Is the asset price jump? 2010-7-22. 1.Introduction Several studies have recently presented strong non- parametric high-frequency data-based empirical.

Is the asset price jump?

2010-7-22

1.Introduction

Several studies have recently presented strong non-parametric high-frequency data-based empirical evidence in favor of jumps in financial asset prices, thus discrediting the classical continuous time paradigm with continuous sample price paths in favor of one with jump discontinuities.

Empirical studies and new statistical procedures include the work of Ait-Sahalia and Jacod (2008), Andersen et al. (2007b), Barndorff-Nielsen and Shephard (2004), Fan and Wang (2006), Huang and Tauchen (2005), Jiang and Oomen (2006), Lee and Mykland (2008), and Mancini (2004).

The tests that we implement here further corroborate this evidence for the presence of jumps at both the individual stock and aggregate market level.

Of particular interest, however, are the contrasts between the outcome of tests for jumps conducted at the level of the individual stocks and the level of the index. Jumps in individual stocks can be generated by either stock-specific news or common market-level news.

Meanwhile, from basic portfolio theory one might expect that jumps in a well-diversified index should only be generated by market-level news that induces cojumps across many stocks.

Jumps are clearly of importance for asset allocation and risk management. A risk averse investor might be expected to shun investments with sharp unforeseeable movements.

Jumps like that are of great importance for standard arbitrage-based arguments and derivatives pricing in particular, as the effect cannot readily be hedged by a portfolio of the underlying asset, cash, and other derivatives.

More formally, the effect of a jump on the price of a derivative is locally nonlinear and thereby cannot be neutralized by holding an ex ante-determined portfolio of other assets.

Of course, not all jumps are sharply ex post identifiable, so that a formal statistical methodology for identifying jumps is needed. In the results reported on below, we start with a univariate analysis of stock-by-stock jump tests based on the seminal work by Barndorff-Nielsen and Shephard (2004) (BN–S). The BN–S theory provides a convenient non-parametric framework for measuring the relative contribution of jumps to total return variation and for classifying days on which jumps have or have not occurred.

The fact that the data reveal a less important role for jumps in the index than in each of its components is not all surprising, since the idiosyncratic jumps should be diversified away in the aggregate portfolio. At the same time, however, there is at best a very weak positive association between the significant jumps in each of the individual stocks and the jumps in the index.

The BN–S jump detection procedure relies on a daily z-statistic, with large positive values of the statistic discrediting the null hypothesis of no jumps on that day. We find the z-statistics for the individual stocks to be essentially uncorrelated with the z-statistics for the index, despite the fact that most of the individual stocks have b’s close to unity with respect to the index.

As one might conjecture, this low degree of association appears to be due to the large amount of idiosyncratic noise in individual returns. This noise masks the cojumps, making them nearly undetectable at the level of the individual return. To circumvent this problem, Bollerslev T. et al (2008) introduce a new cojump test applicable to situations involving large panels of high-frequency returns.

As noted by Andersen et al. (2007b), Barndorff-Nielsen and Shephard (2006), Eraker et al. (2003), and Huang and Tauchen (2005) among others, many, although not all, of the statistically significant jumps in aggregate stock indexes coincide with macroeconomic news announcements and other ex post readily identifiable broad-based economic news which similarly impact financial markets in a systematic fashion.

While macroeconomic events obviously also affect individual firms, individual stock prices are also affected by sudden unexpected firm-specific information that can force an abrupt revaluation of the firms’ stock.

2. Generalized quadratic variation theory

We start by considering the ith log-price process from a collection of n processes evolving in continuous time. We assume that evolves as

where and refer to the drift and local volatility, respectively, is a standard Brownian motion, and is a pure jump Levy process.

A common modeling assumption for the Levy process is the compound Poisson process, or rare jump process, where the jump intensity is constant and the jump sizes are independent and identically distributed.

we adopt the timing convention that one time unitcorresponds to a trading day, so that represents the continuous log-price record overtrading day t, where the integer values coincide with the end of the day.

In practice the price process is only available at a finite number of points in time. Let M + 1 denote thenumber of equidistant price observations each day; i.e., The jth within-dayreturn is then defined by

for a total of M returns per day.

As discussed Andersen et al. (2008), the realized variance

provides a natural measure of the daily ex post variation. In particular, it is well known that for increasingly finer sampling frequencies, or consistently estimates the total variation comprised of the integrated variance plus the sum of the squared jumps

where denotes the number of within-day jumps on day t, and refer to the size of the kth such jump.

In order to separately measure the two components that make up the total variation, Barndorff-Nielsen and Shephard (2004) and Barndorff-Nielsen et al. (2005) first proposed the so-called bipower variation measure

Where . Under reasonable assumptions, it follows that

so that consistently estimates the integrated variance for the ith price process, even in the presence of jumps. Thus, as such the contribution to the total variation coming from jumps may be estimated by , or the relative jump measure advocated by Huang and Tauchen (2005)

3. Jump test statistics

High-frequency data-based testing for jumps in a particular stock or financial instrument has received a lot of attention in the recent literature, and several different univariate test procedures exist for possible use. In this paper, however, we concentrate on the popular BN–S methodology developed in a series of influential papers starting with Barndorff-Nielsen and Shephard (2004).

