^ tUBRAPJSS "^cy T« > ••••^a HD28 .M414 no. ALFRED P. WORKING PAPER SLOAN SCHOOL OF MANAGEMENT High Frequency Data and Volatility in Foreign Exchange Rates by Bin Zhou MIT Sloan School Working Paper 3485-92 September 1992 MASSACHUSETTS INSTITUTE OF TECHNOLOGY 50 MEMORIAL DRIVE CAMBRIDGE, MASSACHUSETTS 02139 High Frequency Data and Volatility in Foreign Exchange Rates by Bin Zhou MIT Sloan School Working Paper 3485-92 September 1992 M.I.T. NOV LIBRARIP? 1 3 1992 High Frequency Data and Volatility in Foreign Exchange Rates Bin Zhou Sloan School of Management Massachusetts Institute of Technology Cambridge, MA 02139 September 1992 Abstract: Exchange rates, like substantial heteroscedasticity. many other financial time series, display This poses obstacles in detecting trend and changes. Understanding volatility becomes extremely important in studying financial time series. Unfortunately, estimating volatility data, such as daily, weekly, or monthly observations, is from low frequency very difficult. Re- cent availability of ultra-high frequency observations, such as tick-by-tick data, to large financial institutions creates a volatile time series. Tick-by-tick new possibility for analysis of data provides us a near continuous observa- much tion of the process that gives us potential to study volatility in However, high frequency data has extremely high negative correlation in its return. In this paper, and US Dollar (DM/$) exchange we use tick-by-tick rates to explore this first detail. order auto- Deutsche Mark new type of data. propose a model to explain the negative autocorrelation and a volatility timator using the high frequency data. DM/$ We es- Daily and hourly volatility of the exchange rates are estimated and the behaviors of the volatility are discussed. Key Words: Financial time series; tick-by-tick data; heteroscedasticity; Introduction 1 There is considerable literature analyzing the behavior of exchange rates. However, structural exchange rate modeling has not been very successful. studying monthly data, Meese and RogofF (1983a,b) have shown that By a random walk model at least as well as fits more complicated structural models. Empirical study (Hsieh 1988) has shown that daily returns are approxi- mately symmetric and leptokurtic are weak but not independent and heavy tion of the (i.e., heavy tailed). is [8] ([6], simple. The argument [4]). The amount is for mean and is is explana- indepen- variance change changing variance of returns of "information flow" that cause not constant over time. Hence there of the price changes One the hypothesis that data is dently distributed as a normal distribution whose over time autocorrelations identically distributed (iid). tailed distribution and The change in prices no reason to beheve that the variance constant over time. Clark (1973) and many others (Mandelbrot and Taylor 1969, Praetz 1972) have argued that observed turns come from a mixture of normal distributions. Xt denotes the daily return of the given information cjt is all number the If random the re- variable price, the conditional distribution of ,Y( is: XtliJt where is -^NifiJiu,)) the information available at time of transactions (Mandelbrot (1) t. The quantity ujt could be and Taylor 1969), or trading volume (Clark 1973). One parametrization ied by of this conditional heteroscedasticity was first stud- Engle (1982) where: p fiut)=ao + J2a,{Xt-,-fi)\ ao>Oa,>0, i = l,...,p (2) «=i It is called the autoregressive conditional heteroscedasticity since the heteroscedasticity Because the in financial ARCH represented in an autoregressiv'e fashion. ARCH model exhibits the conditional heteroscedasticity present time series and used to analyze is (ARCH) model many is mathematically easy to manipulate, financial time series. Previous studies model provides a close it has been found that the