Do central banks lose on foreign-exchange intervention? A review article

Do central banks lose on foreign-exchange intervention? A review article

Journal of Banking & Finance 21 (1997) 1667±1684 Do central banks lose on foreign-exchange intervention? A review article Richard J. Sweeney 1 Scho...

140KB Sizes 1 Downloads 18 Views

Journal of Banking & Finance 21 (1997) 1667±1684

Do central banks lose on foreign-exchange intervention? A review article Richard J. Sweeney

1

School of Business Administration, Georgetown University, 37th and ``O'' Sts., NW, Washington, DC 20057, USA

Abstract Estimates of central bank intervention losses or pro®ts vary widely; some estimates ®nd substantial losses, others pro®ts. In most cases, estimated pro®ts are not risk-adjusted, and risk adjustment can have large e€ects. Furthermore, pro®t estimates involve variables integrated of order one, and because of this test-statistics may have nonstandard distributions; few studies take this into account. Estimates of risk-adjusted pro®ts for the US Fed and the Swedish Riksbank, with allowances for possible nonstandard distributions, suggest that neither made losses and might have made signi®cant profits. Ó 1998 Elsevier Science B.V. All rights reserved. JEL classi®cation: F31; F33; G14; G15 Keywords: Central bank intervention; Intervention losses; Foreign exchange speculation

1. Introduction Friedman (1953) suggests evaluating a central bank's foreign-exchange intervention by examining its pro®ts from intervention±central banks making losses from intervention are likely destabilizing the foreign-exchange market. Many authors have responded to Friedman's views. This paper surveys this

1

Tel.: +1 202 687 3742; fax: +1 202 687 4031; e-mail: [email protected]

0378-4266/97/$17.00 Ó 1997 Elsevier Science B.V. All rights reserved. PII S 0 3 7 8 - 4 2 6 6 ( 9 7 ) 0 0 0 3 3 - 2

1668

R.J. Sweeney / Journal of Banking & Finance 21 (1997) 1667±1684

literature with a view to evaluating where we stand now and where it might be useful to go in the future. The argument that pro®ts imply that intervention is stabilizing, and losses that it is destabilizing, is in doubt. Authors have produced theoretical examples of pro®table speculation that is destabilizing (Baumol, 1957; Kemp, 1963; Johnson, 1976; Hart and Kreps, 1986; Szpiro, 1994). Further, if intervention has no e€ect on exchange rates, as many argue, a central bank may (lose) make money on its intervention without (de)stabilizing the market (Leahy, 1989, 1995; Edison, 1993; Dominguez and Frankel, 1993). The issue has become whether central banks lose money from intervention, not whether such losses imply destabilization (Mayer and Taguchi, 1983, suggest other approaches to whether intervention is stabilizing). In one view, central bank intervention should make expected positive profits: The central bank has an information advantage over other market participants; it also intervenes to straighten out destabilizing behavior such as ``disorderly markets.'' In another view, government interventions in markets are generally misguided and costly; the central bank should lose money by intervening in exchange markets dominated by intelligent, stabilizing speculators. In a third view, the exchange market should be ecient in the sense of the Ef®cient Markets Hypothesis (EMH), and may well be strong-form ecient relative to central bank intervention, implying expected intervention losses of zero. Which view most accurately characterizes exchange markets is then an empirical issue. Taylor (1982a) argues that estimated central bank intervention losses are economically and statistically signi®cant (Schwartz, 1994, ®nds US intervention losses). Many authors attack Taylor's work on various grounds. Estimated losses vary importantly, sometimes dramatically, across time periods (Wilson, 1982; Jacobsen, 1983). Taylor (1982a) does not include interest-rate di€erentials in his calculations of losses (but see his footnote 4, and Taylor, 1982b, 58±60, where both include interest earnings), and their inclusion is sometimes important for results (Bank of England, 1983; Leahy, 1989, 1995; Fase and Huijser, 1989). Some authors use time periods selected to make accumulated net intervention equal to zero at the end of the period (Argy, 1982; Bank of England, 1983; Jacobsen, 1983). Corrado and Taylor (1986) argue that this strategy underestimates central bank losses; they base their argument on a case where intervention has temporary e€ects on the exchange rate, and more generally argue that appreciation and intervention should be viewed as possibly interdependent. Most authors do not adjust central bank intervention losses to re¯ect risk premia earned from taking exposed foreign-exchange positions. Further, few use the implications of the EMH in designing measures of pro®ts and tests of the signi®cance of these measured pro®ts, or draw the conclusions for exchange-market eciency of the estimates and signi®cance of central bank intervention pro®ts (to be sure, some authors o€er exchange-market ineciency as a

