1: 

2: \documentclass{elsart}

3: 

4: \usepackage{graphics}

5: %\usepackage{graphicx}

6: \usepackage{epsfig}

7: \usepackage{amsmath}

8: \usepackage{amssymb}

9: \usepackage{multirow}

10: \usepackage{rotating}

11: 

12: \begin{document}

13: 

14: \begin{frontmatter}

15: 

16: 

17: \title{Value-at-Risk and Tsallis statistics: risk analysis of the aerospace sector}

18: 

19: \author{Adriana P. Mattedi$^1$, Fernando M. Ramos$^1$, Reinaldo R.

20: Rosa$^1$ }

21: \author{Rosario N. Mantegna$^2$ }

22: 

23: \address{$^1$ N\'ucleo para Simula\c{c}\~ao e An\'alise de Sistemas

24: Complexos - Laborat\'orio Associado de Computa\c{c}\~ao e

25: Matem\'atica Aplicada - Instituto Nacional de Pesquisas Espaciais

26: (INPE) - S\~ao Jos\'e dos Campos, Brazil}

27: 

28: \address{$^2$ Istituto Nazionale per la Fisica della Materia, Unit\`a di

29: Palermo and Dipartimento di Fisica e Tecnologie Relative,

30: Universit\`a di Palermo, Viale delle Scienze, I-90128, Palermo,

31: Italia}

32: 

33: 

34: \begin{abstract}

35: In this study, we analyze the aerospace stocks prices in order to

36: characterize the sector behavior. The data analyzed cover the

37: period from January 1987 to April 1999. We present a new index

38: for the aerospace sector and we investigate the statistical

39: characteristics of this index. Our results show that this index

40: is well described by Tsallis distribution. We explore this result

41: and modify the standard Value-at-Risk (VaR), financial risk

42: assessment methodology in order to reflect an asset which obeys

43: Tsallis non-extensive statistics.

44: \end{abstract}

45: 

46: \begin{keyword}

47: econophysics \sep aerospace sector \sep

48: Tsallis statistics \sep Value-at-Risk

49: % PACS codes here, in the form: \PACS code \sep code

50: \PACS 89.90.+n \sep 05.40 \sep

51: 

52: \end{keyword}

53: \end{frontmatter}

54: 

55: \section{Introduction}

56: 

57: Financial institutions are subject to many sources of risk.

58: Financial risk can be broadly defined as the degree of

59: uncertainty about future net returns. Accurate evaluation of

60: risks in financial markets is of crucial importance for the

61: proper assessment and efficient mitigation of risk and,

62: consequently, for optimal capital allocation. The increased

63: volatility of financial markets during the last decade has

64: induced the development of sophisticated risk management tools.

65: In this context, Value-at-Risk (VaR) has become the standard

66: measure that financial analysts use to quantify market risk. VaR

67: is defined as the maximum potential loss in a portfolio value due

68: to adverse market movements, for a given probability (5\% or 1\%,

69: for example) and a fixed time horizon (typically, 1 day). The

70: great popularity achieved by this instrument is essentially due

71: to its conceptual simplicity: VaR reduces the market risk

72: associated with any portfolio to just one number, the loss

73: associated to a given probability, under {\it normal} market

74: conditions.

75: 

76: Here, we analyze the financial risk of the aerospace sector. The

77: aerospace complex "weights" several hundreds of billions of

78: dollars a year, and creates millions of high-wage, high skill

79: jobs worldwide. Moreover, technologically-intensive aerospace

80: production generates extremely important positive spill over

81: effects for the rest of the economy. However, this very fact --

82: the need of large investments in advanced technologies -- makes

83: this industry highly susceptible to political and economic risks.

