A. Channels of Influence

The vast literature on foreign exchange intervention focuses on three main channels of influence: signaling, portfolio balancing, and market microstructure.

• Intervention can be effective through the signaling channel if it is perceived as a credible signal on the future stance of monetary policy. To the extent that intervention, even when sterilized, influences expectations on future money supply, then it can influence the exchange rate.

• According to the portfolio balance channel, domestic and foreign currency denominated bonds assets are imperfect substitutes (and therefore, the “riskier” bond pays a risk premium) and intervention can be effective by modifying the currency composition of agents’ asset portfolios.⁵ Sterilized intervention alters the relative supply of domestic versus foreign currency bonds, leading agents to rebalance their portfolios to equalize risk-adjusted returns, which causes a change in the exchange rate.

• The microstructure approach emphasizes the effects of order flow, market participants, information asymmetries, and price discovery in the foreign exchange market. Central bank trades are assumed to emit information to the market, which modifies exchange rate expectations and ignites a tide of foreign exchange orders, magnified in part by trend-chasing traders (Lyons, 2001). Intervention-induced order flow, in turn, tends to increase short-term exchange rate volatility.

B. Analytical Techniques

A wide array of analytical techniques have been used to evaluate the effectiveness of intervention.⁶ Analyses have examined intervention’s impact on the exchange rate level using OLS regressions of the mean and risk premium equations or through event studies of intervention episodes (Appendix Table 1). Intervention’s impact on exchange rate volatility has been gauged through various forms of generalized autoregressive conditional heteroskedasticity (GARCH) models.

Regression analyses all suffer from simultaneity problems. In particular, the regression of exchange rate changes over intervention fails to disentangle the degree to which intervention reacts to exchange rate movements rather than the other way around. As a result, the coefficient can assume the wrong sign or overstate the impact of intervention on the exchange rate.⁷ Another key shortcoming of existing techniques, including event studies, is the short time horizon—typically the same day—over which intervention’s effectiveness is analyzed.

More recently, attempts have been made to overcome the simultaneity and time horizon problems through a joint analysis of monetary and exchange rate operations. Using an identified vector autoregression framework (IVAR), Kim (2003) and Guimarães (2004), empirically analyze the impact of monetary and intervention operations on the exchange rate and the extent to which intervention occurs in reaction to exchange rate movements. The IVAR framework can estimate the longer-term effects of intervention through an analysis of cumulative impulse responses. Moreover, the use of monthly data eases the limitation imposed by the absence of daily data (though intervention data are still obtained through the aggregation of daily data). However, there are some disadvantages to the use of IVAR, including the limited degrees of freedom (small sample and too many parameters), the validity of the identifying restrictions, and the plausibility of the structural shocks (Guimarães, 2004).

Another strand of the literature on the effects of intervention examines the links between intervention and exchange rate volatility. The potential impact of intervention on volatility is worth studying since many central banks intervene to smooth volatility, even when they are not targeting a particular level of the exchange rate. In addition, exchange rate volatility has often been associated with economic crisis and may be a signal of lack of policy credibility, which gives rise to fear of floating (Calvo and Reinhart, 2002).⁸ Finally, volatility may have harmful effects on trade and capital flows, although evidence supporting this claim is weak (Rogoff, 1999).

The single most important impediment to empirical work on intervention’s effectiveness is the lack of publicly available data on daily intervention. Attempts to use proxies of intervention—e.g., the change in the stock of central bank reserves—have not worked. Neely (2001) has shown that even for G-7 countries, changes in reserves are a poor proxy for intervention: Correlation coefficients between foreign exchange intervention and reserve changes are usually less than 0.4. The use of such proxies in developing countries can be even more misleading since reserve changes may reflect, inter alia, withdrawals of funds from multilateral organizations, government debt repayments, receipts from state-owned companies, and inflows of foreign aid.

C. Empirical Evidence

Empirical studies on the effectiveness of central bank sterilized intervention have focused almost exclusively on advanced countries. The research bias towards advanced countries primarily reflects the availability of data and the depth and sophistication of their foreign exchange markets assumed in many models of intervention.

Empirical tests have found mixed evidence in favor of the signaling and portfolio balance channels. For example, Dominguez and Frankel (1993a) estimate the effect of intervention on contemporaneous exchange rate movements and on forecasts of future exchange rates. Using survey data to measure exchange rate expectations, they find a significant effect of intervention on market expectations, especially if interventions are announced and coordinated. They also show that secret interventions are largely ineffective. Obstfeld (1990) finds that portfolio balance effects are statistically significant, but small in size. The consensus in the literature until recently was that the portfolio effect gives a limited role for intervention to influence the exchange rate. One exception was a study that found a significant and potentially large portfolio effect during the 1984–88 period, using survey data to measure exchange rate expectations and risk premium (Dominguez and Frankel, 1993b).

Recent research using data on order flow, however, identifies permanent price effects through the portfolio balance channel. Evans and Lyons (2001, 2002) found that intervention has a significant price impact in the most liquid currency pair market (before the introduction of the euro), the U.S. dollar-deutsche mark. The permanent effect of a US$1 billion dollar purchase was to appreciate the dollar by about 0.35 percent.⁹ They also found that foreign exchange transactions have the largest impact on the exchange rate when the flow of macroeconomic announcements is high.¹⁰

More generally, in a series of papers using an event study approach, Fatum (2000) and Fatum and Hutchison (2003a, b) find strong evidence in favor of intervention. In analyses of both the US dollar-deutsche mark and U.S. dollar-Japanese yen bilateral exchange rates, they find that sterilized intervention systematically affects the exchange rate level, regardless of whether it is secret or announced. The probability of success is much higher, however, when interventions are coordinated among central banks and when they are conducted on a large scale (i.e., greater than $1 billion). Also using an event study approach, Edison and others (2003) find that the Reserve Bank of Australia’s interventions had some success—albeit a modest one—in moderating the depreciating tendency of the Australian dollar, but interventions also increased exchange rate volatility.

