Identifying Financial Risk Factors

Transcription

1 Identifying Financial Risk Factors with a Low-Rank Sparse Decomposition Lisa Goldberg Alex Shkolnik Berkeley Columbia Meeting in Engineering and Statistics 24 March 2016

2 Outline 1 A Brief History of Factor Models in Finance

3 A Brief History of Factor Models in Finance Market Model and CAPM (1960s) Arbitrage Pricing Theory and Statistical Models (1970s today) Fundamental Models (1970s today)

4 Market Model and CAPM In the 1960s, Jack Treynor (1930 ) and Bill Sharpe (1934 ) developed the Capital Asset Pricing Model, which relates security expected returns to market returns More than half a century after the appearance of [Treynor, 1962] and [Sharpe, 1964], the market model and CAPM remain central to quantitative finance

5 Market Model Security return R is a sum of a component due to a market factor M and a specific component ɛ R = Mβ + ɛ β is the sensitivity of security return to market return Specific returns ɛ are uncorrelated across securities As a consequence of these assumptions and others, the security covariance matrix Σ can be decomposed as a sum of a rank-one factor component and a diagonal security specific return component Σ = σ 2 Mβ β +

6 Arbitrage Pricing Theory and Multi-Factor Models In the 1970s, Stephen Ross (1944 ideas in the CAPM to allow for more factors ) expanded on [Ross, 1976] leads to a security covariance matrix that can be decomposed (using PCA) as a sum of a low-rank factor component and a diagonal security specific return component Σ = X F X +

7 Arbitrage Pricing Theory and Multi-Factor Models [Chamberlain and Rothschild, 1983] built on [Ross, 1976] by developing approximate factor models Their construction relies on asymptotic results and it leads to a covariance matrix decomposition as a sum of a low-rank factor component and a sparse security specific return component Σ = X F X + S but it has not caught on in practice

8 Fundamental Factor Models Barr Rosenberg founded Barra ( Barr and Associates ) in [Rosenberg, 1984] and [Rosenberg, 1985] rely on fundamental factors, reversing the roles of the known and unknown variables in a regression to estimate factor returns from pre-specified exposures (factor betas)

9 Fundamental Factor Models According to the fundamental models, Σ = L + L = X F X has low rank is a diagonal security specific return covariance matrix Barra s fundamental models dominate industry practice today

10 Outline 1 A Brief History of Factor Models in Finance

11 Examples of Risk Factors Market Equity styles Country, industry, currency Creditworthiness Prepayment sensitivity Liquidity Emerging factors: carbon reserves, cyberterrorism, longevity

12 Which Models Can Effectively Estimate Which Factors? Broad Narrow Persistent Easy Fundamental Transient Traditional PCA

13 Which Models are Expensive to Run? Human Intensive Machine Intensive Fundamental X Traditional PCA X

14 Outline 1 A Brief History of Factor Models in Finance

15 Low Rank Plus Sparse Decompositions of Covariance Matrices Inspired by sparse and low-rank decompositions developed in [Candès et al., 2011] and elsewhere as well as graphical lasso decompositions with origins in [Speed and Kiiveri, 1986] and [Yuan and Lin, 2007] [Chandrasekaran et al., 2012] develop a convex optimization that, under hypotheses, provides a latent factor decomposition: ˆΣ 1 Ŝ ˆL

16 Low Rank Plus Sparse Decompositions of Covariance Matrices The routine maximizes an objective function: ( Gaussian likelihood S L, ˆΣ ) λ (γ S 1 + tr(l)) PDC, which we solved with an algorithm developed in [Ma et al., 2013] There is no reliance on asymptotic theory, but there is a normality assumption

17 Low Rank Plus Sparse Decompositions of Covariance Matrices Suppose the inverse of a covariance matrix admits a low-rank plus sparse decomposition: Σ 1 = S L Then the covariance matrix also admits a low-rank plus sparse decomposition: Σ = L + S and the Woodbury formula transforms one decomposition to the other: S = S 1 and L = SLΣ