84: For example, if we consider the volatility as a risk measure, we

85: remark that, in the last decade, the stocks of the main aerospace

86: players presented a volatility 20\% higher than that displayed by

87: major financial indexes such as the Dow-Jones Industrial Index and

88: the Standard \& Poor's 500 Index (see below). Since until

89: presently, there is no single market indicator specially designed

90: for tracking the performance of the aerospace sector as a whole,

91: we introduce a new index, called Comprehensive AeroSpace Index

92: (CASI), based on the stocks negotiated on the New York Exchange

93: (NYSE) and the Over the Counter (OTC) markets. We investigate the

94: statistical characteristics of the CASI and show that this index

95: is well described by Tsallis distribution. We explore this result

96: and modify the standard Value-at-Risk (VaR) risk assessment

97: methodology in order to reflect an asset whose underlying

98: dynamics obeys the non-extensive statistics.

99: 

100: This paper is organized as follows. Following this Introduction,

101: in Section 2 we describe the data set used in this work, how the index CASI is computed, and

102: which are its main statistical

103: characteristics. In Section 3, we briefly introduce the VaR methodology,

104: then generalize it to non normal, Tsallis market conditions, and

105: apply this new methodology to the CASI time-series. Finally, our concluding remarks are

106: presented in Section 4.

107: 

108: \section{The Comprehensive Aerospace Index (CASI)}

109: 

110: The data set used in this paper consists of twelve years of daily

111: prices sets, from January, 1987 to April, 1999, from a

112: comprehensive sample of leading aerospace-related companies with

113: stocks negotiated on the New York Exchange (NYSE) and on the

114: Over-the-Counter (OTC) markets. The number of aerospace companies

115: included in the data set varied from 62 in 1987, to 146 in 1999.

116: The CASI was calculated using a base-weighted aggregate

117: methodology \cite{Nysemet,SPmet} in order to reflect the total

118: market value of each company. Following this approach, the CASI

119: at a given time $t$ is given by

120: \begin{equation}

121: \label{he1}

122: Y(t) = \frac{1}{ID} \, \sum_i^N (\frac {w_i}{W}) \, X_i (t),

123: \end{equation}

124: 

125: where $w_i$ is the number of shares outstanding of the stock $i$,

126: $W$ is the total number of companies shares that compose the

127: index, $X_i$ is the price of the stock $i$, and $N$ is the number

128: of stocks. ID is a normalizing factor called Index Divisor, which

129: permits to set the initial CASI value to an arbitrary level (say,

130: 100 or 10), allowing the index series to be comparable over time,

131: and to adjust for variations in the weights $w_i$. These weights

132: are up-dated every three months, or every time a difference of

133: 5\% or more in the number of shares outstanding of any stocks is

134: detected or when a new stock is included in the index. In any

135: case, the ID value is correspondingly adjusted to assure the

136: continuity in the CASI index values.

137: 

138: As usual, we considered in our analysis the time evolution of

139: CASI returns, defined as the logarithmic differences of CASI

140: values over a period of time $\Delta t$, computed for a sequence

141: of index values as $S_{\Delta t}(t) = \log(Y(t + \Delta t)/Y(t))$.

142: CASI returns, from 1987 to 1999, are shown in Fig. \ref{indice},

143: for a time delay of 1 day, together with the time evolution of

144: the index itself during the same period. The spectral

145: characteristics of both signals display some well-known features

146: of financial time-series, with CASI returns behaving like a

147: white-noise and the CASI index itself following a random-walk

148: pattern. We present in Fig. \ref{casi_pdf}, CASI return

149: histograms (in semi-logarithmic plots), for three time intervals.

