major weekness of the floating exchange rate regime is that volatile exchange rates have discouraged trade- reported in IMF by the Group of Twenty Four and the Group of Ten
However, there is no consensus established in literature theoretically and empirically on how trade flows are affected by exchange rate volatility.
1.2 Objective
The thesis contributes to the existing literature on exchange rate volatility and trade flows in several aspects. Firstly, the thesis uses the most updated disaggregate data of four industries specified according to SITC in period between January 1998 and June 2012 to examine the impact of exchange rate volatility on Japanese export. Secondly,
1.3 The thesis layout
The thesis is developed according to the following structure:
Chapter 2: Literature review: looks at the literature on the impact of exchange rate volatility on export in two main aspects: the theoretical view of the impact and the methodology employed to study the impact.
Chapter3: Data processing and analysis
Chapter 4: Methodology
Chapter 5: Empirical Results
Chapter 6: Conclusion
Chapter 2: Literature review
The aim of this chapter is to explore the impact of exchange rate volatility on export by reviewing literature conducted on trade flows in two main directions: the theoretical views of the impact which varies among studies and the methodology applied to examine the effect and measure the exchange rate volatility. The pros and cons of each method as well as latest econometric method will be discussed to determine the appropriate methodology for this thesis. For the comprehensiveness of the review, studies since 1973 until now will be included to analyse the two main aspects. List of studies with main findings and methodology can be found in appendix 1. The chapter is structured as follows:
Section 2.1 demonstrates the theoretical contributions to explaining the mechanism of the effect of exchange rate volatility on international trade along with its influencing factors. The overall review finds literature takes varying view on this ground
Section 2.2 presents pertinent aspects of methodology employed in empirical studies. The review includes advantage and disadvantage of each method as evidenced in literature finding. Also, the results of the impact are inconclusive throughout literature.
Section 2.3 summarizes the chapter and concludes the appropriate for the thesis.
2.1 Theoretical views of the impact
In attempt to justify the sign and magnitude of exchange rate uncertainty on trade, literature has established various theoretical models which explain nature and determinants of the impact. Unfortunately, as the models leads to diverse results, there is no unique answer to this ground. Main theory and factors attributed to the effect of exchange rate volatility on trade flows are presented as follows:
2.1.1 Degree of risk aversion
One of the traditional views indicates that the impact of exchange rate variability on trade flow is dependent on the attitude of trader towards risks.
The notion that trade flows are adversely affected by exchange rate variability was first introduced by Ethier (1973). Ethier establishes a model where risk-averse company facing exchange rate uncertainty determines the volume of goods to be imported and the amount of forward exchange cover to maintain. According to the research, the amount of trade is unaffected by the level of exchange rate variability while the degree of forward cover has to be taken into consideration given that the import contract is denominated in foreign currency, and the profit for any value of exchange rate is known by the firm and. However, when including uncertainty about firm's revenue with respect to exchange rate, negative impact of exchange rate volatility on trade is found and level of the effect decreases as the firms get more speculative. Ethier also notes that the affect is subject to changes if the denomination of the contract is domestic currency, which means the exporter bears the direct exchange risk while importer faces indirect exchange risk passed through the response of his revenue. The issue about invoicing currency in international trade contract will be discussed later in the thesis.
The finding is consistent with Clark (1973) which investigates behaviours of an exporting company in bilateral trade under the assumption of perfect competition with foreign-denominated currency contract. The finding shows the higher level of exchange rate uncertainty, the more uncertain the profit in domestic currency. After employing a utility function of profit and risk aversion [1] , Clark concludes that in order to maximise utility, a risk-averse firm will decrease its goods supply until marginal revenue is larger than marginal cost to offset additional risks.
A more comprehensive finding is suggested by De Grauwe (1988). The author models the decision of a competitive supplier to trade in domestic or foreign market and finds exchange rate volatility will either encourage or depress trade according to the degree of the firm's risk aversion. Assuming the only source of uncertainty is the export's local currency, the producer's response to increase in exchange rate risks is found to depend on the convexity properties of the function between expected marginal utility of export income and exchange rate. According to the study, producers with low degree of risk aversion will export less because the expected marginal utility of export decreases as exchange rate uncertainty increases. Nevertheless, if they are highly risk-averse and expects for worst case, hence expected marginal utility of export rises as exchange rate risks increase, producers tend to export more to avoid seeing their revenues fall drastically. In other words, De Grauwe indicates that the influence of exchange rate variability on trade is determined by two effects: a substitution effect whereby greater uncertainty discourages trade flows and income effect whereby trade increases to compensate for a decline in the total expected utility. When firm is highly risk-averse, the income effect has preponderance over the substitution effect, which means greater exchange rate uncertainty will stimulate trade.
2.1.2 Invoicing currency
In bilateral trade, exporter or importer can choose to invoice his contract in either foreign or his home currency, which means at the settlement of the contract, he will receive revenue in foreign or home currency, respectively. Researchers have found that selection of invoicing currency can influence the relationship between exchange rate volatility and trade flows.
Study such as Baron (1976a) explores the impact by evaluating the influence of invoicing currency on the exporter's pricing and production decision when facing exchange rate risks under the environment of imperfectly competitive market. The study points out for an exporting firm, the contract with foreign-denominated currency will impose greater price risk [2] while more quantity demand risk [3] is found in case of domestic-invoicing currency. It means that under circumstances of volatile exchange rate, risk-averse firm using foreign-invoiced currency will increase prices. With domestic currency, how the firm maximises utility will depend on the shape of its demand curve. For example: prices will be decreased if demand curve is linear, which leads to a rise in demand and decline in profitability uncertainty.