The impact of intervention on exchange rate volatility has also been extensively researched. Intervention appears to be ineffective in reducing volatility, and oftentimes, increases it.¹¹ Both Dominguez (1998) and Hung (1997) provide evidence that following the Plaza Accord (September 1985) intervention tended to reduce exchange rate volatility among the G-3 currencies, but when the post-Louvre (1987–1989) period is examined, intervention increased volatility. Bonser-Neal and Tanner (1995) use implied volatilities from currency option prices and find that intervention raises exchange rate volatility. Beine and others (2002) study a longer period of interventions spanning 1985–1995 and also find that intervention increases exchange rate volatility in the short run. Cheung and Chinn (1999) conducted a survey with foreign exchange traders, 60 percent of whom view intervention as increasing exchange rate volatility.

Joint analyses of monetary and exchange rate policy actions find that intervention is effective in the case of the U.S. during the period 1973–1996 (Kim, 2003) and Japan (Guimarães, 2004). The approach is based on a structural VAR model similar to those used to study the monetary transmission mechanism. The identifying restrictions used in these models allow the exchange rate to have a contemporaneous impact on intervention, which captures the leaning against (or with) the wind by the intervening authorities. Moreover, the VAR model also permits the estimation of the impact of conventional monetary policy shocks (money or interest rate) on the exchange rate. The results also suggest that intervention in those two countries is sterilized and has an impact (small but significant) beyond the short-term considered in most studies that use daily data.

In contrast with most findings for advanced economies, empirical evidence on the effects of intervention in emerging market economies has been scant. In their empirical analysis of intervention in Mexico and Turkey, Domac and Mendoza (2002) conclude that central bank foreign exchange sales (but not purchases) were highly effective in influencing the exchange rate and in reducing volatility in both countries. In particular, they find that a net sale of US$100 million appreciates the exchange rate by 0.08 percent in Mexico and 0.2 percent in Turkey. A more recent study on Chile found that intervention had a small and generally insignificant effect on contemporaneous exchange rate movements, but in contrast, public announcements on potential intervention had a statistically significant impact on the level and trend of the exchange rate (Tapia and Tokman, 2004).

A. Mexico

Despite the peso’s floatation in 1994, the Mexican central bank has continued to intervene in the foreign exchange market to smooth exchange rate volatility. The concern over exchange rate volatility has stemmed from the exchange rate’s role as a key monetary policy variable, even though it has lost its anchor role to inflation targets since 1999. As in other emerging markets, the exchange rate has remained a determinant of inflationary expectations, even under the inflation targeting framework (Carstens and Werner, 1998, Ho and McCauley, 2003). Under inflation targeting, inflation and interest rates have come down substantially since the mid-1990s (Figure 1).

View Full Size

Mexico: Exchange Rate, Interest Rate, and Inflation

Citation: IMF Working Papers 2004, 123; 10.5089/9781451854640.001.A001

Source: Bank of Mexico and Datastream.

• Download Figure
• Download figure as PowerPoint slide

The authorities intervened to accumulate reserves, given the low level of international reserves in the immediate aftermath of the peso crisis. To this end, the central bank began auctioning put options with the objective of gradually building up reserves in August 1996. The central bank sold put options on the last business day of each month, allowing the holders of the option to sell U.S. dollars to the central bank anytime during the life of the option provided that the exercise price, the exchange rate of the day before, was more appreciated than the 20-day moving average of the interbank spot exchange rate. This condition limited the potential loss faced by the Banco de Mexico since the option could only be exercised if the peso was stronger than its 20-day moving average.¹² The auction of put options continued until June 2001, and resulted in an accumulation of reserves equivalent to 30 percent of reserves when the program ended (about US$14 billion). Intervention amounts and the peso-dollar exchange rate are shown in Figure 2.

View Full Size

Mexico: Exchange Rate and Foreign Exchange Intervention

Citation: IMF Working Papers 2004, 123; 10.5089/9781451854640.001.A001

Source: Bank of Mexico.

• Download Figure
• Download figure as PowerPoint slide

During 1996–2003, the central bank also intervened 14 times in a discretionary fashion, selling foreign exchange to stabilize the exchange rate, but no particular level was targeted. Sales of U.S. dollars were large, totaling US$2.9 billion, but still relatively low compared to the dollar purchases made through auctions of put options.¹³

A significant accumulation of reserves prompted the authorities to start selling foreign exchange directly to the market by May 2003. The amount to be sold to the market in a given quarter is pre-announced and equivalent to 50 percent of the reserves accumulated in the preceding quarter. The daily amount is based on the total amount for the quarter evenly distributed over the number of business days of the period in question. The switch was aimed at reducing the pace of reserve accumulation and the cost of holding additional international reserves, which had reached US$50 billion at end-2002, equivalent to more than 120 percent of short-term debt (by residual maturity) or about 50 percent of the monetary base, up from less than US$20 billion at end-1996. Dollar sales have been comparatively small in magnitude (up to US$32 million-daily), but much more frequent relative to the earlier program. According to the authorities, a major feature of both the options mechanism and the pre-announced sales is that they minimize the impacts on the market mechanism (e.g., pricing decisions) with negligible consequences for exchange rate volatility, claims that can be tested empirically (next section).

B. Turkey

The case of Turkey offers important insights on the challenges and limitations of empirically analyzing the effectiveness of intervention. Among emerging market economies, Turkey is one of the few countries with a (managed) floating exchange rate regime, where daily intervention data, albeit somewhat incomplete, is available. During the period studied here (March 2001–October 2003), the country implemented substantial economic reforms, lived through bouts of domestic political uncertainty, and was hit by contagion from financial market shocks.

Turkey’s exit from a crawling peg in February 2001 shifted the burden of price discovery to the foreign exchange market at a time when it was still undeveloped. During the crawling peg exchange rate regime, the foreign exchange market was heavily influenced by the Central Bank of Turkey (CBT), with most banks trading bilaterally with the CBT rather than among themselves. At the time of the exit from peg foreign exchange market liquidity was low, hedging instruments were virtually nonexistent, and financial institutions were caught with sizable short foreign currency positions and with limited capacity to manage foreign exchange risk. As a result, low turnover in the foreign exchange market may have been a dominant factor in the determination of the exchange rate compared to CBT’s interventions, at least during the early phases of the period analyzed here.