approximation to many financial time series. Since then many other parametrizations, such as the generalized autoregres- sive conditional heteroscedasticity (GARCH) model (BoUerslev 1986), have been proposed. They captured some characteristics of the volatility such as But there also exists problems. The parameters need to these models use historical data. The volatility estimates volatility clustering. be estimated. All are often lagged. The data has opened up new possibility availability of high-frequency estimating the volatility. in Tick-by-tick data provides us with a near contin- uous observation of the process. It gives us potential to study volatility in great detail. Understanding volatility is the key issue in the conditional het- model (1) and any other financial time series model. This paper explores tick-by-tick data and uses the data to estimate and study the eroscedasticity volatility. High-Frequency Data 2 Because of growing computer power, gathering financial data fast than ever. Data began is easier no longer recorded daily or weekly. Many large institutions to collect so called tick-by-tick exchange rate in the early eighties. Different from stock market, the foreign exchange market has graphical location, and no "business-hour" limitations. tiated is and traded over the telephone. volume are not known to the public. The no geo- deals are nego- The transaction prices and trading The exchange rates used for most re- search are the quotes from large data suppliers such as the Reuters, Telerate, or Knight Ridder. pliers. Any market maker can submit new The quotes then are conveyed to data subscriber's screens. suppliers cover the market information worldwide day. The quotes quotes to the data sup- The data and twenty-four hours a are intended to be used by market participants as a general indication of where exchange rates stands, but does not necessarily represent the actual rate at which transactions are being conducted. some participants It is to manipulate indicative prices occasionally possible for and create a favorable market movement. However, since a bank's reputation and credibility as a market maker emerges from favorable relations with other market participants, it match the true is generally felt that these indicative prices would closely prices experienced in the market. A reader who is unfamil- Table PUB. 1: A Sample of Tick-by-tick Exchange Rates Table n 2: Summary Statistics of Tick-by-tick Returns Figure 1: Original Quotes From the Reuters and Telerate adu Figure 2000 2: Validated Quotes 4000 sooo ddu BOOO Figure 3: Sample Kurtosis of Return of n-ticks I 600 400 200 ddu unstable and then decreases. The last stage, decreasing kurtosis, is due to increasing negative autocorrelation in data (see Figure 4). Although we expect a slightly negative first lag autocorrelation as ported in other exchange rate literature, a in tick-by-tick return —47% it re- negative autocorrelation was surprising. To be cautious, we also calculated the The auto- correlation coefficients are —0.4718, —0.4691, —0.4665 and —0.4632. The autocorrelations in four subgroups according to four quarters. negative autocorrelation is consistent. assumed for low frequency data. Our any fundamental difference between the high and low not follow a Brownian motion as question is if there is Obviously, high frequency data does it is frequency data. After further studying the data, we find that the difference in high and low frequency data is level of noises. The noise is neghgible in low frequency data, but becomes very significant in high frequency data. noise may come from many different sources. For example, it The could be round As we know that round off Brownian motion introduces negative autocorrelation in its return. There are also updating noises in quotes. To be visible in the market, traders keep off error. All financial data is quoted in finite digits. Figure First-order Autocorrelations of Return of n-ticks 4: k ^,.