In line with this study is Giovannini (1988) who models the pricing decision of risk-neutral exporter regarding expected profit under the impact of exchange rate volatility. The author reaches the same conclusion with previous study, confirming that using domestic currency denomination in export pricing will lead to different decision in export prices when exchange rate is volatile. It is also stressed that the link between exchange rate uncertainty and exported prices set in foreign currency is dependent on the shape of demand and cost functions. To be specific, if the functions are linear, export prices will be decreased as exchange rate risk increase. Another finding is invoicing prices in either domestic or foreign currency will lead to higher expected profit if the property of exchange rate and profits function is convex or concave, respectively. Hence, the two studies indicate an increase in exchange rate volatility has no clear-cut effect on trade and the effect relies considerably on the invoicing currency of the contract.
2.1.3 Hedging in forward market
Forward market is believed to provide hedging opportunities to deal with exchange rate uncertainty and reduce the impact caused by variation of exchange rate. Therefore, many studies in this field have incorporated the role of forward market in their research on the impact of exchange rate volatility on trade flow to see how the presence of the market affects the relationship. The literature finding on this ground is varied among studies.
Early studies such as Ethier (1973) and Baron (1976b) suggest that in a world where exchange rate fluctuation is the only source of uncertainty, perfect forward market will remove the influence of exchange rate uncertainty on volume of trade. Later studies by Viaene and de Vries (1987, 1992) counter this argument, pointing out even with the availability of forward market, there is still an indirect link between spot exchange rate and the trade volume. In other words, the volume of trade is still influenced by spot exchange rate via its effect on forward rate. According to the study, exporters and importers are on opposite side of the market; therefore, exchange rate volatility will have different effect on them. As stated in the research, which side of traders who will be beneficial from exchange rate risk depends on the sign of trade balance (or the net foreign currency exposed to risks) and the risk aversion. Hence, exporter will lose (or win) and importer will win (or lose) if trade balance is positive (negative). In other words, whether the trade is discouraged or stimulated by exchange rate variability is determined by the net currency position of the trader.
Caporale and Doroodian (1994) agree with previous studies on the availability of hedging in forward market, however, they emphasize that this activity will incur great cost to the firms as the timing and volume of foreign transactions are highly difficult to foresee. Consistent with this view is Obstfeld and Rogoff (1998) study on the firm hedging strategy with respect to its level of risk aversion. According to the study, in case of exchange rate variability, hedging in forward market will be performed by risk-averse company and result in a great cost, which is transferred into export prices, thereby, negatively influences the production output and consumption.
IMF(1984) takes a different view on hedging opportunity in forward market, pointing out as the availability of foreign currency forward contracts are not always guaranteed to all trading firms in every nation, it is doubtful that forward market can take role in the impact of exchange rate volatility on trade in all situation. The study also criticizes forward contract for its high cost, short maturities and large amount while its ability to cover possible exchange rate variations during the intended period is restricted. Therefore, it is concluded that only large-sized company is able to take most advantage of hedging on forward market; small firm, on the other hand, is least beneficial from this.
2.1.4 Other factors
Besides the aforementioned theoretical views, the impact of exchange rate volatility on trade flow can be accounted by other factors such as profit opportunity, openness of the economy or the source and consequence of the exchange rate risks.
An positive hypothesis based on profit opportunities is proposed by Franke(1991) which investigates the exporting strategy of a risk neutral firm. Assuming firm exists in a world monopolistic competition and aims to maximise net present value of expected cash flows from export, Franke contends that the exporter will decide his strategy by weighting the cost to enter (or leave) a foreign market against the export profits (losses). As shown in the study, an increase in exchange rate variability will benefit trade if the cash flow and exchange rate function is convex, hence the cash flows' net present value will increase faster than the entry or exit cost.
With regard to economic condition, Maskus (1986) attributes the variation of the impact of exchange rate uncertainty on trade to the openness of the examined industry or economy to international trade, the levels of industry concentration and the use of long-term contracts. The study estimates the impact on US trade of agricultural product in 10 years period from 1974 to 1984, finding that the exchange rate volatility has greater negative effect on this industry than others because in this sector, openness to international trade is higher with low level of industry concentration and contracts with longer term.
Concerning factors related to exchange rate volatility, Sercu and Uppal (1997) emphasizes the importance of indentifying the source of the increase in exchange rate uncertainty. As presented in the study, under assumption that international commodity markets are imperfect, exchanger rate volatility originating from either commodity markets segmentation or endowment process will have different impact on trade with negative effect in the former and positive in the latter. Brada and Mendez (1988) take a different view, showing that exchange rate volatility can indirectly influence trade flow via government action. The study shows the case where government establishes trade barriers to equalize the destabilized effect which is caused by exchange rate fluctuation and unrelated to changes in determinants of international trade. The volume of trade, interestingly, is found to be adversely affected by this action.
2.2 Methodology
Besides theoretical model, it is proved that methodology employed to conduct the research also determines the results obtained for the impact. According to survey by McKenize (1999) and Bahmani-Oskooee and Hegerty (2007), the problem of mixed results in literature can be attributed to the selection of data approach, measure of exchange rate volatility and method to model the relationship. Details of each aspect are discussed as follows:
2.2.1 Data approach
In investigation of the impact of exchange rate uncertainty on trade, researchers have undertaken their studies under three major data approaches: aggregate, bilateral (or cross-sectional) and disaggregate data.
The use of aggregate data is pervasive in literature and produces mixed results for the impact, from significantly negative (eg: Perée and Steinherr (1989); Quian and Varangis (1994), Bleaney(1992) and significantly positive (eg: Asseery and Peel ( 1991) ) to insignificant and ambiguous (eg: Bailey and Tavlas(1988), Lastrapes and Koray(1990)). The approach does not take into account the effect of exchange rate volatility in sub-industry; in contrast, it assumes that there is uniformity in the magnitude and direction of the impact on all commodities between countries. This assumption is subject to criticism as it can neglect important information of the impact, which leads to biased results. Details of the criticism of aggregate data will be discussed in later part of the thesis.
