Financial Academy under the Government of the Russian Federation Report Singular spectrum analysis for a forecasting of financial time series Speaker: Kozlov Alexander A. Moscow - 2009 Content list: • Introduction to nonlinear dynamics approach • Overview of the main methods (including SSA) • Financial time series analysis and forecasting: • Schlumberger Limited • Deutsche Bank • Honda Motor Co., Ltd. • Toyota Motor Corp. • Starbucks • BP plc. • Conclusions Introduction • • • Time series is a series of variable values taken in successive periods of time. Time series analysis is a part of nonlinear dynamics. Supposition: market of shares is unstable and chaotic. • Objective: analysis and forecasting of stock price time series with nonlinear dynamics methods • In this report the following questions will be considered: • Embedding dimension as “space” characteristic and its estimation • К2-entropy and Lyapunov exponents as “time” characterics and their estimation • SSA forecasting method Overview • The idea of attractor reconstruction [11]: Satisfactory geometry picture of low-dimensional strange attractor can be obtained if instead of x-variables from dynamic system equations somebody use k-dimensional delay vectors: zi {xi , xi 1 ,..., xi ( k 1) } • Takens theorem [2]: There is a transformation that k 2d 1 . which can embed M d to R k on conditions It means that: - k – embedding dimension; - xi F ( xi 1, xi 2 ,..., xi k ) __________________________________________________________________________________________________ [1] Packard N.H., Crutchfield J.P., Farmer J.D., Shaw R.S.,"Geometry from a time series", Phys.Rev.Lett. 45, p.712,1980. [2] F. Takens, "Dynamical Systems and Turbulence", Lect. Notes in Math, Berlin, Springer. №898, 1981, p. 336. Overview •Correlation integral on r<<1 and k>>1 [4]: Grassberger- Procaccia method [3]: C2 ( r ) 1 N2 H r z i zj N i , j 1 -1 -2 k -4 -5 -6 stops 3. This k-number is embedding dimension k -3 1. Find k , having curves for each k, starting with k=1; 2. Starting with certain k-number growing and stabilizes; 0 ln C k • ln Cq (r , w) Dq ln r K q k const -7 -8 de; -4 4. Maximum value of k is a so-called correlation dimension d c(or D2 ) of the attractor. -3 -2 0 1 ln r 2.6 2.4 -1 dc 2.2 2.0 • Limitation [4]: k 1.8 1.6 1.4 d 2 lg N 1.2 1.0 de 0.8 1 2 3 4 k ______________________________________________________________________________________ [3] P. Grassberger, I. Procaccia, "Characterization of Strange Attractors",Phys.Rev.Lett.,50,346, 1983 [4] G.G. Malineckiy, A.B. Potapov, “Actual problems of nonlinear dynamics", М: URSS, 2002 5 6 7 Overview • • Having fixed r and investigating dependence С(r*,k) from k (k>>1), somebody can estimate K2-entropy [5]. K2 defines the time of predictability for the system in “volume” interpretation (growing of the volume in phase space which the system can occupy in the future) C2 (r , w) ~ r D 2 exp( K 2 k ) • T (1) ~ K 2 1 The time of predictability also can be determined from Lyapunov exponents The maximal one is estimated in Wolf method [6]. 1 1 t n t0 i L(ti 1 ) ln L(ti ) i 0 n 1 T ( 2 ) ~ 1 1 ________________________________________________________________________________________________________ [5] Grassberger, I. Procaccia,"Estimation of the Kolmogorov Entropy from a Chaotic Signal",Phys.Rev.A,vol.28,4,1983,p.2591 [6] Wolf A., Swift J.B., Swinney H.L., "Determining Lyapunov exponents from time series", Physica D, 69 (1985), №3, p.285-317. Overview SSA forecasting method [7]: • 1) Construction of the delay matrix from time series and preliminary changes in it (centering and normalization) • • X M ( N M 1) xN M 2 xN M 1 xM 1 x3 x2 xN 2) Finding the components (M) and selection of the most important ones (r) This vM(1) vM( 2) vM( M ) is equal to search of 0 1 eigenvectors and и eigenvalues 2 Λ VM M (1) of the matrix X XT . ( 2) (M ) v v v 2 v (1) 1 2 ( 2) 1 2 (M ) 1 v v 3) Time series reconstruction with r main components ˆ V VT X and taking average on each diagonal. X M r • xM x 2 x 1 0 M M r 4) Forecast constraction with «caterpillar» method: xˆ N 1 Equal to constraction of the new delay vector with one unknown x N ˆ x N 1 coordinate. (xˆ N 1 VM r VMT r xˆ N 1 ) 2 min xˆ N 1 x N M 2 _____________________________________________________________________________________________________ [20] “The main components of time series: “caterpillar“ method”. Col.articles // ed. D.L. Danilov, А.