Since the floatation of the lira, the CBT’s interventions have undergone several phases (Table 1). The CBT initially sold foreign exchange through auctions to sterilize the liquidity injections associated with the Turkish Treasury’s use of external financial resources. These were combined with discretionary interventions to smooth exchange rate volatility related to negative external developments and domestic political problems (CBT, 2001). The CBT began conducting pre-announced (timing and amount) foreign exchange sale auctions in March 2001. Preannounced auctions were designed to enhance the transparency of official intervention and minimize their price impact. Auctions remained the main form of intervention throughout 2001, with brief interludes of discretionary intervention in lieu of or in parallel with preannounced auctions. Throughout 2001, the CBT sold US$6.5 billion in foreign exchange, enabling financial institutions to cover their short positions.

Table 1.

Turkey: Central Bank Foreign Exchange Intervention (March 2001–December 2003)

Source: Central Bank of Turkey, annual reports, press releases, and website (www.tcmb.gov.tr).

¹Data on the CBT’s discretionary interventions is not available.

²From July to December 2002, the CBT intervened three times (purchases) on a discretionary basis.

³The amounts of discretionary purchases were not disclosed.

Table 1.

Turkey: Central Bank Foreign Exchange Intervention (March 2001–December 2003)

Dates	Duration (Days)	Type of Intervention	Frequency	Amounts Sold (+) or Purchased (-) (US$, millions)
2/23/01–3/28/01	…	Discretionary auctions and sales	Ad-hoc	…
3/29/01–5/17/01	37	Preannounced sales	Daily	2,040
5/18/01–7/11/01	39	Discretionary sales1	Ad-hoc	1,678
7/12/01–8/31/01	38	Preannounced and discretionary sales	Bi weekly	1,575
9/4/01–11/30/01	64	Preannounced sales	Daily	1,180
12/1/01–3/31/02	85	No intervention	…	…
4/1/02–4/30/02	22	Preannounced purchases	Daily	(280)
5/1/02–6/28/02	43	Discretionary purchases	Ad-hoc	(515)
7/1/02–5/5/03	212	No intervention2	…	…
5/6/03–8/31/03	83	Preannounced and discretionary purchases3	Daily	(2,960)
9/1/03–10/1/03	29	Preannounced purchases with option for additional purchases	Daily	(1,824)

Source: Central Bank of Turkey, annual reports, press releases, and website (www.tcmb.gov.tr).

¹Data on the CBT’s discretionary interventions is not available.

²From July to December 2002, the CBT intervened three times (purchases) on a discretionary basis.

³The amounts of discretionary purchases were not disclosed.

View Table

Table 1.

Turkey: Central Bank Foreign Exchange Intervention (March 2001–December 2003)

Dates	Duration (Days)	Type of Intervention	Frequency	Amounts Sold (+) or Purchased (-) (US$, millions)
2/23/01–3/28/01	…	Discretionary auctions and sales	Ad-hoc	…
3/29/01–5/17/01	37	Preannounced sales	Daily	2,040
5/18/01–7/11/01	39	Discretionary sales1	Ad-hoc	1,678
7/12/01–8/31/01	38	Preannounced and discretionary sales	Bi weekly	1,575
9/4/01–11/30/01	64	Preannounced sales	Daily	1,180
12/1/01–3/31/02	85	No intervention	…	…
4/1/02–4/30/02	22	Preannounced purchases	Daily	(280)
5/1/02–6/28/02	43	Discretionary purchases	Ad-hoc	(515)
7/1/02–5/5/03	212	No intervention2	…	…
5/6/03–8/31/03	83	Preannounced and discretionary purchases3	Daily	(2,960)
9/1/03–10/1/03	29	Preannounced purchases with option for additional purchases	Daily	(1,824)

Source: Central Bank of Turkey, annual reports, press releases, and website (www.tcmb.gov.tr).

¹Data on the CBT’s discretionary interventions is not available.

²From July to December 2002, the CBT intervened three times (purchases) on a discretionary basis.

³The amounts of discretionary purchases were not disclosed.

The month of April 2002 marked the beginning of the second set of intervention phases, characterized by foreign exchange purchase operations. The move from foreign exchange sales to purchases was driven in part by reverse currency substitution engendered by growing confidence in the policy formulation and implementation, and a pick up in capital inflows (CBT, 2002). Foreign exchange purchases—first through preannounced auctions, then on a discretionary basis—were suspended in July 2002 amidst uncertainties before the November 2002 general elections, but resumed in May 2003 as uncertainties faded and reverse currency substitution continued.

Strong upward pressure on the lira caused the CBT to combine preannounced purchase auctions with discretionary or optional foreign exchange purchases. As a result, the CBT accumulated close to US$5 billion in reserves from May through October 2003. It is also worth noting that there were lengthy periods when the CBT did not intervene in the market at all (December 2001–March 2002 and July 2002–April 2003).

Despite its frequent presence in the market over long periods, the CBT has stated repeatedly through press releases and official policy statements that its interventions do not target a specific exchange rate level (CBT, 2003). These policy pronouncements, in principle, may be interpreted as having been designed to undercut the signaling channel by which intervention can influence the exchange rate.

Nevertheless, the CBT has been concerned about exchange rate misalignment and volatility (CBT, 2001, 2002, and 2003). The exchange rate has remained an important determinant of inflationary outcomes and expectations even after the float, and the authorities have been vigilant against volatility and have resisted it, market conditions and international reserve levels permitting. Figure 3 suggests that CBT interventions tended to “lean against the wind,” particularly during June–October 2001 and May–October 2003, among the two periods of heaviest intervention. Moreover, its interventions appear to have been more successful in tempering exchange rate movements when the lira was under upward rather than downward pressure.

View Full Size

Turkey: Central Bank Intervention in the Foreign Exchange Market March 29, 2001–October 3, 2003

Citation: IMF Working Papers 2004, 123; 10.5089/9781451854640.001.A001

Source: Central Bank of Turkey and Datastream.