Besides aggregate data, studies have adopted bilateral trade flows in expectation of producing reliable conclusion of the impact. Unfortunately, it is evidenced that bilateral data does not improve the quality of results compared to aggregate data. Applying equilibrium market model by Ethier (1973) with relaxed assumption of infinite elasticity supply function , Hooper and Kohlhagen (1978) examine the impact on trade flows and prices using quarterly multilateral and bilateral trade data within UK, US, Japan, Canada, France and Germany. The results exhibit insignificant effect for aggregate data; bilateral data also produces no distinct impact. Another study employing bilateral trade data by Cushman (1988) in attempt to confirm opposite effect in previous literature reaches similar conclusion. The model from Cushman (1983) for bilateral trade flow is extended to new period 1974-1983 along with numerous exchange rate volatility measures. The finding is relatively similar to previous publication, showing that the effect can be mixed from significantly negative on export and import to significantly positive, reported as "puzzling", in some cases.
As bilateral data is not more sufficient than aggregate data in evaluating the impact, further step in empirical research is taken to adopt disaggregate data. Most findings established in this ground support the use of disaggregate over aggregate data. As argued by Bini-Smaghi (1991), disaggregate data can demonstrate various characteristic of the commodities and markets where trade occurs, therefore is not subject to the constraint that income, price and exchange rate risk elasticities throughout the markets should be equal as in case of aggregate data (Goldstein and Khan, 1985). The study employing Hooper and Kohlhagen (1978) model of export function to examine the effect on intra-EMS trade from 1976 to 1984 reveals that disaggregate data produces significant negative impact on trade. In addition, McKenzie (1998) shows the case when aggregate and disaggregate data is used for the Australian trade flows from 1988 to 1995, the former produces restricted and contradictory effects of exchange rate uncertainty; the later, on the contrary, finds significant relationship and reveals that the sign and magnitude of the impact may vary with respect to the nature of commodities market where trade occurs. Given the advantage over aggregate data, disaggregate has been increasingly used in literature (eg: Klein (1990), Grobar(1993), Stokman(1995), Doyle (2001, Ozturk and Kalyoncu (2009)), Bahmani-Oskooee and Bolhassani (2012))
2.2.1 Measuring exchange rate volatility
According to literature surveys by Cote (1994) and McKenzie (1999), in attempt to investigate the impact of exchange rate on trade, research has emphasized the importance of finding the best proxy for exchange rate uncertainty. Concerning this matter, the question centres on the type of exchange rate used in the estimation and the measure of exchange rate which are best appropriate to capture the volatility. Unfortunately, this ground is still open to controversy. As no estimation is proved to be superior to others, there is still no consensus on the best estimation method for exchange rate variability. Details of main aspects concerning measuring exchange rate volatility are given as follows:
Nominal exchange rate versus real exchange rate
The use of nominal or real exchange rate has received a lot of attention from researchers on the sufficiency to capture exchange rate variability.
Opponent of the real exchange rate such as Gotur (1985) claims that real exchange rate produces better measurement for effect of exchange rate uncertainty on trade in long-run prospect. It is argued that real exchange rate should be adopted in medium-term horizon because a large part of the change in revenues and costs of the firm which are caused by nominal exchange rate volatility is subject to offsetting by costs and prices movements.
In contrast, Akhtar and Hilton (1984) supports nominal exchange rate, claiming that there's still a shortage of literature substantiating the existence of purchasing power parity in medium term. In line with this view, Bini-Smaghi (1991) uses OLS and disaggregate data to present the case when nominal exchange rate outperforms real data series. He also points out variability in relative prices, which derives from real exchange rate, is a separate and extra risk to traders. Therefore, nominal rate is a more appropriate proxy in the measurement.
However, studies such as Bailey et al. (1987) and McKenzie and Brooks (1997) argue that the use of nominal or real exchange rate produces no distinct difference in the volatility estimates. Bailey et al. (1987) analyze the impact from period 1962 to 1985 in OECD Big Seven using absolute percentage change and log of standard deviation of both nominal and exchange rate to measure volatility. It is found that in certain cases, nominal and real exchange rate produce effect with same sign and magnitude, which indicates that the choice of exchange rate does not significantly influence the study's outcome. In their study to analyze the impact on US-German trade flows, McKenzie and Brooks (1997) applies Lagrange Multipliers and Ljung-Box test to examine the ARCH effect in data series of both nominal and real exchange rate. The finding reveals no significant difference between the coefficients estimated in ARCH(1) model using the two exchange rates. Therefore, it is unreasonable to conclude one exchange rate can outperform the other in measuring exchange rate volatility.
Measure of exchange rate volatility
According to literature surveys by Cote (1994), McKenzie (1999) and Auboin and Ruta (2011), the choice of exchange rate volatility measure plays a vital role in the estimation of the impact as it greatly influences the sign and magnitude of the studies' outcome. Past studies has tried numerous way to find best proxy for the exchange rate uncertainty, however, no notable method is proved to be superior to the alternatives. List of studies with regards to measure of exchange rate uncertainty is provided in appendix ?.