18 Outline 1 A Brief History of Factor Models in Finance

19 Specification N = 32 securities T = 260 observations (one year of daily data) K = 2 broad factors The market, with annualized volatility of 20% (long only factor: all securities have positive exposure) Creditworthiness with annualized volatility of 8% (long/short factor: have the securities are creditworthy, half are close to default) κ = 4 narrow factors China Argentina India Saudi Arabia

20 Input to Algorithm: Sample Covariance Matrix

21 Low-Rank Component of the Decomposition

22 Sparse Component of the Decomposition

23 True and Recovered Eigenvalues

24 Outline 1 A Brief History of Factor Models in Finance

25 Specification N = 32 securities T = 260 observations (one year of daily data) K =? broad factors Securities drawn from κ = 4 countries China Argentina India Saudi Arabia

26 Input to Algorithm: Sample Covariance Matrix, Oct 2015

27 Low-Rank and Sparse Decomposition: Covariance Matrices

28 Sample Correlation Matrix, Oct 2015

29 Low-Rank and Sparse Decomposition: Correlation Matrices

30 Recovered Eigenvalues

31 Outline 1 A Brief History of Factor Models in Finance

32 : What Would Barra Do if It Were Google? Fundamental risk models dominate the financial services industry PCA-based risk models have not been competitive We explore low-rank sparse decompositions of financial data The approach pioneered in [Chandrasekaran et al., 2012] identified and separated broad and narrow factors in simulated data and in empirical data Unlike traditional PCA, this algorithm does not rely on asymptotic results or rank orderings by eigenvalues

33 Ongoing Research Investigate the impact of the normality assumptions and, if appropriate, generalize the algorithm Automate the search for optimal calibration parameters Benchmark the performance of low-rank sparse models against fundamental and PCA-based models

34 Acknowledgement We thank State Street Global Exchange for financial and intellectual support

35 A Brief History of Factor Models in Finance Thank You Lisa Goldberg Alex Shkolnik Identifying Financial Risk Factors

36 References I [Candès et al., 2011] Candès, E. J., Li, X., Ma, Y., and Wright, J. (2011). Robust principal component analysis? Journal of the ACM, 58(3). [Chamberlain and Rothschild, 1983] Chamberlain, G. and Rothschild, M. (1983). Arbitrage, factor structure and mean-variance analysis on large asset markets. Econometrica, 51(5):

37 References II [Chandrasekaran et al., 2012] Chandrasekaran, V., Parillo, P. A., and Willsky, A. S. (2012). Latent variable graphical model selection via convex optimization. The Annals of Statistics, 40(4): [Ma et al., 2013] Ma, S., Xue, L., and Zou, H. (2013). Alternating direction methods for latent variable gaussian graphical model selection. Neural Computation, 25:

38 References III [Rosenberg, 1984] Rosenberg, B. (1984). Prediction of common stock investment risk. Journal of Portfolio Management, 11(1): [Rosenberg, 1985] Rosenberg, B. (1985). Prediction of common stock betas. Journal of Portfolio Management, 11(2):5 14. [Ross, 1976] Ross, S. A. (1976). The arbitrage theory of capital asset pricing. Journal of Economic Theory, 31:

39 References IV [Sharpe, 1964] Sharpe, W. (1964). Capital asset prices: A theory of market equilibrium under conditions of risk. Journal of Finance, 19(3): [Speed and Kiiveri, 1986] Speed, T. and Kiiveri, H. (1986). Gaussian markov distributions over finite graphs. The Annals of Statistics, 14(1):

40 References V [Treynor, 1962] Treynor, J. (1962). Toward a theory of market value of risky assets. Presented to the MIT Finance Faculty Seminar in 1962, finally published in 1999 in Asset Pricing and Portfolio Performance, Robert J. Korajczyk (editor) London: RiskBooks, [Yuan and Lin, 2007] Yuan, M. and Lin, Y. (2007). Model selection and estimation in the gaussian graphical model. Biometrika, 94(1):19 35.