Clustering Elimination

Ing. Bohumil Stdnk, Ph.D.

bohumil.stadnik@email.cz
The University of Economics in Prague
Faculty of Finance and Accounting
W. Churchill Sq. 4
130 67 Prague 3
Czech Republic

Prague, Czech Republic, August 2013

Abstract

In this financial engineering research we try to recognize if non-normalities in the market price distribution
of government bonds are caused by volatility clustering or also by another non-clustering mechanism. Such findings
allow us to assess according to which rules the market price is actually developing or even make serious
conclusions on market price directional forecasting chances. For such purpose we study European bond futures
(Eurex Euro-Bund Futures) which directly affects prices of appropriate government bonds. The research is done
with respect to the different time periods.

Keywords

Volatility clustering, departures from normality, bond market, Euro-Bund Futures, directional dependence, volatility
dependence

JEL codes: G1, G10, G12, G14

Introduction

The main contribution of this financial engineering study is to resolve the question: Are the departures
from normality in European government bonds price distribution caused by volatility clustering or also by another
non-clustering mechanism distributing the price in a non-normal way? Solution of this question is important for a
general assessment of the bond market functionality depending on the departures responsible mechanisms. Some of
the mechanisms are probably hidden (Stdnk, 2013[1]) and especially directional dependent mechanisms, if they
are presented, allows making interesting assessment for the future market price development forecasting. In this
research we check the price distribution inside 1, 5, 10, 30, 60 minutes periods and also in one day time series as we
expect specific economical processes which are dominating inside the short time periods, processes which are
significant for the longer periods and we also observe processes which are common for all the periods.

Methodology

To make the decision between the clustering or non-clustering mechanism responsible for the non-
normalities we have to, first of all, assess the impact of both the mechanisms on the character of the price
distribution and the departures.
General clustering mechanism causes significant autocorrelation in volatility data series and possibly the
departures from normality in the distribution but we have to mention at this point also the special case of observing
volatility clusters but the resulting distribution is the Gaussian one without the departures as it is simulated in the
figure 14 in the appendix. In spite of this case the volatility clustering is nowadays considering being the main cause
of the leptokurtic departures from normality and the clusters itself could be caused by a pure volatility dependency
effects. As the pure volatility dependence process is considered to be the process which price direction is always
independent on the past but the volatility is dependent. Such process does not allow directional forecasting and it is
closely connected to the size of price steps in the given time period. There are more theories of basic research in the
area of volatility dependence. For example Buckley, 2008, has used in his work the Gaussian mixture distribution.
Gaussian mixture has an acceptable interpretation: financial market performs in two regimes with high and low
volatility. Gaussian mixture can model many departured distributions which depend on the probability of both
regimes and their parameters. If the latent regimes have a Markov law of motion, the mixture is then a hidden
Markov model (Baum, Petrie, 1966), which is also known as the Markov regime switching model. There are many
extensions of Markov switching model (Krolzig, 1997; etc.) Other famous works in this area were done by
Bollerslev, 1986, GARCH process; Engle, 1995, ARCH process. Some new research in the area of volatility
dependence was done by Witzany, 2013 (Estimating Correlated Jumps and Stochastic Volatilities) or Roch, 2011.
While GARCH, ARCH and other stochastic volatility models propose statistical constructions based on volatility
clustering in financial time series, they do not provide any economic explanation. The financial explanation of
volatility clustering is quite difficult. The simplest possible financial clustering mechanism is just switching of the
market between periods of high and low activity or clustering of economical news. The other idea was the
competition between more trading strategies but the simulation does not allow confirming mechanism being
responsible for volatility clustering (Cont, 2005). Some economic works contain examples where switching of
economic agents between two behavioral patterns leads to large volatility. Volatility clustering should also arise
from switching of market participants between fundamentalist and chartist behavior (Lux, Marchesi, 2000).
Fundamentalists expect that the price follows the fundamental value in the long run. Traders using technical
analysis try to identify price trends or other patterns. Agents are allowed to switch between these two behaviors
according to the performance of the various strategies. Chart traders evaluate their investments using historical
development, whereas fundamentalists evaluate their investment opportunity according to the difference between
the market price and the fundamental valuation. According to the Lux-Marchesi model the market price
development follows Gaussian random walk till the moment when some chart traders using certain techniques
surpass a certain threshold value. At this moment a volatility outbreak occurs. According to Cont, 2005, the origin
of volatility clustering can be also caused by threshold response of investors to news arrivals.
Instead of the volatility dependency effects we are able to explain the non-normality using the pure
directional dependency effects. This way considers the price development direction to be dependent on the past and
allows certain forecasting chances in comparison to volatility dependency. There are many case studies based on the
directional dependency but comprehensive modeling of the departures from normality in this way is not so frequent.
For example commonly used technical trading rules are based on a market price direction forecasting according to
the past. We can consider Technical Analysis to be the prediction tool, but its benefit is still under discussion. Some