Barndorff-Nielsen (born 18 March 1935 in Copenhagen) is a renowned Danish statistician who has contributed to many area of statistical science. He became interested in statistics when, as a student of actuarial mathematics, he worked part-time at the Department of Biostatistics of the Danish State Serum Institute. He graduated from the University of Aarhus (Denmark) in 1960, where he has spent most of his academic life, and where he became professor of statistics in 1973. However in 1962-1963 and 1963-1964 he stayed at the University of Minnesota and Stanford University, respectively, and from August 1974 to February 1975 he was an Overseas Fellow at Churchill College, Cambridge, and visitor at Statistical Laboratory, Cambridge University.

The BN–S methodology essentially kick-started the entire line of inquiry, and it is by far the most developed and widely applied of the different methods. In addition to the extensive Monte Carlo validation provided in Huang and Tauchen (2005), empirical studies that have used the approach include the work of Andersen et al. (2007b) and Todorov (2006) among many others. As such, the BN–S methodology arguably represents the standard approach for non-parametric univariate jump detection on a day-by-day basis.

All jump test statistics in the BN–S methodology work by forming a measure of the discrepancy between and and then studentizing the difference.

Barndorff-Nielsen and Shephard (2004) observe that a studentized version of the relative contribution measure can be expected to perform particularly well, since it largely mitigates the effects of level shifts in variance associated with time varying stochastic volatility. This conjecture is corroborated by the Monte Carlo evidence in Huang and Tauchen (2005), which indicates that the test statistic

where

And closely approximates a standard normal distribution under the null hypothesis of no jumps; Moreover, the statistic also exhibits favorable power properties and, in general, provides an excellent basis for univariate jump detection.

4. dataOur original sample consists of 25 component

stocks of the present Shanghai Stock Trade Institute over the August 1, 2006 to March 31, 2008, total include 2,482,760 data.

Throughout the paper our empirical work base on adopting the timing convention that one time unit corresponds to a trading day. For some loss data, we use linear interpolation method to ensure our results building on the basis of calendar day.

5. Empirical analysisIn this paper, we will use the volatility signature plots method advocated by Andersen,Bollerslev,Diebold,Labys (2000) .This method consider the sample mean of RV estimator as a function of sampling frequency over a long period. Long time scale reduce the effect of sampling frequency change, so as the market microstructure influence will essentially cease, the plots for the realized variation measures become flat. We can choose the optimal sampling frequency through it. Fig.1 depicts the changes on the days and average days of jumps of the 25 individual stocks, where, x-axis is the frequency of sampling, y-axis is the days of jumps.

020406080

100120140160180200220240

0 2 4 6 8 10 12 14 16 18 20 22

Fig.1. Number of jump days and the average for the 25 stocks in different sampling

Before presenting the result for all the 25 stocks and Shanghai Stock Composite Index (therefore SSCI), it is instructive to look at the different variation measures and the BN-S test for SSCI. To this end, Fig.2 plots the RV, BV variation measures, the relative jump contribution RJ, along with the BN-S z-statistic. Comparing the BN-S test statistic in the lower plot to the horizontal reference line for the 99.9% significance level included in the plot. We can find 6 large jumps for SSCI across the whole sample, and these 6 jumps are been considered as the co-jumps.

Fig.2. realized variance, bipower variation, relative jump, and z-statistic for SSCI from 2006 to 2008,M=59

(1)2007- 5- 30 val ue of z f or，i ndex=3. 3114

- 1

0

1

2

3

4

5

1 4 7 10 13 16 19 22 25

Ser i al Number Of stock

z

z=3. 08

(2)2007-8-22 val ue of z for，i ndex=3. 9570

- 2

- 1

0

1

2

3

4

5

1 3 5 7 9 11 13 15 17 19 21 23 25

Seri al Number Of stock

z

z=3. 08

(3)2007- 10- 8 val ue of z f or，i ndex=3. 8034

- 2

- 1

0

1

2

3

4

5

1 3 5 7 9 11 13 15 17 19 21 23 25

Seri al Number Of stock

z

z=3. 08

(4)2007-12-5 val ue of z for，i ndex=4. 2609

- 1

0

1

2

3

4

1 3 5 7 9 11 13 15 17 19 21 23 25

Ser i al Number Of stock

z

z=3. 08

(5)2007-12-21 val ue of z f or，i ndex=4. 7582

- 1

0

1

2

3

4

5

6

1 3 5 7 9 11 13 15 17 19 21 23 25

Ser i al Number Of stock

z

z=3. 08

(6)2008- 2- 19 val ue of z f or，i ndex=3. 6140

- 2

- 1

0

1

2

3

4

1 3 5 7 9 11 13 15 17 19 21 23 25

Ser i al Number Of stock

zz=3. 08

0

10

20

30

40

50

1 3 5 7 9 11 13 15 17 19 21 23 25

Ser i al Number Of Stock

Number Of Jump

number of total j umpsnumber of co- j umps

Fig.4. the relationship between the number of co-jumps and the number of total jumps

6. Conclusion We find the price jump be universal whether in

stock index or single stock. But there are distinct differences in the number of significant jumps between a large panel of high-frequency individual stock returns and SSCI. The co-jumps detected by z-statistic in SSCI is not obvious in the component stocks and a host of idiosyncratic jumps exist in single stocks.

We considered that the reason is that the effect of the co-jumps on individual has been involved in Integrated variance, in other word, the respond of individual stock to market information is seriously weaker than the response to own information, it means single stock is more sensitive to their own information.

Thank you