150: 

151: \begin{figure}[ttt]

152: \begin{center}

153:     \mbox{\epsfig{file=f_1.eps,width=7cm,height=5cm}}

154:     \mbox{\epsfig{file=f_1_S.eps,width=6.5cm,height=5cm}}

155:     \caption{The CASI index: (a) daily evolution; (b) daily returns.}

156:     \label{indice}

157: \end{center}

158: \end{figure}

159: 

160: \begin{figure}[ttt]

161: \begin{center}

162:     \mbox{\epsfig{file=f_2.eps,width=9cm,height=7cm}}

163:     \caption{CASI return histograms, for $\Delta t$ (1, 5, 22 days).}

164:     \label{casi_pdf}

165: \end{center}

166: \end{figure}

167: 

168: 

169: These results follow a familiar pattern already found in other

170: financial signals \cite{MantStanN,Ghashghaie,RamosNA,Mattedi}. The

171: transition from an approximately Gaussian behavior, at long time

172: intervals, to a leptokurtotic distribution form, as $\Delta t$

173: decreases, is quite evident. At $\Delta t = 1$ day, the histogram

174: has a kurtosis considerably higher than 3, the value expected for

175: a normal process. This feature, which corresponds to an excess of

176: large fluctuations compared to a Gaussian distribution is a well

177: known signature of intermittence - bursts of intense market

178: volatility.

179: 

180: \begin{table}[bbb]

181: \begin{center}

182: \begin{tabular}{ccccc}

183:             \multicolumn{5} {c} {Table 1: Scaling coefficients} \\

184:             \multicolumn{5} {c} {} \\ \hline

185:            & \multicolumn{4} {c} {Indexes} \\ \cline{1-5}

186:            & CASI    & S\&P500  & DJI     & Nasdaq    \\ \hline

187:  Mean      & 0.96  & 0.99  & 0.96  & 0.98    \\ \hline

188:  Variance  & 1.00  & 0.93  & 0.96  & 1.09    \\ \hline

189:  Kurtosis  & -0.38 & -0.58 & -0.61 & -0.32   \\ \hline

190: \end{tabular}

191: \end{center}

192: \end{table}

193: 

194: Financial markets are extremely complex, with multiple

195: hierarchical structures and a large number of heterogeneous

196: agents interacting in an intricate way. When studying the

197: statistical characteristics of a financial index, an usual

198: approach is to investigate how the underlying distribution moments

199: (or other related quantities) scale for different time intervals.

200: In Fig. \ref{scaling} we present, on a log-log plot, mean,

201: variance and kurtosis as a function of $\Delta t$, for four

202: financial indexes: CASI, Nasdaq, Dow-Jones Industrial (DJI) and

203: Standard \& Poor's (S\&P500). We remark that all curves present a

204: power-law behavior (the corresponding scaling exponents are

205: displayed in Table 1), although the kurtosis results suffer from

206: the finiteness of the sample size. For all indexes, the variance

207: scales linearly with $\Delta t$, the expected result for a

208: standard diffusive regime. Note that CASI and Nasdaq have

209: kurtosis exponents quite different from those displayed by DJI

210: and S\&P500. We may speculate that this result indicates that the

211: dynamics of price formation in technology-intensive companies is

212: peculiar, very different from more traditional industries.

213: Surprisingly, we find that, although CASI and Nasdaq consistently

214: display higher variances, their kurtosis are smaller. In other

215: words, CASI and Nasdaq are more volatile but less intermittent

216: than DJI and S\&P500. This somewhat unexpected result suggests

217: the presence of long memory effects in CASI and Nasdaq volatility

218: patterns. These correlations tend to cluster periods of similar

219: volatility, preventing the excessive alternating of price changes

220: regimes.

221: 

222: \begin{figure}[ttt]

223: \begin{center}

224:     \mbox{\epsfig{file=f_3_mean.eps,width=8cm,height=6.5cm}}

225:     \mbox{\epsfig{file=f_3_var.eps,width=8cm,height=6.5cm}}

226:     \mbox{\epsfig{file=f_3_kur.eps,width=8cm,height=6.5cm}}

227:     \caption{Scaling with varying time intervals

228: of CASI, Nasdaq, DJI and S\&P500: (a) mean; (b) variance; (c)

229: kurtosis.}

230:     \label{scaling}

231: \end{center}

232: \end{figure}

233: 

234: \section{Risk Analysis}

235: 

236: In recent years, VaR has been adopted as the standard tool used by

237: institutions to measure and manage financial market risk.