2
works indicate that several technical indicators do provide a little forecasting improvement and may have some
practical value (Lo, Mamaysky, Wang, 2000). We meet many other interesting detailed works or a case studies in
the area like Henriksson, Merton, 1981; Anatolyev, Gerko, 2005; other works like Vacha, Barunik, Vosvrda, 2009;
Primbs, Rathinam, 2009; study about the connection of liquidity and market crashes done by Huang, Wang, 2010;
Lux, 2011. Price direction development dependence is also taking place in the basic feedback process according to
the behavioral finance concept where upward trend is more likely to be followed by another upward movement
(Schiller, From Efficient Markets Theory to Behavioral Finance, 2003) or in other research as for example
momentum studies (Pesaran, Timmermann, 1995; Chan, Jegadeesh, Lakonishok, 1996; Stankeviien, Gembickaja,
2012), short term trend trading strategy in futures market based on chart pattern recognition (Masteika, Rutkauskas,
2012), development of the conception of sustainable return investment decisions strategy in capital and money
markets (Rutkauskas, Mieinskiene, Stasytyte, 2008). We have to mention also works of Larrain, 1991, which states
that long term memory exists inside the financial market, other similar works of Hsieh 1991, Peters 1989, 1991,
1994 which focus mainly on measurement of probability diversions from normality. Important for us is that the
directional dependency way is able to explain the departures without clustering mechanisms. For example feedbacks
system according to the comprehensive Dynamic Financial Market Model (Stdnk, 2011 [1], [2]) is able to cause
sharpness and fat tails in the distribution. Feedbacks increase the value of probability of next price step up or down
direction (from 50/50 for the pure random walk to for example 51/49) depending on the previous development. The
idea of feedback processes is based on the observations that traders, investors and other market participants dont
only watch present or historical data but according to them they are also placing buy or sell orders and thus
influence future development. Feedback which keeps the movement in the certain direction is described in the
model as the trend stabilizer feedback. For example momentum trading when traders try to find instruments that are
moving significantly in one direction on high volume and try to participate on the profit from this movement is
based on this mechanism. The other important feedback is the price inertia feedback which is pushing the market
price back to a certain level. The certain level can be for example the price level which was set after the last
economic news of high importance. The price inertia feedback contributes to the sharpness in the distribution and
the trend stabilizer feedback contributes to the fat tails according to the model. If the feedbacks work separately
they could distribute the price to the initial value or contrary to the farther distance from the initial value. In such
cases the distribution is non-normal (the leptokurtic one) and the volatility series is without the volatility clusters.
The result of the simulation of such mechanism is in the figure 15 in the appendix. For the assessment we use some
quantitative indicators. The value of kurtosis is not useful quantitative pointer of the sharpness in this case
especially when the price inertia is active separately. There is demonstrated the case of the same sharpness and
different values of kurtosis in the figure 1a and 1b.