• Download Figure
• Download figure as PowerPoint slide

Over the period of interventions analyzed here, market conditions and the CBT’s operating environment became more favorable. First, the CBT gradually regained credibility as it adhered to its monetary program and exercised restraint and transparency in the conduct of foreign exchange intervention. Second, the shift to the new nominal anchors, from base money to the inflation target, have become increasingly entrenched, reducing the pass through from the exchange rate to inflation (CBT, 2003). Third, favorable external finances and reverse currency substitution put upward pressure on the Turkish lira, aiding efforts to reduce inflation, lower interest rates, and bolster debt sustainability. As a result, the supply of foreign exchange in the market consistently exceeded preannounced amounts set for the purchase auctions. Exchange rate volatility also declined, albeit gradually (Figures 4 and 5). Fourth, the emergence of a vibrant interbank foreign exchange market and CBT’s less dominant role in it created more room for market participants to price foreign exchange and manage currency risks.

View Full Size

Turkey: Central Bank Intervention in the Foreign Exchange Market March 29, 2001–October 3, 2003

Citation: IMF Working Papers 2004, 123; 10.5089/9781451854640.001.A001

Source: Central Bank of Turkey and Datastream.

• Download Figure
• Download figure as PowerPoint slide

View Full Size

Turkey: Exchange Rate Trend and Volatility

Citation: IMF Working Papers 2004, 123; 10.5089/9781451854640.001.A001

Source: Datastream

• Download Figure
• Download figure as PowerPoint slide

A. Data Description

Mexico

Data on foreign exchange interventions are publicly available in daily frequency and cover the period August 1996 through June 2003. Both option-based and discretionary interventions are included. The data were obtained from Banco de Mexico’s website (www.banxico.org.mx) and consist of a total of 1800 observations, including no-intervention days. The other data include the spot exchange rate (peso per U.S. dollar), Mexican interest rate (Cetes 90-day), U.S. 3-month T-bill rate, yields on the Brady bond, and turnover and open interest on the Mexican peso contract traded on the Chicago Mercantile Exchange’s International Money Market (IMM). The second part of the data set was obtained from Datastream.

Table 2 presents the descriptive statistics for the exchange rate, the (log) first-difference of the exchange rate, interest rate differential, and spreads on Brady bond yields. The results for the ADF unit root test for the variables used in the regressions indicate that the series are nonstationary (with the exception of the differenced series) and some display other characteristics shared by financial time series, including departures from normality (e.g., fat tails and skewness) and volatility clustering, which are explored in the empirical models used in this paper.¹⁶

Table 2.

Mexico: Descriptive Statistics

Source: Bank of Mexico and authors’ calculations.¹ According to the Augmented Dickey-Fuller Test with a trend term and a maximum number of 15 lags selected according to the Bayesian Information Criteria, at the 90, 95, 99 percent confidence intervals. Stat (Non-stat) refers to stationary (nonstationary), i.e. rejection of the unit root null at the 5 percent level.

Table 2.

Mexico: Descriptive Statistics

Sample Period	Mean	Standard deviation	Skewness	Kurtosis	Jacque-Bera	Stationarity 1
Total Sample (Aug. 1996-Jun. 2003)
Peso-dollar ER	9.1944	0.8490	−0.3226	2.4231	56.2	Non-stat
1st Diff of Log ER	0.0179	0.5636	1.4430	21.1448	25317.4	Stat
Interest Diff (%)	13.3929	7.0010	1.0783	4.2810	471.9	Non-stat
Spreads (bps)	11.0762	1.5447	1.0606	4.1467	436.1	Non-stat
Sub-Sample I (Aug. 1996-Jun. 2001)
Peso-dollar ER	8.9513	0.8086	−0.3583	1.7697	108.3	Non-stat
1st Diff of Log ER	0.0139	0.6008	1.6036	22.4679	20794.5	Stat
Interest Diff (%)	16.2928	6.2505	1.2933	4.9176	553.8	Non-stat
Spreads (bps)	11.2855	1.5988	1.0690	3.8231	280.4	Non-stat

Source: Bank of Mexico and authors’ calculations.¹ According to the Augmented Dickey-Fuller Test with a trend term and a maximum number of 15 lags selected according to the Bayesian Information Criteria, at the 90, 95, 99 percent confidence intervals. Stat (Non-stat) refers to stationary (nonstationary), i.e. rejection of the unit root null at the 5 percent level.

View Table

Table 2.

Mexico: Descriptive Statistics

Sample Period	Mean	Standard deviation	Skewness	Kurtosis	Jacque-Bera	Stationarity 1
Total Sample (Aug. 1996-Jun. 2003)
Peso-dollar ER	9.1944	0.8490	−0.3226	2.4231	56.2	Non-stat
1st Diff of Log ER	0.0179	0.5636	1.4430	21.1448	25317.4	Stat
Interest Diff (%)	13.3929	7.0010	1.0783	4.2810	471.9	Non-stat
Spreads (bps)	11.0762	1.5447	1.0606	4.1467	436.1	Non-stat
Sub-Sample I (Aug. 1996-Jun. 2001)
Peso-dollar ER	8.9513	0.8086	−0.3583	1.7697	108.3	Non-stat
1st Diff of Log ER	0.0139	0.6008	1.6036	22.4679	20794.5	Stat
Interest Diff (%)	16.2928	6.2505	1.2933	4.9176	553.8	Non-stat
Spreads (bps)	11.2855	1.5988	1.0690	3.8231	280.4	Non-stat

Source: Bank of Mexico and authors’ calculations.¹ According to the Augmented Dickey-Fuller Test with a trend term and a maximum number of 15 lags selected according to the Bayesian Information Criteria, at the 90, 95, 99 percent confidence intervals. Stat (Non-stat) refers to stationary (nonstationary), i.e. rejection of the unit root null at the 5 percent level.