Standard deviation based on the level or percentage change of the exchange rate and estimation of the difference between forward and spot rate are two most employed methods in literature (Cote, 1994). However, these methods have been criticized for inappropriately handling the exchange rate data. In the later method, study argues that the results obtained from forward rate may be biased because high correlation is often found between actual rate and forward rate, hence reflecting competitiveness rather than risk (Cushman, 1986). Likewise, standard deviation is claimed to use simple descriptive data assumption, hence ignoring relevant information of random process (Engle, 1983) (Boothe and Glassman, 1987), and be arbitrary in selecting the moving average, which leads to insufficient estimate (Pagan and Ullah, 1988). For example: the measure of exchange rate uncertainty in Akhtar and Hilton (1984), one pioneering study employing standard deviation, is disapproved by Gotur (1985) for its inability to reflect long-run shifts in the exchange rate.
As indicated in studies such as Pozo (1992), Kroner and Lastrapes (1993), McKenzie and Brooks (1997), exchange rate data seems to follow the characteristic of financial time series which is often heteroskedastic, leptokurtic and exhibits volatility clustering (Bollerslev et al,1992), therefore measure such as ARCH suggested by Engle (1982) or its extended form GARCH by Bollerslev (1986) is suggested to be used since it takes into account the data characteristic and provides testing and estimation for conditional variance over time. Nevertheless, this measure also receives criticism for producing conditional variance with low correlation which leads to low level of shock persistence (Baum et al., 2004) (Klaassen, 2004). Additionally, when fitting ARCH and GARCH model into exchange rate data, researchers face several challenges to find the best fitted model for the series. According to Brooks (2008), the number of lag squared error for optimal ARCH model is difficult to determine. In some cases, the optimal lag is found to be very large, which negatively influences the parsimony of the model. In case of GARCH, although the model is claimed to be more parsimonious than ARCH, it is found that the non-negativity and unity assumption of GARCH parameter is often violated when fitting the model in time series (Bollerslev et al., 1992)l. According to McKenzie (1998), among ARCH models such as ARCH, GARCH and EGARCH tried on US-Australian exchange rate data, only ARCH(1) and ARCH(2) are reliable for estimation. The author later applies regression suggested by Pagan and Schwert (1990) to decide the optimal model for the exchange rate volatility. The finding implies that no unique form of ARCH model can immediately fitted to the data series and several ARCH forms should be considered to find best fitted model for the exchange rate volatility.
2.2.2 Modelling the relationship
Beside the selection of data and exchange rate measure, the result obtained from the analysis of exchange rate uncertainty impact on trade flows is greatly influenced by the specification of trade model and its estimation techniques. Full aspects related to modelling the relationship are presented below:
Specification of the model
In terms of export, researchers establish their models in various forms which are commonly comprised of income proxy, relative prices of export, exchange rate and exchange rate volatility. The kind of model is referred to as long-run export demand model, which assumes inelasticity of export supply and demand-determined export quantity in equilibrium as exporters have no market power. Alternative forms which takes into account market power of exporters includes additional variable such as production costs, competitor's price and capacity utilization, however, the results obtained do not differ significantly. (McKenzie, 1999). The model is used widely in multilateral studies such as Perée and Steinherr (1989), Bini-Smaghi (1991), Chou (2000) and Bredin et al. (2003).
Another model commonly used in literature is gravity model. Like the case of long-run export demand model, the gravity one can take different from at the favour of the author. The model established by Thursby and Thursby (1987) to examine bilateral trade flows is incorporated with Linder hypothesis that income per capita between two countries can negatively influence their trade of manufactured goods. To this end, national income and distance between trading partners is included in the model; additionally, its extended version can also have transport costs, consumer tastes, tariff levels and importer hedging variable. The results obtained by this model on 17 countries from 1974 to 1982 indicate statistically significant impact. Another version of gravity model constructed by Brada and Mendez (1988) to study the impact on bilateral export for 30 countries is comprised of foreign income, distance, population and bilateral trade agreements.
Estimation techniques
Empirical studies commonly undertake cross-sectional, time series and panel data analysis to investigate the impact of exchange rate variability on trade.
Cross-sectional method adopted by De Grauwe and Bellefroid (1986), Thursby and Thursby (1987) Brada and Mendez (1988), Frankel and Wei (1993), can help evaluate the level or growth rate of multilateral trade at a specific time. Although integration of structural differences is one of its advantages, it is impossible for cross-sectional analysis to capture the effect of changes in time-variant variables on bilateral trade pattern over time.
Time series data is popular in empirical literature in which OLS regression s has been the most popular techniques employed by researchers. The method is utilised in early studies such as Hooper and Kohlhagen (1978), Caballero and Corbo (1989), Akhtar and Spence-Hilton (1991), Pozo (1992); however, the results obtained are inconsistent with diverse results from negative, positive to unambiguous (see appendix ?)
In explanation to this matter, De Vita and Abbott (2004) argue that stationarity of data is neglected in previous studies. Therefore, non-stationary series are possibly used in the regression, which leads to spurious outcome and imprecise interpretation of the exchange rate volatility impact on trade. To be more specific, research by Cushman (1983) which investigates the relationship in bilateral trade for developed countries under the assumption of profit maximisation and uncertain prices is criticized by later study for not taking the stationerity of data into account. After fixing the problem by first differencing the variables, Qian and Varangis (1994) reveal a contradicting result, showing that the exchange rate exerts no significant impact on trade flow in Cushman's study. In line with this view, the essence of stationary data is emphasized in Asseery and Peel (1991) which performs seasonal testing on data of US, UK, Australia, Japan and Germany in period 1972-1987. After using Augmented Dickey-Fuller unit roots test and correcting the data's non-stationarity, the author finds exchange rate uncertainty has significant adverse impact on exports. Therefore, it is concluded that lack of consensus in empirical work is due to their inconsideration of data properties used in their models.