10
10
8 8
6 6
4 4

2 2

0 0
1
35
69
103
137
171

1
30
59
88
117
146
175

a) b)

Fig. 1: Distributions with the same standard deviation: 47.15, the same sharpness but the different kurtosis: 5.28588 and
0.096698

To avoid this problem we have defined the acuteness as the ratio of maximum value of measured
distribution and maximum value of an adequate normal distribution:

Maxmeasured
acute
Maxnormal

The value of acuteness for normal distribution is 1.

3
 Volatility clustering could be also well explained using the directional dependency effects like the spring
oscillation effect (Stdnk, 2013[2]) when feedbacks may cooperate and under the certain conditions (Stdnk, 2013
[2], Spring Oscillations Effect) cause volatility clusters as the final result. This is the case when we observe
volatility clusters which are not caused by volatility dependence but the directional dependence.
Based on previous we can logically conclude:
If there is no significant volatility autocorrelation and the distribution exhibits certain acuteness the
departures are not caused by the volatility clusters but by a non-clustering mechanism which could be for example
price inertia feedback action.
If there is significant volatility autocorrelation and certain acuteness then the departures may be caused by a
clustering mechanism together with a non-clustering mechanism. In such case we have to decide if the departures
are caused only by the clusters or only by the feedbacks or by the coexistence of both the effects. To answer such
question we suggest the filtering of volatility clusters thus separate from the data series the continuous parts without
the clusters. We continue to make this filtering until the autocorrelation of volatility time series is insignificant but
we also cannot destruct the series (continuous parts without clusters must left). After the filtering we are allowed to
study the price distribution of the data series without the clusters and also inside the clusters separately. If the
volatility series without the clusters is departured from normality the non-clustering mechanism causing the
departures is presented. If the value of kurtosis of such distribution is lower than for the distribution with the
clusters we can conclude on coexistence of both the clustering and non-clustering mechanism.
In the case that volatility clusters is not possible to eliminate without the destruction of data series (we
cannot separate continuous time periods without clusters) we cannot be sure if the departures are cause by the
volatility clustering or by non-clustering effects.

Euro-Bund Futures 1min, 5 min and 10 min Distributions

In case of 1, 5, and 10 minutes price volatility data series (figure 2a, 3a, 4a) we were not successful to
eliminate volatility clusters (to decrease the level of autocorrelation) without the destruction of the time series.

a) b)

Fig. 2: 1 minute price distribution of Euro-Bund Futures (FGBL, EUREX), volatility autocorrelation: 0.20106
average value: 0.000057873, skewness: -0.15601, kurtosis: 29.36968, acuteness: 1.47953, data: May 2013

a) b)

Fig. 3: 5 minutes price distribution of Euro-Bund Futures (FGBL, EUREX), volatility autocorrelation: 0,2292, average value: -
0.00042769, skewness: 0.019039, kurtosis: 20.51322, acuteness: 1.6132, data: May 2013

4
 a) b)

Fig. 4: 10 minutes price distribution of Euro-Bund Futures (FGBL, EUREX), volatility autocorrelation: 0,18373, average
value: -0.00061583, skewness: -0.085219, kurtosis: 9.04905, acuteness: 1.63108, data: April-May 2013

In the case of 1, 5 and 30 minutes price development we cannot be sure about reliable conclusions. The
departures are probably caused mainly by certain clustering mechanism (autocorrelation: 0.20106, 0.2292, 0.18373)
but we are not able to make any conclusion on non-clustering mechanism based on the measurement of volatility
data series in the case. The solution could be made only by the direct observation on the market.