238: Moreover, VaR is the basis for BIS (Bank for International

239: Settlements) market risk-based capital requirements

240: \cite{Jorion,Saunders}. Although, its origins can be trace back

241: as far as 1922, to capital requirements the NYSE (New York Stock

242: Exchange) imposed to member firms, passing through the portfolio

243: theory and a first crude VaR measure published in 1963

244: \cite{Holton}, VaR only became widely used with the J.P.Morgan’s

245: RiskMetrics Technical Document in 1995 (freely available on the

246: J.P.Morgan's website \cite{Morgan}). Great part of its popularity

247: originates from the fact it aggregates several risk components

248: into a single number. VaR measures the worst expected loss under

249: {\it {normal}} market conditions over a specific time interval

250: (which may vary from hours to years), at a given confidence level

251: (for example, 99\% or 95\%).

252: 

253: In general terms, VaR measures the quantile of the projected

254: distribution of gains and losses over a given time horizon. The

255: $q$-quantile (or fractile) of a random variable $X$ is any value

256: $x$ such that $Pr(X \leq x) = q$. If $c$ is the selected

257: confidence level, VaR is the corresponding $1-c$ lower-tail

258: quantile. Analytically, it can be formulated in terms of returns

259: of the portfolio as follows:

260: \begin{equation}

261: \label{varmean}

262: Pr[S_{\Delta t} < VaR_S(\Delta t)] = 1 - c~,

263: \end{equation}

264: where $VaR_S(\Delta t)$ is the VaR, in terms of returns, for a

265: time horizon of $\Delta t$.

266: 

267: It is clear from Eq. (\ref{varmean}), that the precise prediction

268: of the probability of an extreme movement in the value of a

269: portfolio is essential for risk management. By their very nature,

270: extreme movements are related to the tails of the distribution of

271: the underlying data generating process. Several studies, starting

272: with the pioneering work by Mandelbrot \cite{Mandelbrot}, have

273: shown that financial return distributions are leptokurtotic, that

274: is they have heavier tails and a higher peak than a normal

275: distribution. These features are well illustrated by CASI 1-day

276: return distribution, shown in Fig. \ref{casi_pdf}. In practical

277: terms, the assumption of normally distributed returns may lead to

278: serious errors in the estimation of risk, affecting the stability

279: of markets and/or the profitability of financial institutions.

280: 

281: In order to model fat-tailed distributions, the log-normal

282: distribution, generalized error distribution, truncated Levy

283: distribution, and mixtures of normal distributions have been

284: suggested \cite{MantStanlivro,Ramoscondmat,Michael,Ausloos,Liu}.

285: From an econophysics point-of-view, several studies pointed out

286: that the financial market dynamics may efficiently be understood

287: within the statistical mechanical framework and its suitable

288: generalizations \cite{Michael,Ausloos}. In this context, Ramos

289: and collaborators \cite{Ramoscondmat,RamosNA} have recently shown

290: that Tsallis canonical distribution \cite{Tsallis}, given by

291: \begin{equation}

292: \label{he10}

293: p_q(x) \sim (1 - \beta(1 - q)\, x^2\,)^{1/(1 -

294: q)}~,

295: \end{equation}

296: 

297: with $q$ and $\beta$ as parameters, provides a simple and accurate

298: model for financial histograms. Moreover, Tsallis entropic

299: parameter $q$ was found to be an excellent measure of

300: intermittency \cite{Bolzan}. Note that, for $1 < q < 5/3$, Eq.

301: (\ref{he10}) displays fat-tails but still has finite variance.

302: For $q=1$, we recover the Gaussian distribution. Finally, for

303: unit variance, we have $\beta = 1/(5-3q)$.