^M*#**^ I 400 200 600 updating their quotes. The new update previous quotes even the market is 800 1000 often sHghtly different from the Typographical errors are another is still. source of noise. Summarizing these arguments, we assume following process for the exchange rate: = S{t) where S{t) is dit) + + B{Tit)) logarithm of the exchange rate, B{-) motion, both d{-) and 6^ the Brownian motion B{.). is the standard Brownian are assumed deterministic functions, r(-) has t(-) and positive increments, (3) tt is the The mean noise, random zero Cj, is noise independent to combination of several sources which were mentioned above. This extra noise causes most of the negative autocorrelation in high frequency data. = Let X{s,t) S{t) — X{s, where Zt fi{s, t) = is S{s), the return in interval t) = fiis, t) a standard normal d{t) — d{s). The Var(X(3, 0) + ais, t)Zt random ^'(5, t) + variable, a^{s,t) 8 r]\t) + Then + tt-e, variance of the return = [s,t]. (4) = r(f) is: t]\s) + 2c{s, t) — t{s) and where = r/^(0 and Var(e() like cT^{s,t) diminishes. sample random walk. The return, X{s,t), a high-frequency data, the noise -47% \s — t\, noise When |5 — <| is for the DM/$ is — \s t\ increcises, becomes negligible and decreases to near zero, the difference of two noises. order autocorrelation of such series first When Cov(e,,et). For large cr^{s,t) increases as well. X{s,t) behaves = c(s,t) is about —.5. When we The study no longer negHgible. An autocorrelation of exchange rate indicates that level of noises is very high in tick-by-tick data. There are several difficulties is lack of information about r(<) which volatility. (7^{t — = 6,t) T{t) — r(< — or simply 6- volatility. Next section, volatility for 3 analyzing the process difficulties in 6) we call (3). it One of the the cumulative called the ^-increment of volatility is we devote our attention to estimate the any increment. Volatility Estimation we concentrate on estimating the volatility of a given time T{n) — t(0). The function T{t) can be estimated increment by in- In this section [0,n], crement. We first derive an optimal estimator of the volatility based on assumption of constant variance and zero mean. Then we generalize the sults to general process (3). re- Proofs of the theorems are listed in appendix I. Theorem from 1 Assume that {S{t),t 0, l,...,n} is a series of observations the process S{t) where = et,t = distribution l,..,n, are and T[t) = = B{T{t)) + e, (5) independent and identically distributed with normal a^t + b. Let Xt = S{t) — S{t — I). Then the maximum likelihood estimator of a^ is a'MLE = {l/n)±{X! + 2X,X,_,-^,)^-^^-^ where P = ^ ana p = — (6) MLE This An close. ^k^ However, we noticed that p and p' are very unbiased estimator can be obtained by eliminating the factors not unbiased. is and 4: a'u = {Un)Ei^f + 2X,X,.,). (7) 1 Theorem Under 2 Theorem the assumptions of I, mean and variance the of the estimator {!) are: Ea^u = ^^ Var(<7\) = {\/n)a'i6 (8) and From we (9), we apply estimator cLSsume that n is multiple of Theorem — S{i — k),i = The variance is is r? (l/n)a''(6^ minimized ..., n where we (A7,,+2.Y,,X._,,,). (10) given in following theorem: 3 Under the assumptions of theorem = k, 2k, i=k,2k,...n unbiased and the variance Var(a2f;,fc) X down S{i) Let E = " is k. Since aggregation increases the variance = (7) to Xi^k (t2 u,k Estimator (10) rj'^/cr^. (9) be optimized by properly find that variance of a'^u can adjusting the variance ratio (7^, + ie^ + S^^ + ^'/n\ at k = + [-Jj^] IG-^ or k 1 n^ + S/-) + V/"' = [-j^^] + 1, where (H) [x] rounds to the next integer. Since the noise ej is independent, the variance of the estimator can be further reduced by averaging the estimator (7) at different start points: <^' = E(A'.