Euro-Bund Futures 30 min, 60 min Distributions

In the 30 and 60 minutes price volatility series (figures 5a, 6a) the volatility has low autocorrelation (0.10835
and 0.096636), the distributions (figures 5b, 6b) are with the high acuteness (1.76064, 1.76577) and also kurtosis. We
can conclude that departures are caused mainly by the non-clustering mechanism.

a) b)

Fig. 5: 30 minutes price distribution of Euro-Bund Futures (FGBL, EUREX), volatility autocorrelation: 0.10835, average
value: -0.00090405, skewness: -4.23785, kurtosis: 84.03866, acuteness: 1.76064, data: February-May 2013

a) b)

Fig. 6: 60 minutes price distribution of Euro-Bund Futures (FGBL, EUREX), volatility autocorrelation: 0.096636, average
value: 0.00026424, skewness: -2.13688, kurtosis: 34.87858, acuteness: 1.76577, data: February-May 2013

5
Euro-Bund Futures 1 Day Distribution

In the case of 1 day price volatility series (figure 7a) we have been successful to eliminate the volatility
clusters (figure 8a) from the series thus eliminate volatility autocorrelation to the insignificant level (from 0.13817 to
-0.0050248). There is volatility autocorrelation (0.012026) of an independent random walk in the figure 13 in the
appendix for a comparison. We can conclude that the price distribution which does not involve the clusters (figure
8b) has the significant acuteness and therefore there is presented a non-clustering mechanism responsible for the
departures. As the value of kurtosis of the distribution which involves the clusters (figure 7b) is higher we can
confirm coexistence of the clustering and non-clustering mechanisms.
Inside the volatility clusters (figure 9a) where the autocorrelation is also insignificant (0.018581) and the
distribution (figure 9b) has significant acuteness we can also confirm non-clustering mechanisms responsible for the
departures. As the value of kurtosis of the distribution which involves the clusters (figure 7b) is higher we can
confirm coexistence of the clustering and non-clustering mechanisms also in this case.

a) b)

Fig. 7: 1 day price distribution of Euro-Bund Futures (FGBL, EUREX), volatility autocorrelation: 0.13817, average value:
0.010746, skewness: -0.29403, kurtosis: 3.26562, acuteness: 1.59332, data: 1990- 2013

a) b)

Fig. 8: 1 day price distribution of Euro-Bund Futures (FGBL, EUREX) without VOLATILITY CLUSTERS, volatility
autocorrelation: -0.0050248, average value: 0.02, skewness: -0.06136, kurtosis: 0.61863, acuteness: 1.36993, data:1990- 2013

1 day
1 day
1.4

1.2

0.8

0.6

0.4

0.2

0
1 19 37 55 73 91 109 127 145 163 181 199 217 235 253 271 289

a) b)

Fig. 9: 1 day price distribution of Euro-Bund Futures (FGBL, EUREX) inside VOLATILITY CLUSTERS, volatility
autocorrelation: 0.018581, average value: 0.00, skewness: -0.38318, kurtosis: 1.400629, acuteness: 1.4165, data: 1990- 2013

6
 For the clusters elimination we have used special software which detects continuous periods of lower and
higher volatility. Time series were downloaded from Reuters.

S&P500 1 day Distribution

For the comparison we have tried to eliminate clusters from S&P500 price volatility series (figure 10a). In
this case we have been successful to eliminate the volatility clusters and such reduce volatility autocorrelation (from
0.22055 to 0.02482). We can conclude that the price distribution which does not involve the clusters (figure 11b) has
significant the acuteness (1.79542) and therefore there is also presented a non-clustering mechanism which causes
the departures from normality. As the value of kurtosis is higher with the clusters we can confirm coexistence of the
directional and volatility dependency as in the case of Euro-Bund Futures contract.
We have also confirmed significant non-clustering mechanism causing departures inside the clusters (figure
12a), because the value of autocorrelation is insignificant (0.04508) but the acuteness is significantly high (1.8589).

a) b)

Fig. 10: 1 day returns distribution of S&P500, volatility autocorrelation: 0.22055, average value: 0.029437, skewness: 0.93606,
kurtosis: 25.42172801, acuteness: 1.79542, data: 1963- 2013

a) b)

Fig. 11: 1 day returns distribution of S&P500 without VOLATILITY CLUSTERS, volatility autocorrelation: 0.02482, average
value: 0.035793, skewness:, kurtosis 0.035566, acuteness: 1.58969, data: 1963- 2013

a) b)

