extent a quadratic loss function is suJcient, or if higher-order terms corresponding to asymmetric preferences are needed