А. Zhiglyavskiy – St.P.: St.P. University, 1997. - 308 p. Analysis and forecasting • Criteria for the selected companies: - long time on the market of shares (NYSE) – more than 10 years; - publicity; - from different sectors; • Thus the following companies were chosen: • Schlumberger Limited • Deutsche Bank • Honda Motor Co., Ltd. • Toyota Motor Corp. • Starbucks • BP plc. • Forecasting parameters: - delay number - M=20 - number of the main components – • r de During forecasting logarithmic profit is taken in to account: - positive in growth - negative in fall Si ln x(ti ) x(ti 1 ) 1. Schlumberger Limited •Period from 31.12.1981 to 31.12.2008 •Time series consists of 6814 stock price values (on close). k -2 0 3,0 -2,5 2,5 -3 ln C (k) k Analysis and forecasting 2,0 2 4 6 8 10 y = -0,1656x - 2,8552 R2 = 1 -3,5 -4 1,5 -4,5 1,0 -5 1 2 3 de 7 4 5 6 k 7 8 9 10 d c 3,12 11 K2 0,1656 T (1) 6,04 1 0,1597 T ( 2) 6,26 12 Analysis and forecasting 1. Schlumberger Limited 2. Deutsche Bank •Period from 18.11.1996 to 31.12.2008. •Time series consists of 3033 stock price values (on close). k -1 3,5 -1,2 0 2 4 6 8 10 -1,4 3,0 -1,6 2,5 ln C (k) k Analysis and forecasting 2,0 y = -0,1208x - 1,4495 -1,8 2 R =1 -2 -2,2 1,5 -2,4 1,0 -2,6 1 2 3 de 6 4 5 6 k 7 8 9 10 d c 3,26 11 -2,8 K2 0,1208 T (1) 8,28 1 0,1164 T ( 2) 8,58 12 Analysis and forecasting 2. Deutsche Bank Analysis and forecasting 3. Honda Motor Co., Ltd. •Period from 11.08.1987 to 31.12.2008. •Time series consists of 5390 stock price values (on close). 3,0 k -2 0 2 4 6 8 10 -2,5 2,5 -3 y = -0,1495x - 3,1127 k ln C (k) 2,0 2 R = 0,9997 -3,5 1,5 -4 1,0 -4,5 -5 1 2 3 de 7 4 5 6 k 7 8 9 10 d c 2,80 11 K2 0,1495 T (1) 6,69 1 0,1442 T ( 2) 6,94 12 Analysis and forecasting 3. Honda Motor Co., Ltd. Analysis and forecasting 4. Toyota Motor Corp. •Period from 13.04.1993 to 31.12.2008. •Time series consists of 3956 stock price values (on close). k 0 0 2 4 6 8 10 -0,5 2,5 -1 k ln C (k) 2,0 -1,5 y = -0,1175x - 1,8895 R2 = 0,9999 -2 1,5 -2,5 1,0 -3 -3,5 1 2 3 de 6 4 5 6 k 7 8 9 10 d c 2,64 11 K2 0,1175 T (1) 8,51 1 0,1222 T ( 2) 8,18 12 Analysis and forecasting 4. Toyota Motor Corp. Analysis and forecasting 5. Starbucks •Period from 26.06.1992 to 31.12.2008. •Time series consists of 4161 stock price values (on close). k -3 0 2 4 6 8 10 2,5 -3,5 2,0 ln C (k) k -4 1,5 -4,5 1,0 0,5 y = -0,1389x - 3,9295 R2 = 0,9998 -5 1 2 3 de 7 4 5 6 k 7 8 9 10 11 d c 2,34 -5,5 K2 0,1389 T (1) 7,20 1 0,1342 T ( 2) 7,45 12 Analysis and forecasting 5. Starbucks Analysis and forecasting 6. BP plc. •Period from 03.01.1977 to 31.12.2008. •Time series consists of 8076 stock price values (on close). k -2,5 0 -3 2,0 -3,5 k ln C (k) 2,5 1,5 4 6 8 10 y = -0,1403x - 3,0905 R2 = 1 -4 -4,5 1,0 0,5 2 -5 1 2 3 de 6 4 5 6 k 7 8 9 10 d c 2,45 11 K2 0,1403 T (1) 7,13 1 0,1369 T ( 2) 7,31 12 Analysis and forecasting 6. BP plc. Analysis and forecasting •Final results of analysis are in the table: Компания de dc K2 1 Schlumberger Limited 7 3,12 0,1656 0,1597 6,04 6,26 Deutsche Bank 6 3,26 0,1208 0,1164 8,28 8,58 Honda Motor Co. Ltd. 7 2,8 0,1495 0,1442 6,69 6,94 Toyota Motor Corp. 6 2,64 0,1175 0,1222 8,51 8,18 Starbucks 7 2,34 0,1389 0,1342 7,20 7,45 BP plc. 6 2,45 0,1403 0,1369 7,12 7,31 T (1) K2 T ( 2) •Percentage of coincidence between logarithmic profit signs of forecast and real time series Компания 2007 , % 2008 , % Schlumberger Limited 75 75 Deutsche Bank 81 69 Honda Motor Co. Ltd. 75 58 Toyota Motor Corp. 75 81 Starbucks 86 57 BP plc. 71 71 Conclusions • Nonlinear dynamics methods applied to stock price time series led to a “space” and “time” analysis of the trading system. Thus we determined number of the main components (=embedding dimension) and time of predictability (according to K2-entropy and Lyapunov exponents) for each company. • Obtained results have both fundamental and applied sense for economics. • Complex analysis permitted to make a forecast on the basis of SSA method (“caterpillar”). Forecasted values and logarithmic profit fits the real ones quite well. • Thus SSA forecasting method can be a useful instrument in quantitative analysis of any risks connected with financial time series. Thank You for attention!