Fig. 12: 1 day returns distribution of S&P500 inside VOLATILITY CLUSTERS, volatility autocorrelation: 0.04508, average
value: 0.0062996, skewness: 0.12733, kurtosis 1.018144, acuteness: 1.8589, data: 1963- 2013

7
 To support our ideas about impact of non-clustering mechanism on the price distribution we have made the
simulation of price inertia and trend stabilizer feedbacks action (figure 15, appendix). In this simulation we simulate
price inertia and trend stabilizer according to Dynamic Financial Market Model. The simulation is without any
volatility clustering. We can see in the figure that the volatility autocorrelation (0.02364) is insignificant but the
value of acuteness (1.66544) is significantly high.

Conclusion

In this financial engineering research we have tried to recognize if the departures from normality in the
European government bonds distributions are caused mainly by a volatility clustering or by non-clustering effects.
For such purpose we have studied European bond futures (Eurex Euro-Bund Futures) which directly affects prices
of European government bond prices. We have suggested methodology for the recognition between clustering and
non-clustering effects being responsible for the departures which are applicable to the worldwide financial
investment instruments. From the price development we have suggested to filter volatility clusters and study the
price distributions without the clusters and also inside the clusters separately. We have also defined certain
quantitative pointer as the measure of certain non-clustering mechanisms.
For the shorter time periods series (1, 5, 10 minutes) we have not been successful to confirm a non-
clustering mechanism and we have concluded that volatility clustering is probably the key factor in high frequency
distributions. A non-clustering mechanism we can deduce based on the direct empirical observations of behavior of
market participants.
For 30 and 60 minutes distributions we have recognized that the volatility autocorrelation is lower and due
to the significant departures the key factor is probably a non-clustering mechanism.
For the daily distributions we have confirmed coexistence of the clustering and non-clustering mechanisms.
We were successful to eliminate the volatility clusters from the development and we have recognized that such
development is also departured from normality.
The same result we have obtained for S&P 500 distribution of returns.

References

Anatolyev, S.; Gerko, A. 2005. A trading approach to testing for predictability, Journal of Business and
Economic Statistics 23: 455-461.

Baum, L., E.; Petrie, T. 1966. Statistical Inference for Probabilistic Functions of Finite State Markov
Chains, The Annals of Mathematical Statistics 37(6): 15541563.

Bollerslev, T. 1986. Generalized Autoregressive Conditional Heteroskedasticity, Journal of

Econometrics 31: 307-327.

Buckley, I.; Saunders, D.; Seco, L. 2008. Portfolio Optimization When Asset Returns Have the Gaussian
Mixture Distribution. European Journal of Operational Research 185 (3): 1434-1461.

Chan, L., K., C.; Jegadeesh, N; Lakonishok, J. 1996. Momentum strategies, Journal of Finance 51/5:
1681-1713

Cont, R. 2005. Volatility Clustering in Financial Markets: Empirical Facts and AgentBased Models, A
Kirman & G Teyssiere (eds.): Long memory in economics, Springer.

Engle, R.F. 1995. ARCH: Selected Readings. Oxford, UK, Oxford University Press.

Henriksson, R., D.; Merton R.,C. 1981. On the Market Timing and Investment Performance of Managed
Portfolios II Statistical Procedures for Evaluating Forecasting Skills, Journal of Business: 513-
533.

Hsieh, D., A. 1991. Chaos and Nonlinear Dynamics: Application to Financial Markets. Journal of
Finance 46: 18391877.

8
Huang, J.; Wang, J. 2010. Liquidity and Market Crashes, Review of Financial Studies.

Krolzig, H., M. 1997. International Business Cycles: Regime Shifts in the Stochastic Process of Economic
Growth, Applied Economics Discussion Paper 194, University of Oxford.
Larrain, M. 1991. Testing Chaos and Nonlinearities in T-bills Rates, Financial Analysts Journal, September-
October: 51-62.