R.J. Sweeney / Journal of Banking & Finance 21 (1997) 1667±1684

1669

possible explanation of positive measured pro®ts, for example, Leahy, 1989, 1995). Section 2 discusses measurement of risk-adjusted pro®ts/losses from central bank intervention. It shows why failure to adjust for risk can make results sensitive to whether interest-rate di€erentials are included, and to the time period over which pro®ts are estimated. It also discusses why the test-statistics for pro®ts may have nonstandard distributions, and how to handle this. Section 3 compares and discusses empirical estimates of pro®ts for various methods of adjusting for risk including not adjusting. Section 4 discusses ``hidden'' intervention, including forward market intervention. Section 5 brie¯y discusses the issue of whether central bank pro®ts/losses from intervention can explain estimated pro®ts that some authors ®nd from following mechanical trading rules. Section 6 o€ers a summary, some conclusions, and some issues for further research. 2. Measuring risk-adjusted intervention pro®ts In some instances, such as the European Monetary System crises of 1992 and 1993, individual central banks make pro®ts/losses that are obvious to outsiders, though the amounts are uncertain. For major countries, several years may go by without crises. Most studies of intervention losses in practice focus on these more usual periods. Central banks typically assert they make intervention pro®ts rather than losses. Most assert they have goals beyond pro®ts, for example, macroeconomic stabilization and interest-rate management (for example, Bank of England, 1983), with the implication that losses may be worthwhile in pursuit of other goals. Rather than sort this out, one approach supposes that a speculator has information on intervention and mimics intervention proportionately. The speculator's losses are a measure of central bank losses that would have to be justi®ed by goals beyond pro®ts. Estimating pro®ts. From ®nancial studies, a natural approach to measuring the pro®ts from one day's intervention over the window of the following T days is to calculate its abnormal return (AR) each day and to sum the daily abnormal returns to ®nd the cumulative abnormal return (CAR) over the window. (Estimating abnormal returns is discussed below.) For concreteness, take the Fed as the central bank. Let the Fed intervene on day 1 to buy an amount of DEM with the USD value of I1 . The ®rst abnormal return the Fed recieves on I1 is AR2 , from day 1 to day 2; the Fed recieves an abnormal return each day from day 2 to day T + 1. Pro®ts over the window from day 1 intervention are, I1

T X T ˆ1

ARt‡1 ˆ I1 CAR2;T ‡1 ;

1670

R.J. Sweeney / Journal of Banking & Finance 21 (1997) 1667±1684

where CAR2;T‡1 is cumulative abnormal returns for the DEM±USD exchange market. Pro®ts from intervention on days 2,T are found in the same way; total intervention pro®ts from day 1 to day T are then found by summing to give T T t X X X It CARFX:t‡1;T ‡1 ˆ It ARj‡1 : tˆ1

tˆ1

…1†

jˆ1

The formulation Taylor (1982a, b) is similar, though he uses raw rates of return (Rt ) rather than abnormal returns. Some authors measure ARt and Rt as arithmetic rates of return; Taylor (1982a), Leahy (1995) and Spencer (1985, 1989) use compound arithmetic rates of return; Sweeney (1996a) and SjoÈoÈ and Sweeney (1996a, b) use continuously compounded abnormal rates of return; Corrado and Taylor (1986) and Taylor (1989) use ®rst di€erences of the exchange rate. De®ne cumulative intervention in t as CIt ˆ

t X Ij ‡ CI0 ; jˆ1

where CI0 is the central bank's initial exposure from past intervention. Then pro®ts S are Sˆ

T T X X It CARt‡1;T ‡1 ˆ CIt ARt‡1 ; tˆ1

…2†

tˆ1

as in SjoÈoÈ and Sweeney (1996a, b) and Sweeney (1996a). Implications of the ecient markets hypothesis (EMH). If the foreign-exchange market is strong-form ecient relative to Fed intervention, then past It have no (exploitable) information for coming ARt‡j j > 0, or Et …ARt‡j j fIh gthˆ1 † ˆ 0 ˆ Et …ARt‡j j fCIh gthˆ1 †. It may react to current and past ARt under the EMH (or to current and past Rt ); this is an important consideration in forming test-statistics when variables are integrated of order 1 or higher. Testing estimated pro®ts. For US and Swedish data, in conventional tests It appears integrated of order 0, CIt integrated of order 1. Thus, S contains a variable integrated of order 1. Usual statistical theorems are for stationary variables and may well not apply if the variables are integrated. 2 Few papers take account of the integrated properties of CIt in testing estimated pro®ts; the reader may want to reserve judgement on reported signi®cance levels in such cases.

2 From (2), the formulation with It rather than CIt changes the order-1 variable from CIt to CARt‡1;T‡1 but does not remove the problem of dealing with an integrated variable.

R.J. Sweeney / Journal of Banking & Finance 21 (1997) 1667±1684

PT

1671

S divided by its standard error, r^S ˆ ‰^ r2AR tˆ1 …CIt ÿ CI0 †2 Š1=2 , gives the Stest value, S=^ rS . Sweeney (1996a) discusses the range of cases: (a) At one extreme, asymptotically S=^ rS has the standard N(0, 1) distribution, in which case a two-tailed test is appropriate; an example is strong-form eciency with the additional assumption that intervention not linearly related to current or past abnormal returns. (b) At the other extreme, S=^ rS is asymptotically distributed as the negative of the Dickey±Fuller t-value in a unit-root regression under the null of a unit root against the alternative of a root less than unity and an intercept that may be nonzero, here the negative sign arises from leaning-againstthe-wind intervention. Positive test values are to be expected 3 and a one-tailed test is appropriate. This case may arise under strong-form eciency if intervention is related to current or past abnormal returns and cumulative abnormal returns dominate the evolution of cumulative intervention (that is, if CARt and CIt . are cointegrated). (c) In other cases, asymptotically S=^ r is distributed as a weighted average of the two distributions, with one-tail critical values between those in cases (a) and (b). Simulation is needed to ®nd critical values in any given case. SjoÈoÈ and Sweeney (1996a) use simulation to ®nd critical values for di€erent absolute values of q, the correlation of ARt and the unsystematic part of It . For |q| 6 0.25, the one-tail 5% critical values range from 1.65 to approximately 2.0. A related issue is the strategy of choosing sample periods over which ending cumulative intervention is zero, a practice used in Bank of England (1983), Argy (1982) and Jacobsen (1983) and criticized by Corrado and Taylor (1986). In case (a) above, choice of starting and ending points, and hence setting CIT ˆ 0 makes no expected di€erence because ARt and the innovation in It‡j are linearly unrelated for all j, and CIT ˆ 0 has no information for the time path of ARt up to T. In cases (b) and (c), where the central bank reacts to current and past ARt in setting intervention, as Corrado and Taylor argue, CIT ˆ 0 has information about the time path to T of abnormal returns; showing how imposing CIT ˆ 0 a€ects test statistics requires taking explicit account of the fact that CIt is integrated of order 1, which Corrado and Taylor do not do. Risk adjustment. Few papers on intervention pro®ts consider or use risk adjustment. Inferences regarding intervention pro®ts are conditional on the risk model used to form abnormal returns. Among the methods of risk adjustment are: mean, market, market-model and arbitrage-pricing-model adjustment; this paper discusses mean and market-model adjustment. Under mean adjustment, if Rt‡1 is the rate of appreciation of the DEM relative to the USD from t to t + 1, and di€t is the interest-rate di€erential over