^ + ^ kn ,^, 10 2X,^k.kX,.2k.k) (12) In fact, can be proved that: it + 8-j-^) + ^'/n^ 16^ K O Kff < -a\6k + Var(a2) ft The above three theorems assumed However the estimator can that we have observations Then the volatility r{<„) — ^(^0, in) also {5(<,),? r(<o) can 4 —'2k^—'2k = + l,...,n} from process (3). be estimated by: 2X,_,,,X„2k,k) (14) .=1 Assume Cov(ti,ti-k)=0 for E\/(<0,<n) in = ] f^iXlk + ^ Theorem = and constant variances. a more general case. Suppose noises iid be used (13) alii. Then T{t^)-T{to) k-l + ^(i/^')[a2(f,_i_;t,<._fc)-cr2(<„_,+i, <„_,)] i=l 1=0 n +{\lk)Y,[^i\U,U_,) + 2ti{ti,U_k)ti{ti_k^,_2k)] (15) :=0 where r]^{t) = Var{et). Only assumption we made in this theorem is the uncorrelated noises. can be any increasing function and tt is not necessary stationary. Since T{t) n J2Wit„U_k) + 2fi{t„ t,_k)Kt,-k, t,.2k)] < 3 max{|/.(i„ U-k)\Mn) - d{0)] i=k the last term in (15) smooth. Therefore, if is negligible in high frequency data for large n, the no jumps occurred in estimator (14) is if the drift d{t) approximately unbiased time interval [^o^n]- This estimator be updated when new data becomes available. the volatility T{t) dynamically. 11 It is is also easy to will allow us to estimate Estimating Volatility of Exchange Rates 4 In this section, rates. we apply the The data volatility estimator (14) to DM/$ exchange has one discontinuity which was in the week of August 13 when the data base was shutdown due to a power outage accident in lower Manhattan area. The return of that week is set to be zero and so does the volatility. we estimated the variance ratio rj^/a^ to be approximate 6. Minimizing upper bond of variance (13), we have k = 6. Using all available tick-by-tick data, we estimate the volatility of DM/$ exchange rate in entire 1990. The estimate is .010349. To verify this estimate, we compare it to the estimate under the Brownian motion assumption. If the data had no noise and followed a Brownian motion, the quadratic variation To choose the parameters k, {n/k) Qk= Yl x{Uk-k.tikf i=k would be a standard estimator of the is used on data with noise, it volatility. When the quadratic variation overestimates the volatility. along with the sample frequency. The expectation The bias decreases of the quadratic variation is: {n/k) 1=1 {n/k) i=l where c{t,f.,t,k-k) = When Cov(e(,^, e(_^_^). e('s are uncorrelated and /I's are negligible, {n/k) EQ, =^r(^„)-r(<o) + 2 ^tjI (17) 1=1 which decreases as k increases. We axes have a logarithmic scale. When variation is is k = Qk against k in Figure the frequency about the same as our estimate except caused by small sample When plotted I, size. is is 12 Both low, the quadratic for a high variation, In high frequency, the bias the quadratic variation 5. is which tremendous. about thirteen times the size of our Figure 3 Quadratic Variations Using n-tick Returns 5: . s > tidu estimate. Therefore, from (17), the total variance of the noise about is six times the total volatility which confirmed our early estimate of the ratio. To estimate daily volatihties, we need a day since the foreign exchange market market. We is time. GMT is is much less. is as a 7:00pm is more than 7,000. Quotes in For small n, volatility estimate from (14) could be negative. In such a case, we The GMT This twenty-four hour period covers most activities of the weekends or hoHday The Mean Time (GMT) 9:00am Tokyo time and 24:00 world market. Average weekday daily ticks volatility. and the end of a twenty four hour international choose 24 hours from 0:00 Greenwich day because that 0:00 New York to define the start let the estimate to zero to avoid negative daily volatility estimates of DM/$ are plotted in Figure 6. and Aubank gust 2, 3, 1990. On January 4 and 5, the German central surprisingly intervened the foreign exchange market and pushed the dollar lower. On January 