Turkey

Daily data on foreign exchange sale and purchase auctions were obtained from the CBT’s website (www.tcmb.gov.tr). The data set spans March 29^th 2001 through October 3rd 2003 with over 600 observations. It excludes discretionary foreign exchange sales and purchases through brokers and banks, including the CBT’s large discretionary purchases that exceeded those through auctions between May–October 2003. However, the absence of data on discretionary interventions, which we model through the use of a dummy variable, does not seem to be an important handicap in the empirical analysis.

Descriptive statistics on variables used to analyze the effectiveness of intervention are presented in Table 3. Distributions of all variables are asymmetric and nonnormal. Augmented Dickey-Fuller tests indicate that nearly all variables are nonstationary; and thus we take first differences before including them in the regressions.

Table 3.

Turkey: Descriptive Statistics

Source: Central Bank of Turkey, Datastream and authors’ calculations.

¹According to the Augmented Dickey-Fuller Test with a trend term and a maximum number of 15 lags selected according to the Bayesian Information Criteria, at the 90, 95, 99 percent confidence intervals. Stat (Non-stat) refers to stationary (nonstationary), i.e. rejection of the unit root null at the 5 percent level.

Table 3.

Turkey: Descriptive Statistics

Sample Period	Mean	Standard deviation	Skewness	Kurtosis	Jacque-Bera	Stationarity1
Total Sample (3/29/01-10/1/03)
TL-US$ ER	1,468,505	153,143	(0.313445)	2.266825	24.50417	Non-stat
1st Diff of Log ER	0.000489	0.013502	1.201994	15.52918	4279.219	Stat
Interest Diff (%)	48.75443	10.58766	0.449373	3.080784	21.44245	Stat
Spreads (bps)	790.4415	153.7943	0.255735	2.016018	32.38531	Non-stat
Sub-Sample I (3/29/01-6/30/02)
TL-US$ ER	1,375,984	131,347.8	(0.103088)	2.542028	3.300235	Non-stat
1st Diff of Log ER	0.001374	0.017393	0.935264	10.96643	873.3082	Stat
Interest Diff (%)	57.25809	7.534352	0.884976	4.090351	56.54089	Non-stat
Spreads (bps)	809.0573	161.0070	0.24335	1.872148	19.79255	Non-stat
Sub-Sample II (7/1/02-10/3/03)
TL-US$ ER	1,559,861	113,125.6	(0.531649)	1.657559	38.85893	Non-stat
1st Diff of Log ER	(0.000383)	0.007939	0.827176	5.574887	124.1117	Stat
Interest Diff (%)	40.35774	4.983065	(1.095038)	2.969123	63.56535	Non-stat
Spreads (bps)	772.0597	144.2331	0.186389	2.024692	114.44501	Non-stat

Source: Central Bank of Turkey, Datastream and authors’ calculations.

¹According to the Augmented Dickey-Fuller Test with a trend term and a maximum number of 15 lags selected according to the Bayesian Information Criteria, at the 90, 95, 99 percent confidence intervals. Stat (Non-stat) refers to stationary (nonstationary), i.e. rejection of the unit root null at the 5 percent level.

View Table

Table 3.

Turkey: Descriptive Statistics

Sample Period	Mean	Standard deviation	Skewness	Kurtosis	Jacque-Bera	Stationarity1
Total Sample (3/29/01-10/1/03)
TL-US$ ER	1,468,505	153,143	(0.313445)	2.266825	24.50417	Non-stat
1st Diff of Log ER	0.000489	0.013502	1.201994	15.52918	4279.219	Stat
Interest Diff (%)	48.75443	10.58766	0.449373	3.080784	21.44245	Stat
Spreads (bps)	790.4415	153.7943	0.255735	2.016018	32.38531	Non-stat
Sub-Sample I (3/29/01-6/30/02)
TL-US$ ER	1,375,984	131,347.8	(0.103088)	2.542028	3.300235	Non-stat
1st Diff of Log ER	0.001374	0.017393	0.935264	10.96643	873.3082	Stat
Interest Diff (%)	57.25809	7.534352	0.884976	4.090351	56.54089	Non-stat
Spreads (bps)	809.0573	161.0070	0.24335	1.872148	19.79255	Non-stat
Sub-Sample II (7/1/02-10/3/03)
TL-US$ ER	1,559,861	113,125.6	(0.531649)	1.657559	38.85893	Non-stat
1st Diff of Log ER	(0.000383)	0.007939	0.827176	5.574887	124.1117	Stat
Interest Diff (%)	40.35774	4.983065	(1.095038)	2.969123	63.56535	Non-stat
Spreads (bps)	772.0597	144.2331	0.186389	2.024692	114.44501	Non-stat

Source: Central Bank of Turkey, Datastream and authors’ calculations.

¹According to the Augmented Dickey-Fuller Test with a trend term and a maximum number of 15 lags selected according to the Bayesian Information Criteria, at the 90, 95, 99 percent confidence intervals. Stat (Non-stat) refers to stationary (nonstationary), i.e. rejection of the unit root null at the 5 percent level.

The data sample was divided into two subperiods and separate GARCH regressions were run on the subperiods as well as the entire sample. Dividing the sample into two facilitates the analysis of important differences in the type of interventions (sales versus purchases) and the environment in which they were conducted. The first subperiod, spanning March 2001 through June 2002, was dominated by foreign exchange sales and characterized by greater market and political uncertainty. The second subperiod, from July 2002 through October 2003, represented a period of greater confidence in the policy environment in which virtually all of the CBT’s interventions were purchase operations. Compared to the first, interest rates and spreads were substantially lower and their volatility, along with that of exchange rates, was considerably less during the second subperiod.

B. Empirical Model

The effects of intervention on the level and volatility of exchange rates is analyzed within the GARCH framework. The primary advantage of GARCH models is that they provide a unified framework to gauge the impact of intervention on the mean and conditional variance of exchange rate returns simultaneously. The empirical model allows the estimated conditional volatility to enter the mean equation (i.e., the “GARCH-in-mean” effect) and tests for asymmetric effects on volatility of “negative” shocks, defined as unexpected exchange rate depreciations. In addition to its computational simplicity, GARCH models provide relatively good forecasts of realized volatility, and have proved useful for modeling the volatility dynamics of exchange rates and asset prices more generally (Andersen and Bollerslev, 1998).