3 This follows because in a Dickey±Fuller unit-root test against the null of unity, a negative tstatistic is to be expected.

1672

R.J. Sweeney / Journal of Banking & Finance 21 (1997) 1667±1684

P that period, then ARt‡1 ˆ (Rt‡1 ) di€t ) ) (R0 ) di€0 ) and CAR2;T‡1 ˆ Ttˆ1 ARt‡1 , where R0 and di€0 are sample-period means. 4 Because sample-period PT means are used in forming ARt , tˆ1 ARt‡1 ˆ 0 ˆ AR0 , and Sˆ

T T X X CIt …ARt‡1 ÿ AR0 † ˆ …CIt ÿ CI0 †ARt‡1 tˆ1

tˆ1

ˆ …T ÿ 1†cov0 …CIt ; ARt‡1 †;

…3† 0

S is (T ) 1) times the sample covariance (cov ) of CIt and ARt‡1 . For mean adjustment, Table 1 compares the results of measuring pro®ts using Rt‡1 , Rt‡1 ) di€t , Rt‡1 ) R0 , or ARt‡1 . When only the raw return Rt‡1 is used, measured pro®ts exceed S by T R0 CI0 ; the bias in measured pro®ts, relative to S, will be positive or negative as sgn(R0 ) ˆ ,¹ sgn(CI0 ). If S  0, as expected under the EMH, going from one period with sgn(R0 ) ˆ sgn(CI0 ) to another with sgn(R0 ) ¹ sgn(CI0 ) reverses the sign of measured pro®ts. 5 Using Rt‡1 ) di€t , the bias is T (R0 ) di€0 ) CI0 . Inclusion of di€t can change the sign of measured pro®ts relative to those found when using Rt . Use of Rt‡1 ) R0 gives a bias of (T ) 1) cov0 (CIt , di€t ). If cov0 (CIt , di€t )  0 as often happens, the bias in measured pro®ts is close to zero. Thus, failing to adjust for risk can cause serious problems, but failing to adjust for the interest rate di€erential may not seriously a€ect measured risk-adjusted pro®ts. 6 When an S-statistic adjusted for risk by an augmented market model is used, the various returns measures unadjusted for risk have similar biases, mutatis mutandis. The method of risk adjustment may have important e€ects on estimated intervention pro®ts (see below). The usual market model is Rt ˆ a0 + b RMt + et : RMt is the rate of return on the market; beta, the slope b, measures systematic risk (in this case, market risk); and et is an error term orthogonal to RMt . Measured pro®ts found under mean adjustment may arise from association of intervention with movements of systematic (beta) risk. To allow for time-varying systematic risk, Sweeney

4 R0 and di€0 need not be calculated over the same period where the ARt are estimated. Because S is essentially a measure of central bank timing ability, SjoÈoÈ and Sweeney (1996b) argue for using demeaned values of ARt . 5 SjoÈoÈ and Sweeney (1996a) report on the consequences of not adjusting for risk. Over their sample, there was on average appreciation of the SEK relative to the USD, depreciation relative to the DEM. Thus if intervention was mainly in USD, R0 < 0; if mainly in DEM, R0 > 0. Further, Riksbank cumulative intervention generally trended up, or CI0 > 0. If S  0  cov0 (CIt , di€t ), then use of the raw net rate of appreciation instead of risk-adjusted net appreciation gives pro®ts of T R0 CI0 , positive for the DEM, negative for the USD. They ®nd results quite close to those predicted. 6 Sweeney (1996a), Table 2 reports that for periods when S is statistically and economically important, including the interest-rate di€erential or not has minor e€ects on measured pro®ts and their signi®cance.

T (R0 ) di€0 ) CI0

T R0 CI0

0

T (R0 ) di€0 ) CI0

0

T R0 CI0 + (T ) 1) cov0 (CIt , di€t ) zero

S+ (T ) 1) cov0 (CIt , di€t ) (T ) 1) cov0 (CIt , di€t )

Rt‡1 ) R0

zero

zero

S

ARt‡1

Notes: This table assumes mean adjustment for risk, with the mean calculated over the sample. cov indicates sample covariance. Variables: S: Riskadjusted intervention pro®ts; in this case, using abnormal returns found with mean adjustment; Rt‡1 : Rate of appreciation of the DEM relative to the USD, from t to t + 1; di€t : Interest rate di€erential at time t; RN;t‡1 : Net rate of appreciation, Rt‡1 ) di€t ; R0 : Sample mean of Rt , average appreciation rate of the DEM; di€0 : Sample mean of di€t , average DEM interest-rate di€erential; ARt‡1 : Mean-adjusted abnormal return from t to t + 1, Rt‡1 ) di€t ) (R0 ) di€0 ).