Another popular method often employed by studies is Vector Autoregression (VAR). The method is notable for two advantages. Firstly, VAR can handle the general dynamic relationship among data which is argued in Bini-Smaghi (1991) as one of the reason for lack of significant outcomes in literature. Another advantage of VAR is that variables involved in VAR are not subject to any theoretical restriction. Nevertheless, Koray and Lastrapes (1989) points out that as a reduced form model, VAR fails to differentiate structural hypothesis in their test of the impact for bilateral trade from US to five industrialised countries in period 1959-1985.
To examine the long-run relationship of exchange rate uncertainty and trade, studies often employ cointegration test and error correction model (eg: Chowdhury (1993), Arize et al (2002, 2003), Chou (2000), Bredin et al.(2003) However, a relatively similar problem of data characteristic in OLS is encountered in cointegration. As stated in Johansen (1988, 1991) , the variables incorporated in the model should be order one integrated or I(1). However, a great number of studies have found exchange rate uncertainty is often I(0). For example, exchange rate volatility generated from the residuals of ARIMA model in Asseery and Peel (1991) and GARCH model in studies such as Holly (1995), McKenzie and Brooks (1997) is found to be I(0). Thus, it is unavoidable that some empirical findings are subject to problem of spurious regression.
Fortunately, with the development in econometric analysis, a new method is found to resolve the problem found in OLS and cointegration. In recent study, the ARDL model suggested in Pesaran and Shin (1999), Pesaran et al. (2001) is one of the most favourable modern techniques employed to analyze the effect of exchange rate volatility on trade flows. The advanced feature of the method is that cointegration test can be performed even when the regressors are mixed in difference order, which means it allows I(0) or I(1) variables can be taken in the regression simultaneously. Since the approach does not require differencing the data, which is likely to exclude relevant information, it is argued that results derived from this technique are reliable and unbiased (De Vita and Abbot, 2004). Additionally, the long-run and short-run impact of the exchange rate uncertainty on trade can be assessed in only one model, which helps reduce complication of econometric analysis. Given its outstanding advantages, the method is highly recommended by Bahmani-Oskooee and Hegerty (2007) for further study.
Besides cross-sectional and time series data, panel data is a familiar technique used in disaggregate data. Studies such as Rose (2000), Clark et al. (2004), Sauer and Bohara (2001) and Tenreyo (2007) often employ panel data with fixed and random effect model to investigate the relationship of exchange rate volatility on trade, pointing out that the technique contributes certain advantage to the impact analysis. As demonstrated in Sauer and Bohara (2001), panel data analysis can solve the problem of unobservable variables such as structural and policy differences through fixed and random effect models. The finding conducted from 1966 to 1993 on developed and developing countries indicates significantly negative impact of exchange rate uncertainty on export in developing countries while insignificant results are found in developed ones. However, like the case of OLS, panel data is also subject to problem of spurious regression due to non-stationarity of data. Chit (2008) shows that panel-cointegration test which s essential to analyze long-run relationship has been neglected in previous studies. In addition, Baltagi (2001) points out consideration of non-stationary data should be carefully taken with a large number of countries and longer time period to avoid spurious regression. Another disadvantage is related to the assumption of fixed effect model that error terms are homoskedastic. In fact, as the countries response to unobservable shocks is time-varying, heteroskedasticy should be taken into consideration (Chit, 2008).
2.3 Summary
The chapter summarizes the theoretical aspects and methodology employed in literature. With respect to theory on the mechanism of the effect, studies have established numerous hypotheses to explain the relationship, which leads to inconclusive results of the impact's sign and magnitude. In the review of methodology, it is found that data approach, measure of exchange rate volatility and the model as well as estimation techniques can greatly influenced the finding of the impact. Given the pros and cons of each method presented, the chapter emphasizes the importance of data properties consideration when applying estimation techniques and measuring exchange rate volatility.
Considering all discussed aspects of methodology employed in empirical literature, the thesis will conduct analysis of the impact of exchange rate volatility on Japan export according to the following direction:
To avoid the limitations of aggregate data, the data used in this thesis will be sectoral export data from Japan to UK and US. With regards to measure of exchange rate volatility, both ARCH and its extended version GARCH model will be employed to indentify the best fitted model for exchange rate uncertainty based on nominal exchange rate. Additionally, one contribution of the thesis in this aspect is that mean equation will not be predetermined as in most literature but be constructed based on Box and Jenkins (1976)'s approach to find the most appropriate ARMA model for the data series The export model will be specified according to Armington (1969) , including foreign income proxy, relative price and one extra factor exchange rate volatility to estimate for the impact. Given its merit highlighted in the chapter, ARDL method will be the central estimation techniques to analyse the short-run and long-run impact of exchange rate volatility on export.
Chapter 3: Data processing and analysis
In this chapter, the thesis will describe how the data for main variables in modelling the impact is collected and processed. The structure of this chapter is presented as follows:
Section 3.1 introduces the data as well as collection criteria for the study of exchange rate volatility on export
Section 3.2 demonstrates how data is processed for further analysis as well as descriptive statistics
Section 3.3 gives concluding remarks
Data collected and classified according to Rauch (1999), presentation by Chou 2000
3.1 Data
The impact of exchange rate volatility on export in this thesis will be investigated based on bilateral trade data between Japan-US and Japan-UK. The market is selecThe thesis studies US because it is the largest trading partner of Japan from ... to... . UK is selected as it is one of Japan's top trading partners and has established a long-run international trade relationship ... The data covers period from January 1998 to June 2012.
Industrial production is used as a proxy for foreign income as its monthly data is available while GDP or GNP often quarterly or annually
We analyze the effects of real exchange rate volatility on the proportions of bilateral exports of nine categories of goods from the United States to seven major industrial countries. The nine categories of exports, corresponding to l-digit SITC (Klein 1990)
The SITC is used by Doyle (2001) ( Ireland export to UK in 1979-1992 with 2-digit 2SITC deivision levels), Cho et al (2002) (SITC 1 digit divided into machinery, chemicals, other manufacturing and agriculture)
"As the measures of export volumes, the export volume indices (seasonally adjusted) for sectors in manufactures were used at one-digit Standard Industrial Trade Classification (SITC) levels (SITC number in brackets) and includes those of Chemicals (EX5 ), Materialmanufactures (EX6 ), Machinery and transport equipment (EX7), and Miscellaneous finished manufactures (EX8). The data were obtained from the Monthly Review of External Trade Statistics of the UK National Statistics. The relative competitiveness of UK exports was measured as the ratio of the export price index of each manufacturing sector to the world price level, which is proxied by a trade weighted average of the consumer price index of the UK's 13 major export partners.The consumer price indices for the major export
trading partners were obtained from the International Financial Statistics. As the proxy of a scale variable which captures world demand conditions, a trade-weighted average of the industrial production indices (seasonally adjusted) of the UK's major export partners was used.The underlying series were also obtained from the International Financial Statistics.All the indices used in this study are based on 1995 = 100.
For analysis, we use monthly data from January 1976 to January 2000.Prior to esti-mation, all variables were transformed into logarithms.For the variables of Ext and Yt seasonally adjusted series were used, but, for RPt and Vt seasonally unadjusted data were used, since a preliminary investigation shows that they do not exhibit any seasonal fluctua-tions (Cheong et al. 2005)
Frequency
Most studies in the international trad literature have utilised daily, weekly or monthly exchange rate data. Using higher frequency data has the advantage of more observations and thus a more comprehensive information set being used to calculate variability, in particular with respect to the trends and patterns in exchange rate movements over a given period.
Risk exposure, however, may only need to be measured over a period of months, to reflect movements in exchange rates over a contract period
Currency of each variables
3.2 Data analysis
"although evidence for other countries indicates that trade is predominantly invoiced in the exporter's currency (see Grassman, 1973, 1976; Magee, 1974; Hooper and Kohlhagen, 1978 and Magee and Rao, 1980), Australian business survey data indicates that 55% of all export contracts are written against the US dollar, including the majority of commodity contracts such as oil and gold.' The impact of other foreign variables " McKenzie (1998)
Question to answer:
3.3 Summary
Chapter 4: Methodology
In this chapter, procedure of econometric analysis to investigate the impact of exchange rate volatility on export will be presented in details. The process starts with building the measure of exchange rate volatility by specifying mean equation according to ARMA process and determining the best fitted volatility model. In the later part, the export model will be indentified along with cointegration and vector error correction model test. The chapter is presented as follows:
Section 4.1: presents process to model exchange rate volatility with related tests and criteria
Section 4.2: encompasses the specification of export model and procedures to conduct cointegration and vector-error correction model
Section 4.3: gives concluding remark.
4.1 Modelling exchange rate volatility
4.1.1 Unit-root test
To avoid spurious regression, testing for non-stationary data should be performed before further data analysis. Following the framework of Klein (1990), McKenzie (1998), two pairs of exchange rate will be tested for unit-root using Augmented Dickey Fuller (ADF) developed by Dickey and Fuller (1979). The test is based on the following equation with constant and time trend:
Where Δ is the difference operator; t represents the time trend,; α, β and are the equation parameters, ε is the error term and p is number of lag order which is decided by Akaike Information Criterion (AIC)
The null hypothesis of the test is that data has unit root. We can reject the null hypothesis if the test statistic, presented as the t-statistic, exceeds the critical value provided by Dickey and Fuller (1979) in absolute term. If the level I(d) of the series is non-stationary, the data will be differenced to higher order of integration until the null hypothesis can be rejected. To this end, we can use data for further tests.
4.1.2 Autocorrelation test
After the condition of stationary data is confirmed, the next stage is to inspect the autocorrelation and the partial autocorrelation of the series. Test results obtained in this step will help determine the appropriate structure of ARMA model for later volatility measure.
According to Box and Jenkins (1994) the autocorrelation, also known serial correlation, measures the correlation of times series with its own future or past observations. The Autocorrelation Function (ACF) at lag k is calculated as follows:
Where: is the ACF at lag k. and are observations at time t and t-k. s the covariance between and . is variance at time t.
The partial autocorrelation function (PACF) at lag k specifies the correlation between current observation and the observation in the previous k period after managing the impact of correlation of observations between them. The PACF at lag 1 is similar to the ACF, however, from lag 2 onwards, the PACF is computed as follows:
Where: is the PACF at lag k, and , are autocorrelation coefficients at lag k and k-1.
The null hypothesis of the test is that autocorrelation or partial autocorrelation coefficient at lag k equals to zero. If or exceeds the range of , where T is the sample size, the null hypothesis can be rejected at 95% level of confidence. The significant lags obtained from ACF and PACF are essential to determine the lag order of moving average (MA) and autoregressive (AR) for ARMA model respectively in later part of the thesis.
Another techniques to test autocorrelation is using Ljung-Box Q-statistic by Ljung and Box (1978). The statistic is determined as follows:
Where is the autocorrelation coefficient; T is number of observations and m is the maximum lag length.
According to the study, the null hypothesis that autocorrelation coefficients are jointly equal to zero will be rejected if Q-statistic exceeds the critical value at high level of significance. Following Kenourgios and Samitas (2008), the thesis will perform Ljung-Box test at lag level of 10, if no autocorrelation is found, the lag will be increased to 15 and 20. The reason behind this is that small number of lags can fail to capture autocorrelation while too many number of lags can cause dilution in the significance of autocorrelation at one lag. (Harvey, 1993)
4.1.3 ARMA model specification
The ARMA model, often referred to as ARMA (p q), is comprised of two main parts: the autoregressive (AR) of order p and the moving average (MA) of order q. The model is presented by the following equation:
Where is observation at time t. p and q are lag order of AR and MA process, respectively. , and are the parameters of equation, ε is the error term.