30, a wide market followed a rumor that Mr. Gorbachev was consix largest volatilities were on January 4, sidering resigning as secretary of the former Soviet 5, 30, July 12 Communist Party. July 12, the dollar tumbled because possible lower interest rate by the 13 On US Figure 6: Daily Volatility Estimates of 1990 DM/$ 9 a: •r a iiiiiJliiiiltl 1990 J990 Mu On August Federal Reserve. IMO 1990 1990 1990 990 990 ha ;990 hi Auf Scy SI>T Dec Mt/ and 2 3, 1991 Ju the Dollar had another wild ride as the news of Iraq's invasion of Kuwait spread around the world. However, the large volatility does not always have large price change. for above six respectively. The daily change days are -0.0537, 0.0162, 0.0212, -0.0178, 0.0058 and -0.0074 On August 2 and the exchange rate only changed 58 and 74 3, points which are about the average. When we study low frequency data like daily price, noise becomes negli- and the price change, Xi, approximately normal distributed with mean zero and variance erf, or rescaled return ¥{ = X,/ai is a standard normal gible random variable. Therefore we can the normality of rescaled return on Saturdays), we plot return Yi in Figure Table 3. The 95% 7. Y,. test our model assumption (3) by testing Excluding zero volatility estimates Q-Q normal plots The basic statistics for both return A", Xi and Yi of both confidence intervals of the moments (all and rescaled are shown in of Yi are given in the parentheses. Table 3 also shows the Kolmogorov-Smirnov goodness-of-fit test for normality. Comparing the conclude that much Yi is statistics in closer to having a 14 column A', and column normal distribution. It Yi, we indicates Figure Q-Q 7: ReKiled Daily Returns Daily Returni -2-10 QntnolM «f 2 1 Stindud Monnil Table 3: Plot of Daily Returns -2-10 Q^ittdu of 1 Srindard Konnftl Basic Statistics of Daily Returns 2 Table 4: Average Daily Volatility . Figure 4 soE'O: 8: Average Daily Volatilities -r Figure 3 soE'Oe . 3 00E-08 .. 2 SOE'Ot . 2 ooE-oe 1 50E-09 9: Average Hourly Volatilities . 8 8 8 8 8 8 S 17 8 8 8 8 S S 8 Figure 10: Q-Q Plot of Hourly Returns Conclusion 5 High frequency data can be described as a Brownian motion with a This noise brings a strong data. The first noise. lag negative autocorrelation in high frequency autocorrelation decreases as frequency decreases since the role of the noise reduces. High frequency data can be used to estimate the volatility in high frequency in reasonable precision. Different from other volatility estimator, our volatility estimator mainly uses the data within the period we are interested instead of historical data. This allows us to capture the market volatility quickly without delay. The estimate is nearly unbiased when price has no jump. Since trading volume and volatility are highly correlated, this is more important in exchange market than in unknown trading volume. Besides of many other applications of the volatility, we are more interested type of volatility estimation other market because of in modeling and forecasting the time series. Volatility the return to address the heteroscadesticity can analyze the data in volatility scale in eliminate the heteroscadesticity. We like 19 ARCH model. We also stead of calendar time scale to will discuss future papers. the can be used to rescale more of these issues in our ^ Appendix • Proof of Theorems I: Proof of Theorem 1: Under assumptions of the theorem, = Xt is where p random a normal = 77^ variable with the variance of is —Tj^/v^. The Notice that Xt\Xt-i L{s\ + e, mean - c,_i zero and variance v^ Xt has the £(. likelihood function of Xt = L{s\p;Xu...,Xr,) variance s^ aZt = i'^(l p; X, , is first = (7^ + 2r/^, lag autocorrelation is /(A'„|X„_i)/(X„_i|X„_2).../(.Yi|Xo) also a normal variable with mean pXt-i and — p^), we have ..., X„) = (27r.^)-"/^exp(- t ^^^^^^). — (=1 The log-likelihood function i{s\p;X,,...,X„\Xo) The is = -^M2;r.