0

Measured pro®ts if S ˆ 0 Measured pro®ts if S ˆ 0 ˆ cov(CIt , di€t )

Rt‡1 ) di€t S + T (R ) di€ ) CI

0

S + T R CI + (T ) 1) cov (CIt , di€t )

0

Measured pro®ts

0

Rt‡1

Measure of abnormal return

Table 1 Measured pro®ts under alternative measures of abnormal returns

R.J. Sweeney / Journal of Banking & Finance 21 (1997) 1667±1684 1673

1674

R.J. Sweeney / Journal of Banking & Finance 21 (1997) 1667±1684

(1996a) makes beta a linear function of some Fed intervention variable IVt . In a simple case, b ˆ b0 + b1 IVt , giving an augmented market model 7 Rt‡1 ˆ a0 ‡ b0 RMt‡1 ‡ b1 IVt RMt‡1 ‡ et‡1 :

…4†

CIt is used for IVt because explaining pro®ts as due to time-varying risk requires that correlation of CIt and Rt‡1 arise from correlation of CIt with [b0 RMt‡1 + b1 (RMt‡1 IVt )]. 8 Sweeney (1996a) reports economically and statistically important estimates of b1 > 0. 3. Estimates of central bank intervention pro®ts/losses Table 2, Panel A, reproduces the Taylor (1982a) estimates of Fed intervention losses, using the raw rate of return adjusted for neither risk nor the interest-rate di€erential; he uses monthly data on reserve changes to measure intervention. The consistent pattern of losses, and the economic and statistical signi®cance of many of these estimates are apparent. Because Taylor's test-statistics do not take account of the fact that CIt is integrated of order 1, Section 2 suggests the reader may want to reserve judgment on signi®cance. Many authors report positive pro®ts when their estimates include the interest-rate di€erential to give net appreciation, and frequently attribute the di€erence to inclusion of the di€erential. Section 2 shows how estimates that are not risk-adjusted can easily change sign depending on whether the di€erential is included or not; Taylor (1982a), see also 1982b, 58±60 argues that inclusion of interest rates does not substantially a€ect the sum of intervention losses across countries. Critics of Taylor's results report that estimates are sensitive to the time periods used; Taylor (1982a) shows (in his Table 2) and discusses this sort of instability. This type of instability may be expected: A certain amount of random variation is to be expected; Section 2 shows how failure to use risk-adjusted pro®ts may easily lead to instability in results; and see below for comparing whole-period and yearly results. Table 2, Panel B, shows the Leahy (1995) pro®t estimates that include interest rates, but are not risk-adjusted, for Fed daily intervention in both DEM and JPY. His estimates are for a longer period that Taylor's (March 1973±December 1992 versus April 1973±January 1980), and use di€erent data with different frequency (daily intervention versus monthly changes in international reserves, daily versus monthly average exchange rates). Taylor's estimated profits for the Fed are USD ) 2351 m, Leahy's USD 16573 m, a startling and eco-

7 8

Sweeney (1996a) discusses more general augmented market models. The risk explanation of the positive mean-adjusted S-statistics in Table 3 requires b1 > 0.

R.J. Sweeney / Journal of Banking & Finance 21 (1997) 1667±1684

1675

Table 2 Estimates of pro®ts/losses from intervention Country

Period Begin.

Period End

Pro®ts (+) (USD millions)

Panel A. Taylor (1982a): Estimates of losses from ocial intervention a Canada June '70 Dec. '79 )82 France Apr. '73 Dec. '79 )2,003 Germany Apr. '73 Dec. '79 )3,423 Italy Mar. '73 Dec. '79 )3,724 Japan Mar. '73 Dec. '79 )331 Spain Feb. '74 Dec. '79 )1,367 Switzerland Feb. '73 Dec. '79 )1,209 United Kingdom July '72 Dec. '79 )2,147 United States Apr. '73 Jan. '80 )2,351 DEM

t-statistic

)0.20 . . .b )0.70 )3.86 )0.14 )3.58 )0.26 )1.93 . . .c

JPY

Panel B. Leahy (1995): Estimates of pro®ts from US intervention, March 1973±December 1992 Pro®ts 12,329 4,244 (millions of USD) Proceeds of purchases/sales

Interest proceeds

a

Total proceeds

Panel C. Fase and Huijser (1989): Estimates of pro®ts from Nederlandsche Bank intervention, 1974±1989 a (millions of guilders) 370 1324 1693 Sources: Panel A ± Taylor (1982a); Panel B ± Leahy (1995); Panel C ± Fase and Huijser (1989). a Taylor takes ocial intervention as monthly changes in international reserves. Taylor corrects of®cial reserves series by including hidden intervention through, for example, state-controlled companies. He does not include interest rates in the calculations he presents ± but see his footnote 4, and also Taylor, 1982b, pp. 58±60. Leahy uses Fed daily intervention data and includes interest rates in his calculations. Fase and Huijser use monthly intervention data in their calculations. None of these studies adjusts for risk. b Assumes intervention is done in DEM. c No t-statistic is shown ``since the data cannot be put on a comparable basis. . .'' (Taylor, 1982a, Table 1).

nomically important di€erence. Leahy does not give statistical measures of the signi®cance of these pro®ts; however, he performs three types of tests of whether cumulative intervention today is useful in forecasting net appreciation from today to tomorrow, though he does not account for the statistical problems introduced because cumulative intervention is integrated of order 1. 9

9 One of the Leahy (1995) tests of market eciency is a regression of net appreciation on yesterday's cumulative intervention. This is equivalent to an S-test under mean adjustment (Appendix A to Sweeney, 1996a), but the OLS t-value is generally not distributed N(0, 1), as Section 2 argues.

1676

R.J. Sweeney / Journal of Banking & Finance 21 (1997) 1667±1684

Fase and Huijser (1989) report estimated pro®ts for the Bank of the Netherlands from January 1974±June 1989 for monthly intervention data, including the e€ects of interest rates, but without risk adjustment (Table 2, Panel C). For spot-market intervention they separately report the pro®ts from currency appreciation and from di€erential interest earnings; across the Bank's intervention currencies, the totals are 370 m and 1324 m guilders, for total pro®ts of 1693 m guilders. In this case, inclusion of interest rates has a major e€ect, as can happen when estimates are not risk-adjusted (Section 2). The authors present statistical tests based on the Taylor (1982a) method and hence their test results are subject to the same reservations. 10 Table 3 shows some results on Fed and Riksbank pro®ts/losses from daily intervention where risk-adjusted abnormal returns are used. During the 1985±1991 period the Sweeney (1996a) examines, the USD ¯oated with intervention on only about 10% of the days for each currency. During the 1986± 1990 period SjoÈoÈ and Sweeney (1996a) examine, the Riksbank kept the value of the SEK in terms of a basket of currencies within a band around a preannounced central target, and intervened on almost 60% of days. Thus, results for the Fed and Riksbank are for quite di€erent exchange-rate management strategies. Panel A shows results from Sweeney (1996a) for the Fed for mean adjustment. Pro®ts are measured over the whole period, and by using the sum-ofthe-years pro®ts; these measures di€er because they use risk measures estimated over di€erent periods and because in the former the previous year(s)'s bets are allowed to ride and in the latter the investor starts each year with zero exposure (compare Taylor, 1982a). For both currencies, for both types of risk-adjustment, and for both the whole-period and sum-of-the-years's measures, pro®ts are positive and in some cases signi®cant at conventional significance levels. For both currencies and both measures, pro®ts fall in going from mean to market-model adjustment (not shown). In going from the whole-period to the sum-of-the-years' measure, pro®ts rise for the DEM, fall for the JPY. Panel B, shows comparable results from SjoÈoÈ and Sweeney (1996a) for the Riksbank for 1986±1990. The Riksbank did not reveal the intervention currency, the USD or the DEM, it used on a given day (authoritative observers report  and that over this period, there was very little intervention using DEM). Sj oo Sweeney present results under the assumption that intervention is all in USD, all in DEM, or in a 50±50 ratio. Further, for comparison, they suppose that the investor takes intervention as a signal to trade a basket of currencies, with all currencies in the same proportion as the basket the Riksbank targets. In Panel B, pro®ts strongly tend to be positive ± only for the DEM for whole-period 10

For estimates of Canadian intervention pro®ts, see Murray et al. (1989).

a

Sum of years

b

2143.28 (1.84)

Whole period

3300.57 1132.65 2711.60 1122.72

(2.07) (2.79) (NR) (NR)

822.00 (1.51)

Sum of years

The whole period for the US is 1985±1991, for Sweden 1986±1990. Whole-period test-statistics are valid if either: the test-statistic is distributed N(0, 1); or is a weighted average of a normal distribution and (the negative of) a Dickey±Fuller unit-root test t-value of the estimated slope, on the assumption that the correlation of ARt and innovations in It have a correlation q such that |q|  0.25; for both the DEM and JPY, |q| is less than 0.20. b The ``sum of the years'' is the sum of the S-statistics for each separate year. The test statistic is valid if each year's S=^ rS is distributed N(0, 1). The distribution of the test statistic is unknown if each year's S=^ rS as distributed as a weighted average of a normal distribution and (the negative of) a Dickey±Fuller unit-root test t-value of the estimated slope. c The Fed intervenes in both the DEM and JPY and reports interventions separately for each. The Riksbank's two intervention currencies are the USD and the DEM, but it does not report the currency in which each day's interventions are made (or the split if both are used). The USD results suppose that the informed investor mimics Riksbank intervention but all in USD; the DEM results that the mimicing is all in DEM; the 50±50 results that the mimicing is half in USD, half in DEM; the basket results that the mimicing goes to the di€erent currencies in the basket around which the SEK ¯uctuates with the proportions equal to the weights of the currencies in the basket. NR: Not reported in SjoÈoÈ and Sweeney (1996a).    , , Signi®cant at the 10%, 5% and 1% levels.

a

Panel B. Risk-adjusted pro®ts from Riksbank intervention (SjoÈoÈ and Sweeney, 1996a), millions of SEK: S (S=^ rS ) 2800.04 (0.38) 400.05 (1.89) DEM 477.84 (0.10) Mean Adj. USD c 50±50 2002.20 (0.57) 2070.03 (1.90) Basket 951.78 (0.79) Mkt.-Mod. Adj. USD 8428.46 (1.23) 1828.00 (NR) DEM )407.67 ()0.09) 50±50 3874.73 (1.14) 2599.58 (NR) Basket 1202.47 (1.00)

Panel A. Risk-adjusted pro®ts from fed intervention (Sweeney, 1996a), millions of USD: S (S=^ rS ) 785.67 (0.44) 1805.92 (3.85) JPY Mean-Adj. DEM c

Whole period

Table 3 Risk-adjusted intervention pro®ts for the US and Sweden

R.J. Sweeney / Journal of Banking & Finance 21 (1997) 1667±1684 1677

1678

R.J. Sweeney / Journal of Banking & Finance 21 (1997) 1667±1684

measurement and market-model adjustment are pro®ts negative (and losses are relatively small). In going from mean to market-model adjustment, pro®ts rise for some ``intervention'' currencies, fall for others. In going from whole-period to sum-of-the-years' measures, pro®ts rise for some, fall for others. For both Fed and Riksbank intervention, choice of risk-adjustment method (mean or market-model adjustment) and choice of whole-period or sum-of-theyears' measurement have important e€ects on estimated pro®ts. Confusingly, the e€ects di€er across various intervention currencies. For both Fed and Riksbank intervention, a strong case can be made that neither central bank makes losses. It is possible, but a harder case, to argue that the banks make pro®ts. Forward-market intervention. Some studies have also looked at the pro®tability of forward-market intervention. Fase and Huijser (1989) report losses on the Netherlands' central bank intervention in forward markets over the period 1984±mid-1989, using monthly data. Because there is no risk adjustment, these measured pro®ts cannot be interpreted. Beenstock and Dadashi (1986) investigate Canadian forward intervention and report pro®ts close to zero; Longworth (1980) argues that the dominant part of their forward data is overnight spot transactions rather than forward positions. 4. Disguised intervention Many central banks go to substantial lengths to make it dicult to tease intervention information out of published data. Taylor (1982a) 357, 358 argues that ``each country's monetary authorities direct other institutions, such as the treasury, an exchange stabilization fund, nationalized industries, and commercial banks in their intervention,'' and data is needed on all ``institutions that buy and sell foreign currency to in¯uence the exchange rate.'' He argues that in the 1970s, France, Italy, Japan, Spain and the UK used some of these methods to avoid reporting changes in reserves that truly re¯ect intervention, and that Canada, Germany and the US either report intervention losses only selectively or use gross rather than net reserves to disguise intervention. Taylor (1982b) documents the lengths he went to in building his corrected reserves series (see also Wilson, 1982). Unfortunately, some exchange-rate studies that use reserves take ®gures uncritically from ocial sources. 11

11 Leahy (1995) and SjoÈoÈ and Sweeney (1996b) note that use of monthly reserves can have other problems, among them that ocial borrowing of foreign currencies moves reserves but not intervention, for example, and that reserves do not capture intra-monthly exchange rate movements and intervention.

R.J. Sweeney / Journal of Banking & Finance 21 (1997) 1667±1684

1679

Neither the US nor Sweden makes regular, daily, quantitative reports on its central bank's intervention on the current day or the recent past; instead observers must wait some time to get data on actual intervention or data from which intervention can be inferred with noise. Knowledgeable observers say that neither country uses the devices Taylor detailed to hide actions that are equivalent to reported intervention. Many authors have used daily data on the purchases and sales of Canada's Exchange Stabilization Fund, sometimes to evaluate pro®ts, sometimes to judge whether such changes a€ect the exchange rate relative to the USD, starting as early as Pippenger and Phillips (1973). These data contain transactions beyond intervention and hence are dirty though not disguised intervention (see Murray et al., 1989). Forward-market intervention. The Swedish Riksbank intervenes in the forward market, mainly to hide part of its intervention from the market. The bank believes that the market might overreact if provided true intervention ®gures; in particular, relatively large out¯ows might feed speculation of further out¯ows and ultimately devaluation. The Riksbank says it has followed the principle that its total intervention, spot plus forward, must have the same sign as the spot intervention that can be estimated from its weekly releases on ocial capital ¯ows; this series gives an estimate of spot market interventions, but not interventions on the forward market. Riksbank forward interventions began in 1982 on a relatively small scale. The size grew somewhat 1986, becoming larger during 1987. In 1987 the Riksbank decided to publish its end-of-month forward position, with a three month delay. The announcement made headline news in Sweden. Many observers, even dealers, were unaware that the Riksbank washed its data in this way; this tactic is not mentioned in any article or textbook before 1987. The average total monthly intervention during 1986±1990 was 2318 million SEK, the average forward intervention was 435 million SEK; average forward intervention is perhaps 20% of total intervention. This mean is not very informative since there is no rule for the share of forward interventions; the Riksbank decides that share each day, based on its view of what the public might regard as large unexpected capital ¯ows. The Riksbank's open forward market position was relatively constant from mid-1993±mid-1994 (H assel and Norman, 1994), implying that over this period Riksbank forward intervention often served mainly to keep its open position constant over time. 5. Do speculators gain from central bank intervention losses? A number of authors report risk-adjusted pro®ts (net of transactions costs) from following a mechanical buy-and-sell rules in the foreign-exchange market