After the stationarity of data and autocorrelation are confirmed, the ARMA can be employed to specify the mean equation from exchange rate data series. The model construction will be based on Box-Jenkins method suggested by Brooks (2008) and Gao and Kling (2005). At first, the maximum lag order for p and q will be determined by the lag number that exhibits significant autocorrelation and partial autocorrelation, given that all previous lags also have autocorrelation and partial autocorrelation at significant level. Based on this finding, various ARMA model will be generated with alternative p and q to find the best fitted one for data series. Brooks (2008) and several studies such as Baker et al. (2008) and Gujarati (2003) suggest the use of information criterion to select the appropriate model. According to this approach, the model which minimizes information criteria will be selected in further estimation.
The thesis will consult three major information criteria namely: Akaike's information criterion (AIC), Schwarz's Bayesian information criterion (SBIC) and the Hannan-Quinn information criterion (HQIC). The estimation equation for each information criterion is found below:
Where: RSS is residual sum of squares. T is the number of observations, k is the number of parameters estimated in total.
4.1.4 Testing for ARCH Effects
After the suitable ARMA model is specified for the series, a test for ARCH effect , known as ARCH Largrange Multiplier (LM), is performed to confirm the heteroskedasticity in the residuals of the process. Through the test, the following regression of squared residuals is calculated:
Where: k is the number of lag, is the squared of residual from ARMA model.
The test null hypothesis is that the squared residuals exhibit no ARCH effect up to the order of lag k. The test statistic, which is computed as the product of obtained from the above regression and the number of observations, follows a distribution. We can reject the null hypothesis if the test statistic exceeds a critical value of at a certain level of confidence.
The ARCH LM test is essential to the adoption of volatility model such as ARCH and GARCH. Not only does it provide information to select the lag order in the model, it is also a post-estimation test to evaluate the efficiency of the model generated. As indicated in McKenzie and Brooks (1997), if there still exists ARCH effect in the residual of volatility model detected by ARCH LM test, the lag order in ARCH, GARCH model can be adjusted to higher lag until the null hypothesis is totally rejected, which means no ARCH effect left in the model. The test will be initially conducted at lag 1, if ARCH effect is detected, ARCH (1) or GARCH (1,1) will be estimated. After each estimation, the test procedure will be repeated to see whether the model parameters should be improved or not until no ARCH effect is found in the model.
4.1.5 ARCH and GARCH model specification
The variance equation of ARCH (q) introduced by Engle (1982) has the following form:
Where is the conditional variance at time t, is the residual from mean equation, q is the number of lag order, and β are parameters of the model subject to non-negativity assumption ( >0, β >0) .
According to Brooks (2008), ARCH model is subject to certain pitfalls. In particular, the number of lag order q can be very large, which adversely influences the model's parsimony by generating large conditional variance. Another drawback is the condition that the parameters are greater than zero is often violated as more parameter is added to the model.
In such case, GARCH model developed by Bollerslev (1986) is stated to resolve the problems encountered in ARCH. According to GARCH, the conditional variance in ARCH model can also be explained by its own previous lag order, which is presented as follows:
Where , and must be greater than 0, the sum of and should be less than 1 so that the process can be stationary
Following McKenzie (1998) in finding the appropriate ARCH, GARCH model for exchange rate volatility, when more than one model is found to fit in the series, the best fitted model which has most predictive power will be determined by Pagan and Schewrt (1990)'s approach. According to the studies, the model which has the highest R2 value obtained from the regression below will be selected to estimate exchange rate volatility:
where is the squared residual from the mean equation at time t. is the fitted value of conditional variance, and is the parameter of the equation, is the error term.
4.2 Export model specification
4.2.1
Export demand model: (Arize et al. 2000)
Where X is the sectoral real export of Japan to UK or US, Y denotes income of UK and US for the sector which is proxied by industrial production, P denotes the sectoral relative prices of Japan to UK or US, V is the terms for measure of exchange rate volatility adopted in the model and is the disturbance term.
The model assumption
Based on the equation, the long-run relationship for export demand can be determined. (Bailey et al., 1986). According to traditional theory, a rise in foreign income will increase the demand for export, therefore, the sign of its coefficient is expected to be positive. Conversely, the relative prices' coefficient is believed to be negative since an increase in export prices can discourage foreign countries to import. (Arize, 1995). The coefficient of exchange rate volatility, as discussed in literature review is uncertain.
4.2.2 ARDL approach
Chapter 5: Empirical results
summary
5.1 Measuring exchange rate volatility
5.1.1 Unit-root test
After performing ADF on two pairs of exchange rate, it is found that both series have t-statistic exceeds the 1% critical value, which is -4.01262 for JPY/GBP and -4.01361 for JPY/USD, in absolute term. Therefore, the null hypothesis cannot be rejected and a higher level of integration should be considered.
Following Chou (2000), we generate the first deference from the natural logarithm of the two series as follows:
Where Et is the exchange rate at time t.
After being tested for unit root using ADF, the new series data is confirmed to be stationary. The t-statistic for first difference of JPY/GBP and JPY/USD both exceed the 1% critical value in absolute term at 1% level of confidence as the statistic probability is smaller than 0.01.
Summary of test results are demonstrated in the below table while the complete results can be found in appendix 3.