The empirical analysis of the effectiveness of intervention is based on the Asymmetric Component Threshold GARCH (ACT-GARCH) specification, which jointly estimates the impacts of intervention on volatility at different time horizons. The volatility part of the model allows for asymmetric responses of the conditional volatility to unexpected exchange rate depreciations. Furthermore, the model is consistent with stylized facts in asset pricing empirics, including persistence and volatility clustering.

The baseline model is given by:

where Δs_t is the (log) first-difference of the exchange rate (expressed in terms of local currency per US dollar and in log form), ${{I}}_{{t}}^{-}\left({{{I}}_{{t}}}^{+}\right)$ denotes sales (purchases) of foreign currency in millions of US dollars by the central bank for intervention purposes, d_t is the interest rate differential (domestic minus foreign), sp_t is the yield spread on a sovereign foreign currency bond over a comparable U.S. treasury bond. The log exchange rate, interest differential, and yield spreads are first differenced to obtain their stationary forms. The error term is the unexpected return which is used to model the conditional volatility of the exchange rate in the volatility equations (2) and (3).

Equation (1) of the empirical model (the “mean” equation) analyzes changes in the exchange rate return (depreciation or appreciation against the dollar) as a function of intervention, interest rate differentials, and yield spreads on a sovereign bond. The interest differential aims to capture the possible impact of monetary policy actions and local money market conditions on the exchange rate. Yield spreads on sovereign external debt over a comparable US Treasury bond are included as a measure of country risk and foreign investor sentiment, which are possibly key determinants of demand for local currency. It is hypothesized that a higher interest differential appreciates the domestic currency (Δs_t < 0), net purchases of foreign exchange depreciate the domestic currency, and higher yield spreads are associated with depreciations of the domestic currency.

In equation (2), h_t is the conditional volatility of the exchange rate (log returns), z_t is a dummy variable indicating unexpected exchange rate appreciations (i.e. if ε_t < 0, then z_t > 0). The model allows mean reversion of the short-term volatility, h_t, to a time-varying longer-term volatility, given by q_t, in contrast to the constant long-term volatility assumed in the standard GARCH model.^17,18

Equation (2) models the short-term conditional exchange rate volatility, h_t, as a function of a time-varying long-term volatility, q_t, lagged unexpected shocks relative to lagged long-term volatility, given by the term (${{{\epsilon }}^2}_{{t}-1}$ – q _t-1), lagged volatility relative to lagged long-term volatility, given by the term (h_t-1 - q_t-1), and the set of regressors of explanatory variables that were included in the mean equation (intervention, interest rate differentials, and sovereign spreads). Lagged unexpected shocks and volatility are included to capture volatility clustering (as in standard GARCH-type models), since high volatility periods tend to be clustered over time. The equation also includes a term ((${{{\epsilon }}^2}_{{t}-1}$ – q _t-1)z _t-1) that allows for asymmetric impacts of past shocks (relative to long-term volatility) on short-term volatility. If the estimated τ is less than zero, then unexpected depreciations increase short-term volatility.

The model departs from the standard GARCH representation by assuming that the long-term volatility is not constant. The long-term volatility equation is given by (3), and like its short-term counterpart, it depends on a set of explanatory variables (intervention, interest differentials, sovereign spreads), its own lagged value (q_t-1), and past shocks (${{{\epsilon }}^2}_{{t}-1}$). Unlike the short-term volatility, q_t converges to a constant ω.

Some general features of the model above are noteworthy. First, it allows for asymmetric shocks in the conditional (short-term) variance equation. In particular, if τ < 0, then the impact of “negative” shocks (unexpected domestic currency depreciation, ε_t > 0) on short-term volatility is given by α, greater than the impact of “positive” shocks (unexpected appreciation), which is given by (τ + α).¹⁹ Second, the short-term impact of foreign exchange intervention on exchange rate volatility may differ from the long-term impact. The empirical model may also be augmented with other exogenous variables, such as sovereign spreads, market turnover, order flow, and other relevant variables. For instance, in the case of Mexico, futures market turnover (and open interest) are included in the exchange rate volatility equations.²⁰

This paper also analyzes the effects of volatility on intervention, given the evidence in favor of reverse causation between exchange rate returns and intervention (policy reaction function). Following Dominguez (1998) and Baillie and Osterberg (1997), we apply the Probit model to evaluate whether excessive volatility, defined as the deviation of estimated volatility from its recent trend, increases the probability that the central bank will intervene in the foreign exchange market.²¹ In addition to excessive volatility, the estimated Probit model also includes the deviation of the current exchange rate from its recent moving average. Although other variables could be included in the model, the estimated model appears to perform quite well and makes our analysis more comparable to previous work.

The estimated model is given by:

where Pr denotes probability and Ф(.) denotes the standard normal transformation. If the estimated α₁ (α₂) is statistically significant (different from 0), then deviations from the k-day exchange rate trend (volatility) affect the probability of intervention.

C. Estimation Results

Mexico

The first set of regressions on Mexico (not reported here) highlight the importance of model specification and properly accounting for the simultaneity problem. The regressions also show that the effects of intervention on the exchange rate vary according to the sample period and whether intervention is lagged or contemporaneous.²² For example, the regression of exchange rate returns on contemporaneous central bank purchases indicate that a US$100 million U.S. dollar purchase by the Banco de Mexico (BoM) appreciates the peso by 0.2 percent against the U.S. dollar, while an equivalent sale of U.S. dollars depreciates the peso by 1.4 percent (both statistically significant at the 1 percent level). This result could be erroneously interpreted as “leaning against the wind,” but in fact, it is consistent with the rationale that investors tend to exercise their put options when the domestic currency appreciates. This, in turn, would make the error term correlated with the explanatory variable, which requires the application of instrumental variables/generalized method of moments or the use of lagged intervention to account for the correlation.²³

In order to redress the simultaneity problem, the model was re-specified by using two-day lagged intervention. The estimates from the second set of regressions using lagged intervention are presented in Table 4. The upper half of the table shows the estimates for the mean equation.