1680

R.J. Sweeney / Journal of Banking & Finance 21 (1997) 1667±1684

(for the USD relative to other currencies, Sweeney, 1986; Surajaras and Sweeney, 1992; Levich and Thomas, 1993, and others; Lee and Mathur, 1996 ®nd little pro®t from cross rates). If these potential net pro®ts are taken seriously, one possible source is from beating intervening central banks, a possibility that a number of authors raise. Szakmary and Mathur (1996) explore this possibility using monthly data on reserve changes and conclude that potential speculator pro®ts are explained by speculators taking positions opposite the Fed's. LeBaron (1996) examines this issue on Fed daily intervention data, for a speculator using a 250-day moving average trading rule. He ®nds that removing from the sample those days when the Fed was intervening causes a serious reduction in pro®ts but does not eliminate them. Sweeney (1996b) also uses Fed daily intervention data, but for a 25-day moving average, as suggested by Surajaras and Sweeney (1992). He tests whether speculator pro®ts are linearly a€ected by the Fed's intervention, measured in various ways, and ®nds no signi®cant e€ect. 6. Summary and conclusions The issue of whether central banks make pro®ts or losses on their intervention can have serious e€ects on observers' views of the wisdom of intervention. Losses need not imply that the intervention destabilizes the market, but may incline some observers to that view. Further, central banks making losses may have an awkward time justifying continued intervention. Dean Taylor (1982a, b) reports estimated losses from intervention for each of nine central banks, including the Fed. Critics note that his results are sensitive to starting and ending points for the sample, and to whether the rates of appreciation are net of the interest-rate di€erential. Leahy (1989, 1995) and Fase and Huijser (1989) use appreciation net of the interest-rate di€erential and report positive pro®ts from Fed and Nederlandsche Bank intervention, respectively. These studies do not adjust for risk. Failure to adjust for risk creates possibly serious biases in estimated pro®ts. Under mean adjustment in the case where the interest-rate di€erential is omitted, the per-day bias is (CI0 R0 ), where R0 is the sample average rate of foreigncurrency appreciation and CI0 the sample average of cumulative intervention. If the di€erential is included, the bias is [CI0 (R0 ) di€0 )], where di€0 is the sample average of the di€erential. Clearly, the bias can change importantly with inclusion of the di€erential or with change in time periods such that either CI0 or R0 ± or (R0 ) di€0 ) ± changes sign, and these changes need have nothing to do with risk-adjusted intervention pro®ts. Studies that adjust for risk (Sweeney, 1996a and SjoÈoÈ and Sweeney, 1996a) ®nd results that do not support the view that central banks make

R.J. Sweeney / Journal of Banking & Finance 21 (1997) 1667±1684

1681

losses and might be argued to support pro®ts. Pro®ts would be evidence of strong-form ineciency relative to intervention; strong-form ineciency would not cast doubt by itself on semi-strong form or weak-form eciency of the foreign-exchange markets. Their results allow for possible association of predetermined intervention with market betas and hence with time-varying systematic risk. Measures of intervention pro®ts over time include a variable integrated of order 1, either cumulative intervention or cumulated rates of return. Many standard statistical theorems, valid if the variables are stationary, are possibly invalid if a variable is integrated of order 1. The distribution of test-statistics for the statistical signi®cance of estimated pro®ts depends on whether and how the central bank intervention reaction function depends on the (abnormal) rate of return. The ecient market hypothesis requires in strong-form that the central bank cannot react to (unforecastable) coming abnormal returns, but of course it can react to current and past returns. If the reaction is negligible, the test statistic has the standard N(0, 1) distribution; if the in¯uence is dominant, the distribution is nonstandard and of a type familiar from Dickey±Fuller ttests of unit roots. In practice, the distribution appears to be an average of the two extremes, so simulation is needed to ®nd critical values. For the US and Sweden, abnormal returns appear to have an important but rather small e€ect on intervention. Some conclusions are clear. Studies should use intervention rather than changes in reserves, and should use high frequency data rather than data observed less frequently, both for intervention and exchange rates. Studies should adjust returns for risk; the method of risk adjustment can make an important di€erence, not necessarily in the same direction across central banks and intervention currencies. Studies should take account of the distribution of test statistics in light of the fact that cumulative intervention (or cumulative abnormal returns) is likely to be integrated of order 1. There is much work to be done on government intervention reaction functions. Because of the interaction of pro®ts and the intervention function, ideally both should be estimated together. A major problem is lack of daily data on variables that intuitively seem important, for example, political pressures arising from currency movements that some observers see as ''unwarranted'' by fundamentals. In Dominguez and Frankel (1993) in estimated daily reaction functions, intervention typically depends on current and lagged exchange rates and lagged intervention, with no allowance for the episodic nature of much intervention (but see Lewis, 1995); for example, the Fed intervenes on roughly 10% the Riksbank on roughly 60% of trading days. It is important to ®nd estimates of the pro®tability of intervention when more sophisticated functions with better explanatory ability are used. Of course better intervention functions are desirable in their own right and also for estimating the e€ect of intervention on exchange markets.

1682

R.J. Sweeney / Journal of Banking & Finance 21 (1997) 1667±1684

Detailed studies of how di€ering exchange-rate regimes a€ect pro®tability would be desirable. Regimes vary widely across countries and for a given country can vary widely over time. Such studies ought to account for the insights of the target-zone literature, for example, for some countries. Much work, theoretical and empirical, is needed on modelling the association of intervention with systematic risk. How does this association vary across countries? Is it stable over time? Is a multi-factor model superior to the market model? Desirable research depends in great part of availability of intervention data and exchange-rate data ± the cleaner and more detailed the data the better. Central banks o€er various defenses for not releasing their data. Fair or not, the inevitable question is, What are they hiding? Acknowledgements  for many long, helpful discussions. Michael Thanks are due to Boo Sj oo Leahy, Ike Mathur and Dean Taylor kindly commented on this paper. Georgetown University and the Georgetown Business School provided summer research and sabbatical support. The major part of this paper was written at the Gothenburg School of Economics, Sweden. References Argy, V.E., 1982. Exchange Rate Management in Theory and in Practice. Princeton Studies in International Finance, vol. no. 50. Princeton University, Princeton, New Jersey. Bank of England, 1983. Intervention, stabilization and pro®ts. Quarterly Bulletin 23, 384±391. Baumol, W., 1957. Speculation, pro®tability, and stability. Review of Economics and Statistics 39, 263±271. Beenstock, M., Dadashi, S., 1986. The pro®tability of forward currency speculation by central banks. European Economic Review 30, 449±456. Corrado, C.J., Taylor, D., 1986. The cost of a central bank leaning against a random walk. Journal of International Money and Finance 5, 303±314. Dominguez, K.M., Frankel, J.A., 1993. Does Foreign Exchange Intervention Work? Institute for International Economics, Washington, DC . Edison, H., 1993. The e€ectiveness of central bank intervention, Princeton studies in international ®nance. Special Papers in International Finance, vol no. 18., Princeton University, Princeton, New Jersey . Fase, M.M.G., Huijser, A.P., 1989. The pro®tability of ocial foreign exchange intervention: A case study for the Netherlands 1974±1989, Research Memorandum WO&E nr 8821, De Nederlandsche bank. Friedman, M., 1953., The case for ¯exible exchange rates, essays in positive economics, University of Chicago Press, Chicago.