304: 

305: Relaxing the assumption of normal returns intrinsic to the

306: original VaR method, and assuming that market return

307: distributions follow Eq. (\ref{he10}), $VaR_S(\Delta t)$ can be

308: computed from

309: 

310: \begin{equation}

311: \label{he12}

312: Z_q^{-1} \, \int_{-\infty}^{-VaR_S(\Delta t)}

313: [1 - \beta(1 - q)S^2]^{1/(1 - q)}dS = 1-c~,

314: \end{equation}

315: where $q$ is a function of $\Delta t$ and, for unit variance,

316: $\beta = 1/(5-3q)$. The normalizing factor is given by

317: \begin{equation}

318: Z_q = \sqrt{\frac{\pi}{\beta (q-1)}} \,

319: \frac{\Gamma((3-q)/2(q-1))}{\Gamma(1/(q-1))}~.

320: \end{equation}

321: It is important to remark that, since the entropic parameter $q$

322: only depends on the kurtosis of the experimental histogram

323: \cite{Bolzan}, its value can be easily estimated for longer time

324: horizons. For this, it suffices to extrapolate the kurtosis value

325: of the 1-day histogram, assuming some known scaling with $\Delta

326: t$, as suggested by Fig. \ref{scaling}c.

327: 

328: Figure \ref{cdf_fig}a compares two theoretical cumulative density

329: functions (CDF), computed using Gaussian and Tsallis ($q=1.51$)

330: standardized return distributions, with CASI 1-day experimental

331: histograms. We observe that Tsallis distribution provides better

332: VaR estimates, for confidence levels within the range of 95 to

333: 99.5\%. This comparison may be further enhanced plotting the

334: difference between experimental and theoretical quantiles as a

335: function of the confidence level (\ref{cdf_fig}b). Note that for

336: high confidence levels (above 99\%), the Gaussian model grossly

337: underestimates the risk, while for low confidence levels (below

338: 97\%), it leads to sub-optimal capital allocation due to

339: systematic risk overestimation.

340: 

341: \begin{figure}[ttt]

342: \begin{center}

343:     \mbox{\epsfig{file=f_5b.eps,width=6.7cm,height=6.2cm}}

344:     \mbox{\epsfig{file=dif.eps,width=6.7cm,height=6.2cm}}

345: \caption{(a) Experimental ($\diamond$) and theoretical (Gaussian:

346: dashed line; Tsallis: solid line, $q=1.51$) cumulative density

347: functions (CDF); (b) Experimental and theoretical quantile

348: logarithmic differences as a function of the confidence level.}

349: \label{cdf_fig}

350: \end{center}

351: \end{figure}

352: 

353: 

354: \section{Conclusion}

355: 

356: In this paper, we studied the financial risk of the aerospace

357: sector. In spite of its undeniable economic, social and even

358: strategic importance, until presently, there was no single market

359: indicator specially designed for tracking the performance of the

360: aerospace sector as a whole. This fact led us to develop a new

361: index, called Comprehensive Aerospace Index (CASI), based on the

362: stocks negotiated on the New York Exchange (NYSE) and the Over

363: the Counter (OTC) markets. Our data set consisted of twelve years

364: of daily prices sets, from January, 1987 to April, 1999, from a

365: comprehensive sample of leading aerospace-related companies.

366: 

367: We investigated the statistical characteristics of the CASI and

368: found that it is more volatile but less intermittent than other

369: traditional market indicators, such as the DJI and S\&P500. This

370: results suggests the existence of long memory correlations

371: impacting the volatility clustering patterns of the CASI index.

372: We also shown that CASI histograms, for different time intervals,

373: are well described by Tsallis canonical distribution. We explored

374: this result and modified the standard Value-at-Risk (VaR) risk

375: assessment methodology in order to reflect an asset whose

376: underlying dynamics obeys Tsallis non-extensive statistics. Our

377: results have shown this approach provides better VaR estimates,

378: for confidence levels within the range of 95 to 99.5\%.