Table 4.

Mexico and Turkey: Asymmetric Component GARCH Model Estimates

Notes: An * denotes significant at 10 percent, and ** denotes significant at 5 percent. The results reported above are based on equations (1-3) after dropping the weekend dummy and the moving average parameter, which were insignificant in all specifications estimated. The coefficient on intervention in the mean equation gives the impact for a US$100 million purchase/sale of foreign exchange in percent. The sample sizes are respectively 1800 and 1278.

Table 4.

Mexico and Turkey: Asymmetric Component GARCH Model Estimates

$\begin{array}{l}{\Delta }{{s}}_{{t}}={{\beta }}_0+{{\beta }}_1{{{I}}_{{t}}}^{+}+{{\beta }}_2{{{I}}_{{t}}}^{-}+{{\beta }}_3{\Delta }{{d}}_{{t}}+{{\beta }}_4{\Delta }{s}{{p}}_{{t}}+{{\epsilon }}_{{t}}\\ {{h}}_{{t}}-{{q}}_{{t}}={\alpha }({{{\epsilon }}^2}_{{t}-1}-{{q}}_{{t}-1})+{\tau }(({{{\epsilon }}^2}_{{t}-1}-{{q}}_{{t}-1}){{z}}_{{t}-1}+{\delta }({{h}}_{{t}-1}-{{q}}_{{t}-1})+{{\gamma }}_1{{{I}}_{{t}}}^{+}+{{\gamma }}_2{{{I}}_{{t}}}^{-}+{{\gamma }}_3{\Delta }{{d}}_{{t}}+{{\gamma }}_4{\Delta }{s}{{p}}_{{t}}\\ {{q}}_{{t}}={\omega }+{\rho }({{q}}_{{t}-1}-{\omega })+{\phi }(({{{\epsilon }}^2}_{{t}-1}-{{h}}_{{t}-1})+{{\gamma }}_5{{{I}}_{{t}}}^{+}+{{\gamma }}_6{{{I}}_{{t}}}^{-}+{{\gamma }}_7{\Delta }{{d}}_{{t}}+{{\gamma }}_8{\Delta }{s}{p}_{{t}}\end{array}$
	Mexico	Turkey
Exchange rate level (mean) equation
β1 (Intervention-sale)	−0.4319**	−3.38E-05
β2 (Intervention-purchase)	0.0122	7.15E-06
β3 (Interest differential)	0.0369**	0.000203
β4 (Spreads)	0.2163**	0.000172*
Short-term volatility equation
τ (Negative shock)	−0.2363**	0.040596
γ1 (Intervention-sale)	0.0595**	−0.011062***
γ2 (Intervention-purchase)	0.0002	−2.08E-07
γ3 (Interest differential)	0.0162	−1.25E-05
γ4 (Spreads)	0.1886**	3.66E-07
Long-term volatility equation
γ5 (Intervention-sale)	0.0115**	0.011062***
γ6 (Intervention-purchase)	−6.73E-05	2.72E-08
γ7 (Interest differential)	−0.0211**	3.17E-06
γ8 (Spreads)	0.0167	4.38E-07

Notes: An * denotes significant at 10 percent, and ** denotes significant at 5 percent. The results reported above are based on equations (1-3) after dropping the weekend dummy and the moving average parameter, which were insignificant in all specifications estimated. The coefficient on intervention in the mean equation gives the impact for a US$100 million purchase/sale of foreign exchange in percent. The sample sizes are respectively 1800 and 1278.

View Table

Table 4.

Mexico and Turkey: Asymmetric Component GARCH Model Estimates

$\begin{array}{l}{\Delta }{{s}}_{{t}}={{\beta }}_0+{{\beta }}_1{{{I}}_{{t}}}^{+}+{{\beta }}_2{{{I}}_{{t}}}^{-}+{{\beta }}_3{\Delta }{{d}}_{{t}}+{{\beta }}_4{\Delta }{s}{{p}}_{{t}}+{{\epsilon }}_{{t}}\\ {{h}}_{{t}}-{{q}}_{{t}}={\alpha }({{{\epsilon }}^2}_{{t}-1}-{{q}}_{{t}-1})+{\tau }(({{{\epsilon }}^2}_{{t}-1}-{{q}}_{{t}-1}){{z}}_{{t}-1}+{\delta }({{h}}_{{t}-1}-{{q}}_{{t}-1})+{{\gamma }}_1{{{I}}_{{t}}}^{+}+{{\gamma }}_2{{{I}}_{{t}}}^{-}+{{\gamma }}_3{\Delta }{{d}}_{{t}}+{{\gamma }}_4{\Delta }{s}{{p}}_{{t}}\\ {{q}}_{{t}}={\omega }+{\rho }({{q}}_{{t}-1}-{\omega })+{\phi }(({{{\epsilon }}^2}_{{t}-1}-{{h}}_{{t}-1})+{{\gamma }}_5{{{I}}_{{t}}}^{+}+{{\gamma }}_6{{{I}}_{{t}}}^{-}+{{\gamma }}_7{\Delta }{{d}}_{{t}}+{{\gamma }}_8{\Delta }{s}{p}_{{t}}\end{array}$
	Mexico	Turkey
Exchange rate level (mean) equation
β1 (Intervention-sale)	−0.4319**	−3.38E-05
β2 (Intervention-purchase)	0.0122	7.15E-06
β3 (Interest differential)	0.0369**	0.000203
β4 (Spreads)	0.2163**	0.000172*
Short-term volatility equation
τ (Negative shock)	−0.2363**	0.040596
γ1 (Intervention-sale)	0.0595**	−0.011062***
γ2 (Intervention-purchase)	0.0002	−2.08E-07
γ3 (Interest differential)	0.0162	−1.25E-05
γ4 (Spreads)	0.1886**	3.66E-07
Long-term volatility equation
γ5 (Intervention-sale)	0.0115**	0.011062***
γ6 (Intervention-purchase)	−6.73E-05	2.72E-08
γ7 (Interest differential)	−0.0211**	3.17E-06
γ8 (Spreads)	0.0167	4.38E-07

Notes: An * denotes significant at 10 percent, and ** denotes significant at 5 percent. The results reported above are based on equations (1-3) after dropping the weekend dummy and the moving average parameter, which were insignificant in all specifications estimated. The coefficient on intervention in the mean equation gives the impact for a US$100 million purchase/sale of foreign exchange in percent. The sample sizes are respectively 1800 and 1278.

The results indicate that the impacts of intervention on the level of the exchange rate are nontrivial. In particular, a two-day lagged US$100 million sale appreciates the peso (against the U.S. dollar) by 0.4 percent (statistically significant at the 5 percent level), but purchases of foreign exchange do not have a statistically significant impact on the value of the peso.²⁴ The results also underscore the importance of estimating the impacts of sales and purchases of foreign exchange separately, particularly when there are systematic differences between purchases versus sales.

The effects of intervention on volatility are also significant and are shown in the lower half of Table 4. Several factors account for the time-varying nature of exchange rate volatility, including the yield spread on Brady Bonds, interest rate differential, and intervention. The impact of intervention on exchange rate volatility is also estimated separately for sales and purchases.

The model indicates that foreign exchange sales and changes in the Brady bond yield spread increase the short-term volatility of the exchange rate (h_t) at the 5 and 1 percent significance levels, respectively. The estimates indicate the existence of asymmetric shocks to the conditional variance (significant at the 10 percent level in both samples): Unexpected domestic currency depreciations have a larger effect on volatility than unexpected appreciations. By contrast, neither changes in the interest rate differential nor foreign exchange purchases affect short-term exchange rate volatility. Foreign exchange sales also increase the long-term component of exchange rate volatility (significant at the 5 percent level). The estimated effect of purchases is negative, but not statistically significant.²⁵ Changes in the Brady yield spread did not have a significant impact on long-term volatility, while changes in the interest rate differential has a negative effect (significant at the 5 percent level).

The main empirical findings for Mexico may be interpreted as follows.²⁶ First, intervention seems to have a non-negligible effect on exchange rate changes, with US$100 million sale of foreign exchange by the central bank estimated to appreciate the peso (against the U.S. dollar) by 0.4 percent. However, foreign exchange purchases, which constitute the bulk of interventions in Mexico during the period covered here, did not appear to have had a statistically significant impact on the value of the peso. This is consistent with the authorities’ objective of accumulating international reserves in a floating exchange rate regime.

Nevertheless, intervention has a nontrivial impact on exchange rate volatility. The results indicate that foreign exchange sales increase both the short- and long-term volatility of the exchange rate, which in part may be because sales are seen as less credible.²⁷ Other factors (e.g., interest rate differentials) also appear to be important in explaining exchange rate volatility at longer horizons. This finding has important implications for exchange rate policy, especially if (short-term) volatility-enhancing interventions can influence the expectations of market participants, as described by Hung (1997).

The probit estimations for Mexico indicate that “excessive” exchange rate volatility decreases the probability of intervention, in contrast with the findings of Dominguez (1998) and Baillie and Osterberg (1997b) (Table 5).²⁸ The estimations also reveal that exchange rate depreciations increase the probability of intervention. The results, which are based on a 21-day window (k=21) and the full sample, are generally not robust to the choice of k and the sample period. For instance, when the sample period 1996–2001 is considered, increases in volatility relative to its recent trend raise the probability of intervention. Nonetheless, the probit estimation results underscore the importance of controlling for simultaneity effects when estimating the impacts of intervention on the level and volatility of the exchange rate.

Table 5.

Mexico and Turkey: Probit Model Estimates

$\mathrm{Pr}\left\{\left|I_t\right|\ne 0\right\}={\Phi }({\alpha }_0+{\alpha }_1(s_{t-1}-{\displaystyle {\sum }_{j=1}^k{s}_{t-j}}/k)+{\alpha }_2(\sqrt{h_t}-{\displaystyle {\sum }_{j=1}^k\sqrt{h_{t-j}}/k))}$

Notes: *, **, and *** denote significance at 10, 5, and 1 percent levels. Huber-White standard errors are used. The results are based on k=21; s_t is the log of the nominal local currency-U.S. dollar rate, h_t denotes the estimated conditional volatility of the exchange rate return.

Table 5.

Mexico and Turkey: Probit Model Estimates

$\mathrm{Pr}\left\{\left|I_t\right|\ne 0\right\}={\Phi }({\alpha }_0+{\alpha }_1(s_{t-1}-{\displaystyle {\sum }_{j=1}^k{s}_{t-j}}/k)+{\alpha }_2(\sqrt{h_t}-{\displaystyle {\sum }_{j=1}^k\sqrt{h_{t-j}}/k))}$

Coefficients	Mexico 1996–2003	Turkey 2001–2003
α 0	−1.36**	−0.38***
α 1	8.58*	11.28
α 2	−0.03**	−2.42

Notes: *, **, and *** denote significance at 10, 5, and 1 percent levels. Huber-White standard errors are used. The results are based on k=21; s_t is the log of the nominal local currency-U.S. dollar rate, h_t denotes the estimated conditional volatility of the exchange rate return.

View Table

Table 5.

Mexico and Turkey: Probit Model Estimates

$\mathrm{Pr}\left\{\left|I_t\right|\ne 0\right\}={\Phi }({\alpha }_0+{\alpha }_1(s_{t-1}-{\displaystyle {\sum }_{j=1}^k{s}_{t-j}}/k)+{\alpha }_2(\sqrt{h_t}-{\displaystyle {\sum }_{j=1}^k\sqrt{h_{t-j}}/k))}$

Coefficients	Mexico 1996–2003	Turkey 2001–2003
α 0	−1.36**	−0.38***
α 1	8.58*	11.28
α 2	−0.03**	−2.42

Notes: *, **, and *** denote significance at 10, 5, and 1 percent levels. Huber-White standard errors are used. The results are based on k=21; s_t is the log of the nominal local currency-U.S. dollar rate, h_t denotes the estimated conditional volatility of the exchange rate return.