Lo, A., W.; Mamaysky, H.; Wang, J. 2000. Foundations of Technical Analysis: Computational
Algorithms, Statistical Inference, and Empirical Implementation, Journal of Finance.

Lux, T. ; Marchesi, M. 2000. Volatility clustering in financial markets: a micro simulation of

interacting agents, International Journal of Theoretical and Applied Finance, 3: 675702.

Lux, T. 2011. Sentiment dynamics and stock returns: The case of the German stock market
Empirical Economics 41 (3): 663-679.

Masteika, S.; Rutkauskas, A.,V. 2012. Research on futures trend trading strategy based on short term chart
pattern, Journal of Business Economics and Management 13 (5): 915-930.

Pesaran, M.,H.; Timmermann, A. 1995. Predictability of stock returns: Robustness and economic
significance, Journal of Finance 50: 1201-1228

Peters, E. 1989. Fractal Structure in the Capital Markets, Financial Analysts Journal/July-August: 32- 37.

Peters, E. 1991. Chaos and Order in the Capital Markets: A New View of Cycles, Prices, and Market
Volatility, John Wiley & Sons, New York.

Peters, E. 1994. Fractal Market Analysis: Applying Chaos Theory to Investment and Economics, John
Wiley & Sons, New York.

Primbs, J.,A.; Rathinam, M.2009. Trader Behavior and Its Effect on Asset Price Dynamics, Applied
Mathematical Finance 16/2: 151-181.

Rutkauskas, A., V.; Mieinskien; A. Stasytyt, V. 2008. Investment Decisions Modelling along
Sustainable Development Concept on Financial Markets, Technological and Economic Development of
Economy, 14(3): 417427.

Roch, A., F.2011. Liquidity risk, price impacts and the replication problem, Finance and Stochastics
15/3: 399-419

Schiller. 2003. From Efficient Market Theory to Behavioral Finance, Yale University.

Stdnk, B. 2011. [1]. Dynamic Financial Market Model and Its Consequences. Available at SSRN:
http://ssrn.com/abstract=2062511.

Stdnk, B. 2011. [2]. Explanation of S&P500 Index Distribution Deviation from a Gaussian
Curve(Dynamic Financial Market Model), Journal of Accounting and Finance, vol. 11(2), USA,
North American Business Press, ISSN: 2158-3625

Stdnk, B. 2013. [1]. Market Price Forecasting and Profitability - How To Tame Random Walk?, Verslas:
Teorija ir Praktika / Business: Theory and Practice, Volume 14, Issue 2, 2013: 166-176

Stdnk, B. 2013. [2]. Spring Oscillations Effect, http://pozemstan.cz/sposef.pdf

9
 Stankeviien, J.; Gembickaja, N. 2012. Market Behavior: Case Studies of NASDAQ OMX Baltic,
Business, Management and Education 10(1): 110127, ISSN 2029-7491 print / ISSN 2029-6169 online.

Vacha, L.; Barunik, J.; Vosvrda, M. 2009. Smart Agents and Sentiment in the Heterogeneous Agent
Model, Prague Economic Papers 2, Praha.
Witzany, J. 2013. Estimating Correlated Jumps and Stochastic Volatilities, to appear in Prague
Economic Papers 2013

Appendix

a) b)
Fig. 13: Random walk, volatility autocorrelation: 0.012026, average value: -0.035625, skewness: -0.0092008, kurtosis: -
0.25217

a) b)
Fig. 14: Clustering with the Gaussian distribution, volatility autocorrelation: -0.56357, skewness: -0.0092008, kurtosis: -
0.25217
1 day
1.6

1.4

1.2

0.8

0.6

0.4

0.2

a) 1 148 295 442 589 736

b)
Fig. 15: 1 day returns distribution of S&P500 SIMULATIONS USING FEEDBACKS (WITHOUT VOLATILITY
CLUSTERING), volatility autocorrelation: 0.02364, skewness: -1.057, kurtosis: 5.25904, acuteness: 1.66544

10