379: 

380: Summarizing, we explored in this paper the idea that the

381: comprehension of financial time-series statistical features may be

382: the first step to grasp the behavior of a highly complex system

383: such the financial market. Investment professionals have at their

384: disposal a wide array of sophisticated tools to aid their

385: analyzes, like the Value-at-Risk method. Many of these tools

386: depend on the assumption that the asset returns are normally

387: distributed. Today we know they are not. To emphasize the

388: importance of this point, we showed in our results that the

389: relaxation of normality premise can improve a risk measure, with

390: positive impacts on the stability of markets and the

391: profitability of financial institutions.

392: 

393: \section*{Acknowledgments}

394: This work was supported by FAPESP, CAPES and CNPq, Brazil.

395: 

396: \begin{thebibliography}{00}

397: 

398: \bibitem{Nysemet} New York Stock Exchange, {\it Nyse Composite Index methodology guide}. Jan. (2003).

399: \bibitem{SPmet}  Standard \& Poor's., {\it S\&P analysts' handbook} - 3o. chapter (1999).

400: \bibitem{MantStanN} R.N. Mantegna and H.E. Stanley, Nature 376, 46-49 (1995).

401: \bibitem{Ghashghaie} S. Ghashghaie, W. Breymann, J. Peinke, P. Talkner and Y. Dodge, Nature 381, 767-770 (1996).

402: \bibitem{RamosNA} F.M. Ramos, R.R. Rosa, C. Rodrigues Neto, J.A. Bolzan, L.D.A. S\'a, NonLinear Analysis 47,3521 (2001).

403: \bibitem{Mattedi} A.P. Mattedi, F.M. Ramos, R.R. Rosa, M.M. De Marchi, L.S. Bins. Anais do {\it{XXXIV}} Simp\'osio Brasileiro de Pesquisa Operacional. ISSN 1518-1731 (2002).

404: \bibitem{Jorion} P. Jorion, {\it Value at Risk: the new benchmark for controlling market risk}. New York: McGraw-Hill Companies Inc.(1997).

405: \bibitem{Saunders} A. Saunders, {\it Credit Risk Measurement: new approaches to Value at Risk and other paradigms}. John Wiley \& Sons Inc., New York, (1999).

406: \bibitem{Holton} G.A. Holton, Contigency Analysis, Working Papers http://econwpa.wustl.edu/eps/mhet/papers/0207/0207001.pdf (2002).

407: \bibitem{Morgan} J.P.Morgan. RiskMetrics: technical Document. Fourth Edition, New York, (1996).

408: \bibitem{Mandelbrot} B. Mandelbrot, The Journal of Business 36, 394-419, (1963).

409: \bibitem{MantStanlivro} R.N. Mantegna and H.E. Stanley, {\it An Introduction to Econophysics: Correlations and Complexity in Finance,\/}(Cambridge Univ. Press, Cambridge UK,2000).

410: \bibitem{Ramoscondmat} F.M. Ramos, C. Rodrigues Neto, R.R. Rosa, Preprint cond-mat/9907348 (1999).

411: \bibitem{Michael} F. Michael and M.D. Johnson, cond-mat/0108017 (2001).

412: \bibitem{Ausloos} M. Ausloos and K. Ivanova. Phys. Rev. E 68, 46122 (2003).

413: \bibitem{Liu} Y. Liu, P. Gopikrishnan, P. Cioeau, M. Meyer, C.-K.Peng, H.E. Stanley, Phys. Rev. E 60, 2, 1399 (1999).

414: \bibitem{Tsallis} C. Tsallis, J.Stat.Phys. 52 (1988).

415: \bibitem{Bolzan} J.A. Bolzan, F.M. Ramos, L.D.A. S\'a, C. Rodrigues Neto, R.R. Rosh, J. Geoph. Res. 107, 8063, (2002).

416: 

417: \end{thebibliography}

418: 

419: \end{document}

420: