E-Leader Tallinn, 2009

FINANCIAL MARKETS DURING ECONOMIC CRISIS Mária Bohdalová and Michal Greguš Comenius University, Faculty of Management Bratislava, Slovak Republic

Abstract Contemporary financial depression of the financial markets proves high fluctuations of the prices of the stocks. These fluctuations have considerable impact on the values of the financial portfolios. Classical approaches to modeling of the behavior of the prices of the stocks may produce wrong predictions of their future values. That is the reason why we introduce in this paper the fractal market analysis. Fractal structure accepts global determinism and local randomness of the behavior of the financial time series. We will use R/S analysis in this paper. R/S analysis can distinguish fractals from other types of time series, revealing the self-similar statistical structure. Key words: financial time series modeling, fractal, R/S analysis, Hurst exponent

1.

Introduction

The financial markets are an important part of any economy. For that reason, financial markets are also an important aspect of every model of the economy1. Markets are “efficient” if prices reflect all current information that could anticipate future events. Therefore, only the speculative, stochastic component could be modeled, the change in prices due to changes in value could not. If markets do not follow a random walk, it is possible that we may be over-or understanding our risk and return potential from investing versus speculating.

2.

Introduction to Fractals and the Fractal dimensions

The development of fractal geometry has been one of the 20-th century’s most useful and fascinating discoveries in mathematics ([2], p.45). Fractals give structure to complexity, and beauty to chaos. Most natural shapes, and time series, are best described by fractals. Fractals are self–referential, or self–similar. Fractal shapes show self–similarity with respect to space. Fractal time series are random fractals, which have more in common with natural objects than the pure mathematical fractals we will cover initially. We will be concerned primarily with fractal time series, but fractal shapes give a good intuitive base for what “selfsimilarity” actually means. Figure 1 shows daily and weekly Bank of America Corporation prices2 for consecutive observations from march 2007 to may 2009. With no scale on the X and Y axes, we are not able to determine which graph is which. Figure 1 illustrates selfsimilarity in a time series.
1
2

http://www.economymodels.com/financialmarkets.asp Data were retrieved from www.yahoo.com

p. 1951.50. Instead. or degree of freedom.E. then classical geometry offers little help. Time series is only random when it is influenced by a large number of events that are equally likely to occur. Fractal time series can have fractional dimensions. Random fractals are combination of generating rules chosen at random for different scales. Only probabilities can help us to understand and take advantage of the process. The underlying philosophy implies that randomness and determinism cannot coexist. the series have similar statistical characteristics. In statistical term. hence. we will find that fractal time series are qualitatively self similar in that. can make fractals useful in capital market analysis. p. Combination of randomness coupled with deterministic generation rules. the time series will be fractal. The data will clump together. the causal process that created the time series has many component parts. Nothing would keep the points in the same vicinity. A random series would have no correlation with previous points. The fractal dimension of a time series is important because it recognizes that process can be somewhere between deterministic (a line with fractal dimension of 1) and random (a fractal dimension of 1. and the interaction of those components is so complex that deterministic explanation is not possible ([1]. the fractal dimension of a line can range from 1 Hurst. These systems are not well-described by standard Gaussian statistics. its fractal dimension is 1. they will fill up whatever space they are placed in.51) do not necessarily have pieces that look like pieces of the whole. May be. to reflect the correlations inherent in its influences. In statistical term. we introduce the term fractal dimension. Random fractals ([2]. The Long-Term Storage Capacity of Reservoirs.57). The fractal dimension describes how a time series fills its space. we must measure how the object clumps together in its space. p. The simplest way is to take a generating rule and iterate it over and over again. A nonrandom time series will reflect the nonrandom nature of its influences. Standard statistical analysis begins by assuming that the system under study is primarily random. a straight line has a fractal dimension of 1. at different scales. Instead. time series is a random walk – a system so complex that the prediction becomes impossible. In fact. 2009 Close 60 Close 60 50 50 40 40 30 30 20 20 10 10 0 2006 2007 2008 2009 2010 0 2006 2007 2008 2009 2010 Date Date Figure 1: Daily and weekly prices of stock Bank of America Corp.16). 3 . we need a probability theory that is nonparametric. a random walk has 50–50 chance of rising or falling. The fractal dimension of a time series measures how jagged the time series is ([1]. To determine the fractal dimension. the number of degrees of freedom or factors influencing the system is very large. However. that is. Fractal shapes can be generated in many ways.E.53). In order to study the statistics of these systems and create a more general analytical framework. is the product of all factors influencing the system that produces time series ([2]. p. In: Transaction of the American Society of Civil Engineers 116. In this paper we introduce nonparametric methodology that was discovered by H. In other words. H. As would be expected. it has a high number of degree of freedom. they may be qualitatively related. In advance.50). In the case of time series.E-Leader Tallinn. Hurst3. to preserve their dimensionality. or “causality”. If we would like to understand the underlying causality of the structure of time series.

This adjusted range. Z. The adjusted range. r=1. for Rn now signifies that this is the adjusted range for x1. Yn). For markets. is independent for increasing values of n. …. by definition. n. by how high H is above 0.nH (5) The subscript. Brownian motion became the primary model for a random walk process. R/S analysis and Hurst exponent Hurst was aware of Einstein`s4 work of Brownian motion.E-Leader Tallinn.…. will always be nonnegative. Y2.50. …. the maximum value of Y will always be greater than or equal to zero. (1) where R is the distance covered and T is a time index. to represent n consecutive values. Financial economists use it to annualize volatility or standard deviation. to allow diverse phenomena and time periods to be compared. To apply this concept to time series that are not in Brownian motion. Because Y has been adjusted to a mean of zero. and the minimum will always be less than or equal to zero. Einstein found that the distance that a random particle covers increases with the square root of time used to measure it. The R/S value of equation (5) is referred to as the rescaled range because it has zero mean and is expressed in terms of local standard deviation. xn}. by rescaling the data to zero mean and standard deviation of one. which many of characteristics of the time series.…. 4 . Einstein. …. we can apply equation (1). However. now has a mean of zero. the R/S value scaled as we increase the time increment. that is. The strength of the trend and the level of noise could be measured by how the rescaled range scales with time. (4) The subscript. At values 1. In general. xn.3. Uber die von der molekularkinetischen Theorie der Warme geforderte Bewegung von in ruhenden Flüssigkeiten suspendierten Teilchen. The next step creates a cumulative time series Y: Y1=(Z1+Zr).56) reformulated Hurst`s work for a general time series as follows. Y2.…. we need to generalized equation (1) and take into account systems that are not independent. provided that the time series. Hence. the adjusted range.…. 2009 to 2.2. equation (1) applies only to time series that are in Brownian motion (they have zero mean. the analysis is called rescaled range analysis (R/S analysis). and we convert it into the fractal dimension d in this paper. or: R=T0. X={x1. We introduce methodology of the Hurst exponent H. x2. Hurst found that most natural phenomena follow a “biased random walk” – a trend with noise. by a power–law value equal to H. Hence. x2. Annals of Physics 322. A. To standardize the measure over time. Rn. prices present a problem because of inflationary growth. is the maximum minus minimum value of the Yr: Rn=max(Y1. 3. and variance is equal to one). is the distance that the system travels for time index n.50. and it is commonly used in statistics. In comparing stock returns of the 1920s with those of the 1980. Rn. Peters ([1] p. n. 1908. r =2. n. They are many ways of calculating fractal dimensions. Rescaling allows us to compare periods of time that may be many apart.n (3) Note that. it can be the daily changes in price of a stock index. xn and c is a constant. Hurst decided to create a dimensionless ratio by dividing the range by the standard deviation of the observations. Equation (1) is called the T to the one–half rule. Yn)–min(Y1. We begin with a time series. The normal distribution has an integer dimension of 2. the last value of Y (Yn) will always be zero because Z has a mean of zero.50<d<2. Rn. The rescaled range was calculated by first rescaling or “normalizing” the data by subtracting the sample mean xm: Zr=(xr–xm). If we set n=T. x2. Rescaling minimize this problem. for (R/S)n refers to the R/S value for x1. generally called the Hurst exponent. X.n (2) The resulting series. Hurst found that the following was a more general form of equation (1): (R/S)n=c. a time series is more jagged than a random series.

log(n) (6) If a system is independently distributed.66). but tends to understate long-run correlation for non-Gaussian series (full mathematical explanation we find in [5].E-Leader Tallinn. p. no mater what the shape of the underlying distribution. We compare the behavior of our process. 2009 Rescaled range analysis can also describe time series that have no characteristic scale.64): 1.50. p. or persistence.0.50. What happens today influences the future. A high H value shows less noise. (7) where C is correlation measure and H is Hurst exponent.71) 8 Variance was calculated by Feller ([1]. 2.50. This time series is more volatile than a random series. p. or ergodic. increases as H approaches 1. Each observation carried a „memory“ of all the events that preceded it. we calculate expected value of the adjusted range7 E(R/Sn) and its variance8 Var(E(R/Sn)). if it was down before. Persistent series are fractional Brownian motion. We will test this null hypothesis: “The process is independent.70). it is more likely to be up in the next period. The strength of this antipersistent behavior depends on how close H is to zero. Equation (7) equals zero. identically distributed and is characterized by a random walk”6. or negative correlation. Mandelbrot called them fractional brownian motions. then the chances are that it will continue to be positive (negative) in the next period. The impact of the present on the future can be expressed as a correlation: C=2(2H-1)–1. 0≤H<0. The ACF works well in determining short-run dependence. Biased random walks were extensively studied by Hurst in the 1940s and again by Mandelbrot in the 1960s and 1970s. Testing R/S analysis To evaluate the significance of R/S analysis. p. A high H means less risk. The closer it is to zero. or near-Gaussian. 4. Its probability density function can be normal curve. There are three distinct classifications for the Hurst exponent ([2]. The ACF assumes Gaussian. When H differed from 0. events are random and uncorrelated. properties in the underlying distribution. R/S analysis can classify an independent series. the noisier it will be.50.([2]. It is important to remember that this correlation measure is not related to the Auto Correlation Function (ACF) of Gaussian random variables ([2].50≤H<1. we calculate expected value of the R/S statistics and the Hurst exponent. p. To verify this hypothesis.5. 3. more persistence and clearer trends than do lower value. the closer C in equation (7) moves toward –0. The present does not influence the future. If the time series has been up in the previous period. The closes H is to 0. H=0. described by R/S analysis with an independent and random system and gauge its significance. but it does not have to be. The strength of the trend-reinforcing behavior. The strength of the bias depends on how far H is above 0.00: time series have a persistent or trend-reinforcing character.61) 6 This process has Gaussian structure (see [1].66) 5 . or biased random walk5. p. Trend is apparent. This is a characteristic of fractals. If the series has been up (down) in the last period. 0.50: time series is antipersistent. 7 This formula was derrived by Anis and Lloyd ([1]. it is more likely to be down in the next period. the observations are not independent. Conversely. Time is important.50: time series is random. Where we are now is a result of where we have been in the past. The Hurst exponent can be approximated by plotting the log(R/Sn) versus the log(n) and solving for the slope through an ordinary least squares regression: log(R/Sn)=log(c)+H. and the less defined its trends will be. then H=0.

then the graph would be upwardly sloping. (8) where St is logarithmic return at time t and Pt is stock price at time t. In other words. Conversely.E-Leader Tallinn.1986 to 7. Hurst. normally distributed and therefore we would expect that the values of H would also be normally distributed (see Peters [1]. the long-memory process has dissipated. the “breaks” would occur when the V chart flattens out. E.50). R/S analysis is capable of determining periodic cycles. 2009 n − 0 . At those points. but it depends on the total sample size T.finance. H. We have 5787 observations for daily frequency (only trading days) and 1197 observations for weekly frequency. R/S values are random variables.com (see Figure 2). 9 . if the process was antipersistent (H<0.5. defined as follows: St=ln(Pt/Pt-1). For financial purpose.5  π  E ( R/Sn ) = ⋅n ⋅  n  2 −0. a plot of V versus log(n) would be flat if the process was an independent. we use logarithmic returns.72): 1 Var ( H n ) = .2009 and the data follow from www. The Long-Term Storage Capacity of Reservoirs. By plotting V on the vertical axis and log(n) on the horizontal axis. Using this statistic we give a more precise measure of the cycle length. The expected Hurst exponent will vary depending on the values of n we use to run the regression. (10) T where T is total number of observations in the sample. Now t-statistics will be used to verify of the significance of the null hypothesis. R/S analysis needs a long time intervals. we will begin with n=10.    Using the results of equation (8) we can generate expected values of the Hurst exponent. p. The statistics is called V and it is defined as follows ([1]. random process. Empirical study We apply R/S analysis to the daily and weekly closing stock prices Bank of America from 29. Note that the Var(Hn) does not depend on n or H. 5.yahoo. p. If the process was persistent and R/S was scaling at a faster rate than the square root of time (H>0. Transactions of the American Society of Civil Engineers 116. Finding Cycles Hurst9 was the first to realize that an underlying periodic component could be detected with R/S analysis ([1].5 n −1 r =1 ∑ (n − r ) r (8) π 2 π  Var ( E ( R / Sn )) =  (9)  6 − 2 ⋅n . 1951. Any range will be appropriate as long as the system under study and the E(R/Sn) series cover to the same values of n.05. 88) and used simple statistic to test stability. the graph would be downward sloping. When analyzing markets.50). p. even when they are superimposed. 6.92): (R / S )n (11) Vn = n This ratio would result in a horizontal line if the R/S statistics was scaling with the square root of time. The real power of R/S analysis is in finding nonperiodic cycles. The final value of n will depend on the system under study.

p. 2009 Daily prices of BAC (29. However. The regression yielded H=0.86 to 7.86−7. a significant result.53520 and E(H)= 0. the Hurst exponent is not significant.83) and for n=1445 observations (log(1445)≈7. V-statistics is decreasing from V68=1. identically distributed and they are characterized by a random walk.28). The variance of E(H). The Hurst exponent is –1. while percentage changes do not (see [2]. see ([1].56213 for daily returns (see Table 3).0132.22). we cannot increase the time frame.57486 and expected H equals to 0. The H value for weekly returns is –1. The standard deviation of E(H) is 0. for 10≤n<70. Unfortunately. logarithmic returns are more appropriate than the more commonly used percentage change in prices. 10 We need T=4/(H-E(H))2 points. We need 3396 observations to achieve significance10. as shown in equation (10) is 0.63978.E-Leader Tallinn.0132. H=0. Figure 4 and Table 2 show the results of R/S analysis.53540 and E(H)=0. The regression yielded H=0.56952 for weekly returns (see Table 4).0003 and standard deviation of E(H) is 0.83). Unfortunately. (BAC).5.54422 for 70<n≤2890. To estimate precisely where this break occurs. The variance of E(H) is 0.13 to V85=1.0002. we have only T=1188 observations for weekly return data of stock Bank of America. it means that weekly returns are independent.5.51900 and E(H)= 0.52939 and they are excessively closely and therefore H is insignificant.8418 standard deviations bellow its expected value.5. Hurst exponent equals to 0. from 29.153) .86−7.56952 (see Table 4). We see a systematic deviation from the expected values on the Figure 3.8418 standard deviation bellow its expected value. The range used in R/S analysis is the cumulative deviation form the average.50). for Gaussian random variables.09) Close 130 120 110 100 90 80 70 60 50 40 30 20 10 0 1980 1985 1990 1995 2000 2005 2010 2015 Date Date Figure 2: Daily and weekly prices of Bank of America Corp. The next subperiod is 70<n≤1445.05.53520 and E(H)=0. H=0. for n=340 observations (log(340)≈5.09 For R/S analysis. Process became persistent. The H value for daily returns is –2. where the slope appeared to follow the E(R/S) line.5. p. Table 1 show both the R/Sn values and the Vn. Also plotted is E(R/S) (calculated using equation (8)) as a comparison against the null hypothesis that the system is an independent process.09) Close 130 120 110 100 90 80 70 60 50 40 30 20 10 0 1980 1985 1990 1995 2000 2005 2010 2015 Weekly prices of BAC (29. Figure 3 (on the left) shows the log R/S plot for daily return data for T=5775 observations.11. a break in the R/S graph appears to be at n=68 observations (log(68)≈4. We will examine the behavior of H over different time increments.5. The series exhibits persistence (H>0.5. and logarithmic returns sum to cumulative return. we calculate the V-statistics using equation (11) (Vstatistics versus log(n) is plotted in right Figure 3). a non significant result for confidence level α=0. Stocks Bank of America were not come off until 1986.0313 standard deviation bellow its expected value. for daily and weekly returns of stock Bank of America Corp. Hurst exponent were estimated from the R/S plot and the E(R/S) and H equals to 0.

68908 3.833213 2.94454 28.110165 1.219042 1.986735 0.890372 3.47719 V statistic BAC E(R/S) 0.945871 0.682131 4. Bank of America: weekly returns R/S analysis.E-Leader Tallinn.978472 3.990546 0.14748 1.981414 6.276594 4.77524 29.71537 12.204843 1.551656 2.694515 6.602551 5.526361 4.10369 1.496508 3.988984 4. 2009 N 10 17 20 34 68 85 170 289 340 578 1156 1445 2890 log N 2.135798 5.091042 3.583519 3.648838 6.034132 1.2055 Table 2: R/S analysis and V-statistics.252023 1.852481 2.116562 1.1 2 1.19011 1. Bank of America Corp.147031 5.39936 23.027127 0.137727 1. daily returns log_RS 5 V−statistics.628198 5.60723 24.322625 4.222661 1.098083 1.868894 2.160488 1.07206 2.63842 1.142953 1.084408 1.82832 17.192693 4.920869 10.037391 4. daily returns V_stat 1.876819 0.176707 1.981267 1.21137 1.983337 1.967042 1.796379 R/S 3. Bank of America: daily returns .264156 7.442651 5.052721 7.0 1 log_RS log_ERS 0.28237 28.122582 1.275865 log R/S 1.80396 13.135837 1.120197 1.969012 4.2472 11.373262 2.3806 BAC 0.231924 2.371679 2.78419 3.22053 66.546395 1.5396 E(R/S) 2.065127 1.44118 40.42308 33.947224 11.167522 1.295837 3.968737 8.666427 5.433343 1.078642 1.09341 1.484907 2.009434 1.941009 0.129947 1.031087 1.101265 10.182707 1.386879 log R/S 1.325834 2.032329 4.65775 21.463863 1.940635 6.998763 1.154199 1.858808 0.73145 1.44743 46.995732 3.941168 8.00803 44.057879 3.166956 1.90229 19.4 4 1.4285 20.182997 1.288267 5.049725 1..19614 E(R/S) 2.975899 3.10602 21.065256 7.818525 1.072779 4.212199 3.693732 5.101327 1.219508 4.048391 1.879877 4.9 V_stat V_ERS 0 2 3 4 5 6 7 8 0. Bank of America Corp.23522 14.324742 6.245172 5.189655 4.13091 20.32771 15. Bank of America: daily returns V statistic N 11 12 18 22 27 33 36 44 54 66 99 108 132 198 297 396 594 log N 2.183691 1.882802 5.988227 0.380623 Table 1: R/S analysis and V-statistics.307041 74.532168 1.188393 2.8 2 3 4 5 6 7 8 log_N log_N Figure 3: R/S analysis and V-statistics.3 1.397895 2.73235 10.195528 3.044026 1..156338 1.971698 1.2 3 1.171689 0.625041 7.968515 0.61753 19.890857 2.502434 R/S 3.302585 2.080656 1.37562 9.186805 1.162762 6.317779 10.050077 9.347838 3.650278 3.59512 4.362054 E(R/S) 0.838092 0.848343 3.828946 6.359574 7.701515 2.96639 41.227306 1.23164 7.950429 0.130386 1.03758 1.18354 16.194751 1.

01963 0.9977 19 0.0001 -0.2 1..90 Pr > |t| 0.0001 113. weekly returns log_RS 4 V−statistics.9976 0.0002 0.00635 -6.16 <.23026 0.58 <.02633 0.87 Pr > |t| 0.05695 0.3 3 1.20089 0.72 <..53540 0.8418 -0. weekly returns V_stat 1.0099 136. Bank of America Corp.9955 0.0313 -0.0003 0.56952 0. Bank of America: weekly returns .03035 0.03168 0.02350 0.0001 88. Bank of America: weekly returns Parameter Estimates for daily returns of BAC Parameter Standard DF Estimate Error Intercept Hurst exponent H R-Square for H Adj R-Sq for H Expected Intercept expected Hurst exponent E(H) R-Square for E(H) Adj R-Sq for E(H) Number of Observations Var(E(H)) s(E(H)) significance 1 1 0.1 2 1.9978 0. 2009 R/S analysis.4 1.00501 -9.56213 0.8 2 3 4 5 6 7 log_N log_N Figure 4: R/S analysis and V-statistics.9975 33 0.34 <.00647 t Value -0.66 <.9 log_RS log_ERS V_stat V_ERS 1 2 3 4 5 6 7 0.0001 -0.00393 t Value -2.0001 Table 4: Hurst exponent for R/S analysis. Bank of America Corp.53520 0.9953 1 1 0.9990 1 1 0.80 <.0186 -1.0132 -2.0 0.3922 82.E-Leader Tallinn. Bank of America: daily returns Parameter Estimates for weekly returns of BAC Parameter Standard DF Estimate Error Intercept Hurst exponent H R-Square for H Adj R-Sq for H Expected Intercept expected Hurst exponent E(H) R-Square for E(H) Adj R-Sq for E(H) Number of Observations Var(E(H)) s(E(H)) significance 1 1 0.9991 0.0001 Table 3: Hurst exponent for R/S analysis.

HURST..E. [1] [2] [3] [4] [5] [6] Bibliography PETERS.asp MANDELBROT. 110 s. Annals of Economic Social Measurement 1. Inc.6 may 2009 0. Yesterday`s events do not impact today. 1972.56213 0. 8.E.com/financialmarkets. 9. B. 1996.47579 0.117302494 -3.393130167 Table 5: Hurst exponent for Bank of America daily return 7. These results may be caused by insufficient number of data.50 implies a random walk.019181178 0. ISBN 80-245-0598-3. 315 s. grant numbers 1/0500/09.53540 0. Today`s events do not impact tomorrow. H=0. E. p.56067 0. 1994.50 implies that today`s events do impact tomorrow.58554 Standard deviation 0. The Long-Term Storage Capacity of Reservoirs.55842 0. 1994.56174 0. it is a longer memory function. Statistical methods and capital markets.economymodels. we divide time series to three parts: from 29 may 1986 to 25 February 1997. 2003. J.27 aug 2004 Hurst coeficient 30 aug 2004 . 1/4024/07 and 1/0373/08. Acknowledgement The work on this paper has been supported by Science and Technology Assistance Agency under the contract No. Old news has already been absorbed and discounted by the market. From Table 5 we see that time series had persistent character (time series is fractal random walk) from 29 May 1986 to 6 May 2009 and it is significant result. H greater than 0. PETERS. 1996. 1972. Information received today continues to be discounted by the market after it has been received. from 26 February 1997 to 27 August 2004 and from 30 August 2004 to 6 may 2009. H. Praha: Oeconomica. New York: John Wiley & Sons. E. and by the VEGA grant agency. Estimated H Hurst coeficient 29 may 1986 -6 may 2009 Hurst coeficient 29 may 1986 -25 feb 1997 Hurst coeficient 26 feb 1997 . This is not simply serial correlation. APVV-0375-06.033237792 0.023014365 0. ISBN 0-471-58524-6. Chaos and order in the capital markets. and it goes across time scales. Transactions of the American Society of Civil Engineers 116. 274 s. Statistical Methodology for Non-Periodic Cycles: From the Covariance to R/S Analysis. Fractal market analysis. TREŠL. For next period.013146519 0.E-Leader Tallinn. Information can impact the future for very long periods. E. Conclusion In this paper we have shown how it is possible to measure the impact of information on the time series by using Hurst exponent H([2]. New York: John Wiley & Sons. ISBN 0-471-13938-6.54495 Expected H 0. These dates are corresponding to the rapid changes in the prices of the stock. 2003. 1952 . from 26 February 1997 – 27 August 2004 time series is significantly antipersistent (time series is more volatile than a random series) and last period indicate random walk.. Inc. Time series is random walk from 29 |May 1986 to 25 February 1997. 1951.029135827 significance -2.734623933 -1. http://www.102). 2009 Now. The events are uncorrelated.