Table 5.1: Results of unit-root test exchange rate
5.1.2 Autocorrelation test
After the stationarity of data is confirmed, the Ljung Box test is performed along with analysis of autocorrelation and partial correlation. As the number of observations is 174, the critical range which is the benchmark for the choosing significant lag is = 0.149.
According the correlogram graph presented in appendix 4 as the p-value is smaller than 0.01 and 0.05, we conclude that Q-statistics for 10 lag exceed the critical value at 1% and 5% level of confidence for JPY/GBP and JPY/USD respectively. Therefore, the null hypothesis of no autocorrelation can be rejected.
In attempt to indentify the appropriate lag for ARMA model in later process, the autocorrelation and partial correlation of each lag is summarized as follows:
In case of UK, the autocorrelation is found statistically significant in lag 1 (0.343), lag 2 (0.203), lag 5(-0.175), lag 6 (-0.192), lag 9 (0.153) and lag 10 (0.170) while the partial autocorrelation is shown in lag 1 (0.343), lag 4(-0.246), lag 8 (0.142) and lag 9 (0.118).
With regards to US, we detect significant autocorrelation in lag 1(0.213), lag 5 (-0.276), lag 6 (-0.194), lag 9 (0.142) and lag 10 (0.196); the partial autocorrelation is found significant in lag 1 ( 0.213), lag 5 (-0.232) and lag 8 (0.116.
5.1.3 Specification of ARMA model
As demonstrated in chapter 3, the alternative ARMA model will be estimated from ARMA (0,0) to ARMA (p,q) to find the most appropriate model which minimizes information criteria. The order p and q is determined by analyzing significant lags found in the above autocorrelation test.
For UK, as the autocorrelation and partial autocorrelation is consecutively significant up to lag 2 and 1 respectively, value of information criteria AIC, SBIC, HQIC is considered from ARMA (0.0) to ARMA (1,2). Though ACF and PACF are still significant at higher lag, we ignore autocorrelation from lag 3 and partial autocorrelation from
lag 2 onwards as these lag lie within the critical range.
According to the table in appendix 5, the result is consistent in AIC, SBIC and HQIC that AR (1) model can minimize information criteria. Using Least Square method, the model is specified as follows: (see appendix 5)
Xt= -0.00319 + 0.3446 Xt-1 + εt (5.1)
Similarly for US, lag 2 onwards in both autocorrelation and partial autocorrelation will be omitted as the coefficient in lag 2 does not exceed the critical range in absolute term. In this case, only ARMA(0.0) to ARMA (1,1) will be evaluated to find the best fitted ARMA model. Based on the table of information criteria in appendix 5 all criteria indicate that MA(1) is best fitted ARMA process for the series. To this end, the model is identified using Least Squares method as follows:
Xt= -0.0028 + 0.193891 εt-1 + εt (5.2)
Following in the procedure in chapter 4, the residual of these two equation will be test for ARCH effect in the next step.
5.1.4 ARCH effect test and selection of volatility model
After the ARMA structure of mean equation is indentified, the ARCH LM test is conducted on the squared residuals of equation 5.1 and 5.2. The results from table 5.2 show that the F-statistic from equation AR(1) and MA (1) are significantly greater than the critical value at 1% level of confidence as probability is smaller than 0.01. Thus, we can reject the null hypothesis that there is no ARCH effect at lag 1 for both equation.
Based on this result, volatility model such as ARCH (1) and GARCH (1,1) is fitted into the data.. After estimating the parameters for variance equation, ARCH LM test is again performed to detect ARCH effect in the residual of each model. As presented in table 5.2 , the null hypothesis of no ARCH effect cannot be rejected because the probability of F-statistic is greater than 0.05. Therefore, it can be concluded that ARCH (1) and GARCH(1,1) model is sufficient in capturing the volatility movement and no higher lag order of the model should be considered.
Table 5.2: Result of ARCH LM test
Studying the parameter of ARCH(1) and GARCH(1,1) estimated in two markets, the thesis finds that all the coefficients generated follow the constraint of ARCH and GARCH model. As shown in table 5.3, coefficients of ARCH (1) term are positive and statistically different from zero at 1% or 5% level of confidence. The same result is found in the coefficient of GARCH (1,1), additionally, the sum of coefficients is smaller than 1, which means the process is stationary. Details of these model are shown in appendix 6.
Table 5.3: Coefficients of ARCH (1) and GARCH (1,1) in two markets
To choose the best fitted model for measure of exchange rate volatility, the residual and conditional variance of ARCH (1) and GARCH (1,1) are regressed based on the model suggested by Pagan and Schwert (1990). The R2 values obtained from the regression of ARCH(1), GARCH (1,1) in two markets presented in table 5.4. is a benchmark to compare efficiency of the two models. As ARCH (1) has higher R2 value than GARCH (1,1), the thesis determines that ARCH(1) is the appropriate measure of exchange rate volatility for later economic analysis. The regression details can be found in appendix 7.
Table 5.4: Pagan and Schwert (1990) R2 value
ARCH(1)
GARCH(1,1)
JPY/GBP
0.272636
0.152648
JPY/USD
0.033664
0.023047
Accordingly, the volatility of for each pair of exchange rate is estimated as the fitted conditional variance from the model.
Chapter 6: Conclusion
Unlike previous study use ..., the thesis use.... the contribution is....
Look at Cheong et al for policy implication.
This would seem to suggest that sectoral differences do exist in explaining the different impact of volatility on trade and may be based on the characteristics of the markets in which they trade. We believe this is an important finding and may be the key to understanding why so many studies have not found a clear-cut empirical relationship between exchange rate volatility and trade when using ag gregate trade data. It also suggests that a greater degree of disaggregation, at the industry, product or firm level, may provide further worthwhile results
Reference
Appendices