R.J. Sweeney / Journal of Banking & Finance 21 (1997) 1667±1684

1683

Hart, O.D., Kreps, D.M., 1986. Price destabilizing speculation, Journal of Political Economy 94, 927±952. Jacobsen, L.R., 1983. Calculations of pro®tability for US dollar±deutsche mark intervention. Sta€ Series No. 131, Board of Governors of the Federal Reserve System, Washington, DC . Johnson, H.G., 1976. Destabilizing speculation: A general equilibrium approach. Journal of Political Economy 84, 101±108. LeBaron, B., 1996. Technical trading rule pro®tability and foreign exchange intervention. Working Paper, Department of Economics, University of Wisconsin, Madison, Wisconsin. Leahy, M., 1989. The pro®tability of US intervention. Board of Governors of the Federal Reserve System, International Finance Discussion Paper No. 343, Washington, DC. Leahy, M., 1995. The pro®tability of US intervention in the foreign-exchange markets. Journal of International Money and Finance 14, 823±844. Lee, C.I., Mathur, I., 1996. Trading rule pro®ts in European currency spot cross-rates. Journal of Banking and Finance 20, 949±962. Levich, R.M., Thomas, L.R., 1993. The signi®cance of technical trading-rule pro®ts in the foreign exchange market: A bootstrap approach. Journal of International Money and Finance 12, 451± 474. Lewis, K., 1995. Occasional intervention to target rates. American Economic Review 85, 691±715. Longworth, D., 1980. Canadian intervention in the foregin exchange market: A note. Review of Economics and Statistics 62, 284±287. Mayer, H.W., Taguchi, H., March 1983. Ocial intervention in the exchange markets: Stabilizing or destabilizing? BIS Economic Papers, No. 6. Murray, J., Zelmer, M., Shane, W., 1989. Measuring the Pro®tability and E€ectiveness of Foreign Exchange Market Intervention: Some Canadian Evidence, Bank of Canada, Ottawa. Pippenger, J.E., Phillips, L., 1973. Stabilization of the Canadian dollar: 1952±1960. Econometrica 41, 797±815. Schwartz, A., 1994.Correspondence with author . , B., Sweeney, R.J., 1996a. Foreign exchange intervention: New methods and results on the Sj oo pro®tability of central bank intervention. Working paper, Gothenburg School of Economics, Gothenburg, Sweden. , B., Sweeney, R.J., 1996b. The pro®tability of foreign-exchange intervention: Evidence from Sj oo Sweden. Working paper, Gothenburg School of Economics, Gothenburg, Sweden. Spencer, P.D., 1985. Ocial intervention in the foreign exchange market. Journal of Political Economy 93, 1019±1024. Spencer, P.D., 1989. How to make the central bank look good: A reply. Journal of Political Economy 97, 233±235. Surajaras, P., Sweeney, R.J., 1992, Pro®t-making speculation in foreign exchange markets, Westview Press, Boulder, CO. Sweeney, R.J., 1986. Beating the foreign exchange market. Journal of Finance 41, 63±82. Sweeney, R.J., 1996a. Does the foreign-exchange market beat the Fed? Working paper, Georgetown School of Business, Georgetown University, Washington, DC. Sweeney, R.J., 1996b. Do foreign-exchange speculators beat the Fed? Working paper, Georgetown School of Business, Georgetown University, Washington, DC. Szakmary, A.C., Mathur, I., 1996. Central bank intervention and trading rule pro®ts in foreign exchange markets. Carbondale, Ill: School of Business, University of Southern Illinois, . Szpiro, G.G., 1994. Exchange rate speculation and chaos-inducing intervention. Journal of Economic Behavior and Organization 24, 363±368. Taylor, D., 1982a. Ocial intervention in the foreign exchange market, or bet against the central bank. Journal of Political Economy 90, 356±368.

1684

R.J. Sweeney / Journal of Banking & Finance 21 (1997) 1667±1684

Taylor, D., 1982b, The mismanaged ¯oat: Ocial intervention by the industrialized countries. In: M.B. Connolly (Ed.). The international monetary system: Choices for the future. Praeger, New York, pp. 49-84. Taylor, D., 1989. How to make the central bank look good. Journal of Political Economy 97, 226± 232. Wilson, J.F., 1982. Comments on Dean Taylor: The mismanaged ¯oat: Ocial intervention by the industrialized countries. In: M.B. Connolly (Ed.), The international monetary system: Choices for the future, Praeger, New York, pp. 297±306.