^)-X: ^^'7f-^^' partial derivative of the likelihood function with respect to p di dp ^^ iXt-pXt-,)Xt-i h Setting the derivative equal to zero and solve for « P Similarly, for 5^, _ - Y^^-\ r-n /), XjXt-i v-2 we have d^ ^ n ^ {Xt-pXt-,Y ^ 20 we have is or s 1 ^ = -j2{Xt-px,_,f 2 "«=i ^ Substituting p hy p m (=1 (=1 t=l above formula, we have " (=1 (=1 = -t-'^ii^-PP') where It is easy to show that a' Therefore, the = v'{l+2p) = s'{l+2p)/{l-p') maximum = • Proof of Theorem hkelihood estimator of a^ i|:,.v?..v,A-,..A,lL^ 2: = (l/n)f;(Var(A^) + 2EX,X,_i) E{l/n)f2iXf + 2X,Xt-,) 1 1 = {l/n)±{a' + 1 and Vara\ = is (l/n)2Var X:(A? + 2A^X,_i) 1 21 2rj^-2r,^) = a\ 1 +2aZt-xtt - 2aZt-itt-x n a^ Proof of Theorem Since VarX,,i — Ws.v{ka^u,,) 2e(e(_2 + 2e,_ift_2) + a^ 3: ka^ (9) implies that , = {k/n}{ka')\6 + 16-^ + Var(aV) = 8-^) + ^Vin/kf (l/n)(<72)2(6A: + IG^ + sf^) + V/n^ ka* CT'' • variance reaches Proof of Theorem EVito,tr.) = minimum when 4: (\/k)J2[a'it,_,J,) +fl'^{t„t,_k) = - n^ or The e^ n^ n • - [r{tn) - + + r]'{U)-ri'{U_k) 2^l{t„t,_k)fi{t,_k,t,_2k)] T{to)] k-1 1 =1 k + {Uk)j:iv'{t,-k)-v'itn-.)] 1=0 n 1=0 22 tl] Appendix II: Validation Program There are many reasons to have outliers in original data set shown in Figure 1. Most quotes are typed in by human, there are unavoidable keying errors. Most outliers we found are this type of errors. Outlier also could be caused by large bid and ask spread. In such case, at least one of bid or ask and becomes an Following program price does not reflect true market price electronic error also makes outliers. outlier. Occasionally, is designed to remove above three types of outhers: A i) ii) iii) quote is considered as an outlier and removed from the time series a rate is more than bid and ask spread a rate is 5 or less than is 1, more than 50 above or below than its if or points, or neighbor prices more than a certain threshold. (iii), we of the data. If For carry out two regressions using ten bid prices on each side the current bid price both regressions, it is is higher or lower than c-point from considered as an outlier, where c is range from 15 to 30 points dependent on the variances of the neighbor points. Whenever an is detected, we go back ten steps and repeat above procedure. About 0.73% of data have been removed by this validation program. The first ten removed data due to (iii) are listed in Table 6 with possible outlier explanation of error. 23 Table 6: The First Removed Data Ten Outliers Removed From 1990 DM/$ [6] Friedman, Daniel and Stoddard Vandersteel (1982), "Short-run fluctuation in foreign exchange rates." [7] Goodhart, C.A.E. and Money and [11] Karatzas, loannis and Steven E. Shreve (1988), Brownian Motion and New York. Mandelbrot, B. and H. Taylor (1969), "On the distribution of stock price differences." Operations Research, 15, 1057-1062. Meese, R. A. and K. Rogoff (1983a), "Empirical exchange rate models of the seventies: do they [12] fi- rates: 1974-1983." J. of Inter. Econ., 24, 129-145. Stochastic Calculus, Springer- Verlag, [10] in Finance, 10, 23-52. Hsieh, David A. (1988), "The statistical properties of daily foreign ex- change [9] "Every minute counts L. Figliuoli (1991), nacial markets." J. of Inter. [8] Econ., 13, 171-186. J. Intern. fit out of sample?." J. of Inter. Econ., 14, 3-24. Meese, R. A. and K. Rogoff (1983b), "The out of sample failure of empirical exchange rate models: in J. Frenkel, ed.. sampling error or misspecification?." Exchange Rates And International Microeconomics, Chicago: University of Chicago Press. [13] Praetz, P.D. (1972), "The distribution of share price changes." Journal of Business, 45, 49-55. [14] Taylar, Stephen (1988), Modeling financial time series, Sons. 25 John Willey & V928 122 Date Due "Wr 2 7 1991 7\9^ OCT.?- Lib-26-67 MIT LIBRARIES .DUPL 3 9080 02879 8350