TEDAS portfolios: parameters optimisation and out-of - Hu
Transcription
TEDAS portfolios: parameters optimisation and out-of - Hu
TEDAS - Tail Event Driven ASset Allocation: equity and mutual funds’ markets Wolfgang Karl Härdle Sergey Nasekin Alla Petukhina Ladislaus von Bortkiewicz Chair of Statistics C.A.S.E. – Center for Applied Statistics and Economics Humboldt–Universität zu Berlin http://lvb.wiwi.hu-berlin.de http://case.hu-berlin.de Motivation 1-1 Strategies comparison: hedge funds’ indices Comparison of Strategies 9 8 Cumulative Return 7 6 5 4 3 2 1 0 2006 2007 2008 2009 2010 2011 2012 2013 2014 Time Figure 1: Strategies’ cumulative returns’ comparison: TEDAS Basic, S&P500, TEDAS Naïve, RR Comparison of Strategies 1.6 Cumulative Return 1.5 TEDAS - Tail Event Driven Asset allocation 1.4 1.3 1.2 1.1 1 0.9 March May July Time September Motivation 1-2 ⇢ Härdle et al. (2014) I TEDAS applied to hedge funds’ indices performs better than benchmark models ⇢ Limitation of using hedge indices as portfolio assets ⇢ Application of TEDAS approach to global mutual funds’ data and German stock market Comparison of Strategies 1.6 Cumulative Return 1.5 TEDAS - Tail Event Driven Asset allocation 1.4 1.3 1.2 1.1 1 0.9 March May July Time September Motivation 1-3 Core & Satellites Mutual funds, SDAX, MDAX and TecDAX constituents ⇢ diversification - reduction of the portfolio risk ⇢ construction - a more diverse universe of assets ⇢ allocation - a higher risk-adjusted return. Comparison of Strategies 1.6 Cumulative Return 1.5 TEDAS - Tail Event Driven Asset allocation 1.4 1.3 1.2 1.1 1 0.9 March May July Time September Motivation 1-4 ⇢ Comparison of the TEDAS with more benchmark strategies: I 60/40 portfolio I Risk Parity (equal risk portfolio contribution) I Mean-Variance strategy ⇢ TEDAS parameters optimisation Comparison of Strategies 1.6 Cumulative Return 1.5 TEDAS - Tail Event Driven Asset allocation 1.4 1.3 1.2 1.1 1 0.9 March May July Time September Motivation 1-5 Tail Risk −0.25 −0.2 −0.15 −0.1 −0.05 0 S&P Returns 0.05 0.1 0.15 0.2 Figure 2: Estimated density of S&P 500 returns Comparison of Strategies 1.6 Cumulative Return 1.5 TEDAS - Tail Event Driven Asset allocation 1.4 1.3 1.2 1.1 1 0.9 March May July Time September Motivation 1-6 The TLND challenge ⇢ Tail dependence ⇢ Large universe: p > n ⇢ Non normality ⇢ Dynamics Comparison of Strategies 1.6 Cumulative Return 1.5 TEDAS - Tail Event Driven Asset allocation 1.4 1.3 1.2 1.1 1 0.9 March May July Time September Motivation 1-7 TEDAS Objectives ⇢ Hedge tail events I Quantile regression I Variable selection in high dimensions ⇢ Improve Asset Allocation L N T D I Higher-order moments’ optimization I Modelling of moments’ dynamics Comparison of Strategies 1.6 Cumulative Return 1.5 TEDAS - Tail Event Driven Asset allocation 1.4 1.3 1.2 1.1 1 0.9 March May July Time September Outline 1. Motivation X 2. TEDAS framework 3. Data 4. Empirical Results 5. Discussion: choice of ⌧ -spine 6. Conclusions TEDAS framework 2-1 Tail Events ⇢ Y 2 Rn core log-returns; X 2 Rn⇥p satellites’ log-returns, p>n ⇢ def q⌧ (x) = FY |x1 (⌧ ) = x > (⌧ ) = arg minp EY |X =x ⇢⌧ {Y 2R ⇢⌧ (u) = u{⌧ X > }, I(u < 0)} ⇢ L1 penalty n kˆ ! > k1 to nullify "excessive" coefficients; and ! ˆ controlling penalization; constraining 0 yields ALQR Details ˆadapt = arg min ⌧, n p 2R n X ⇢⌧ (Yi i=1 Xi> ) + ! n kˆ > n (1) k1 Comparison of Strategies 1.6 Cumulative Return 1.5 TEDAS - Tail Event Driven Asset allocation 1.4 1.3 1.2 1.1 1 0.9 March May July Time September TEDAS framework 2-2 TEDAS Step 1 Initial wealth W0 = $1, t = l, . . . , n; l = 120 length of the moving window ⇢ Portfolio constituents’ selection 1. determine core asset return Yt , set ⌧t = Fbn (Yt ) ⌧j=1,...,5 = (0.05, 0.15, 0.25, 0.35, 0.50) 2. ALQR for ˆ⌧t , n using the observations X 2 Rt Y 2 Rt l+1,...,t l+1,...,t⇥p , 3. Select ⌧j,t according to the right-side q̂⌧j,t in: rt q̂⌧1,t or q̂⌧1,t < rt q̂⌧j,t Comparison of Strategies 1.6 Cumulative Return 1.5 TEDAS - Tail Event Driven Asset allocation 1.4 1.3 1.2 1.1 1 0.9 March May July Time September TEDAS framework 2-3 TEDAS Step 1 TEDAS framework TEDAS Example TEDAS framework 2-6 TEDAS Example 1. Suppose = 86 (Feb. 2007), W86 = framework 1. Suppose t = 86 (Feb. TEDAS 2007), W86 =t $1.125, accumulated W86 = Example wealth W86 = $1.429, TEDAS Y161wealth = 1.36% < 0$1.429, Y161 = 1.3 𝒀 = −𝟎. 𝟎𝟏 → 𝝉 𝟏𝟐𝟎 Suppose t = 86 (Feb. 2007), W86 = $1.125, b=n𝟏𝟐𝟎 �𝒏 (𝒀161 2.1. Fwealth (Y 0.35, so estimate ˆ0.35 2. Fbn (Y161 ) = 0.35, so estimate 𝑭 𝟎. 𝟒𝟓 𝟏𝟐𝟎 ))== W86 = $1.429, Y161 = 1.36% < 0 choose assets with 60⇥163 60 y bn (Y161 )on 3. ALQR X 2 R , Y 2 R 2. F = 0.35, so estimate TEDAS framework ˆ ,0.35 < 0 > ˆchoose assets with ( 0.77, 1.12, 0.41) , La 0.25 = TEDAS Example ˆ ,0.3560< 0 3. ALQR on X 2 R60⇥163 , YNorth 2 R America yields Macro, Emerging M ˆ with �for ALQR Suppose t= 86American W86 = $1.12 > ,𝜷 ,0.35 ALQR for 𝑿(Feb. 2007), ˆ0.25 = ( 0.77, 1.12,1. 0.41) 𝟎.𝟒𝟓,𝝀 Latin Arbitrage, Futures with XW and Y 𝒀 from the and wealth = from $1.429, Y161 = 1.36% < 0 86 moving North America Macro, Emerging Markets window the moving windowCTA/Managed 4. FbTEDAS CF-CVaR optimization De 2. 0.35, so estimate Futures n (Y161 ) = 60⇥163 60 3. choose ALQR on X 2with R with Choose assets Y 2R yields >; X b87 = (0 b86 = (0.22, 0.16, ,0.62) w =<(Details 1.12, >0.41)> , Latin Amer 4. TEDAS CF-CVaR optimization yields ˆˆ0.25 � 0 0.77,(1 ,0.35 = 𝜷 b86 𝟎.𝟒𝟓,𝝀 W W +w Xb87 ) =Markets $1.438 87 < 𝟎 86 Macro, Emerging b86 = (0.22, 0.16, 0.62)> ; North w Xb87 =America (0.38%, 0.45%, 0.76%)> , CT Futures Y 3. ALQR 2 R60⇥163 , Y 2 R60 yields 87 =on X1.53% > b b86 X87 ) = $1.438 (return of 0.62%) >while 87 = W86 (1 + w Window'length=120' ˆ0.25 Window length l W = 60 = (CF-CVaR 0.77, 1.12, 0.41) , Latin Am 4. TEDAS optimization Details yields Y87 = 1.53% North America Macro, >Emerging Markets b TEDAS Event 0.16, Driven0.62) Asset allocation b86- Tail w = (0.22, ; X87 = (0.38%, 0.4 TEDAS - Tail Event Driven Asset allocation >X Futures b87 )Asset TEDAS - Tail allocation b86 W87 = WEvent w = $1.438 (return of 86 (1 + Driven allocation Y87 = CF-CVaR 1.53% 4. TEDAS optimization Details yie Comparison of Strategies 1.6 Cumulative Return 1.5 Comparison of Strategies 1.6 TEDAS - Tail Event Driven Asset Cumulative Return 1.5 1.4 1.3 1.2 1.1 1 1.4 0.9 March 1.3 May July September Time 1.2 1.1 1 0.9 March May July September Time b86 = (0.22, 0.16, 0.62)> ; Xb87 = (0.38%, 0 w TEDAS framework 2-4 TEDAS Step 2 ⇢ Portfolio selection bt 2 Rk 1. apply TEDAS Gestalt to Xj , obtain w 2. determine the realized portfolio wealth for t + 1, def bt> Xbt+1 ) Xbt+1 = (Xt+1,1 , . . . , Xt+1,k )> : Wt+1 = Wt (1 + w Comparison of Strategies 1.6 Cumulative Return 1.5 TEDAS - Tail Event Driven Asset allocation 1.4 1.3 1.2 1.1 1 0.9 March May July Time September TEDAS framework 2-5 Rebalancing of portfolio: ⇢ one of inequalities in step 3 holds I sell the core portfolio and buy satellites (step 4) with estimated weights (step 5) I stay "in cash" if there are no adversely moving satellites (step 4) ⇢ no one of inequalities holds: invest in the core portfolio ⇢ period (t+1), if no one of inequalities (step 3) holds, we return to the core portfolio Comparison of Strategies 1.6 Cumulative Return 1.5 TEDAS - Tail Event Driven Asset allocation 1.4 1.3 1.2 1.1 1 0.9 March May July Time September TEDAS framework 2-6 TEDAS Example 1. Suppose t = 161 (May. 2011), accumulated wealth W161 = $2.301, Y161 = 1.36% < 0 2. Fbn (Y161 ) = 0.35, so estimate ALQR for ˆ ,0.35 3. ALQR on X 2 R120⇥583 , Y 2 R120 yields ˆ0.35 = ( 1.12, 0.41)> , Blackrock Eurofund Class I, Pimco Funds Long Term United U.S. States Government Institutional Shares b161 = (0.014, 0.026)> ; 4. TEDAS CF-CVaR optimization w > X b161 ) = $2361 b161 Xb162 = (0, 1)> , W162 = W161 (1 + w Comparison of Strategies 1.6 Cumulative Return 1.5 TEDAS - Tail Event Driven Asset allocation 1.4 1.3 1.2 1.1 1 0.9 March May July Time September TEDAS framework 2-7 TEDAS Gestalten TEDAS gestalt TEDAS Naïve Dynamics modelling NO Weights optimization Equal weights TEDAS Hybrid NO Mean-variance optimization of weights Details DCC volatility CF-VaR optimization TEDAS Basic Details Details Comparison of Strategies 1.6 Cumulative Return 1.5 TEDAS - Tail Event Driven Asset allocation 1.4 1.3 1.2 1.1 1 0.9 March May July Time September Data 3-1 Small and mid caps German stocks ⇢ MDAX I 50 medium-sized German public limited companies and foreign companies primarily active in Germany from traditional sectors I Ranks after the DAX30 based on market capitalisation and stock exchange turnover ⇢ SDAX I The selection index for smaller companies from traditional sectors I 50 stocks from the Prime Standard ⇢ TecDAX I Comprises the 30 largest technology stocks below the DAX Comparison of Strategies 1.6 Cumulative Return 1.5 TEDAS - Tail Event Driven Asset allocation 1.4 1.3 1.2 1.1 1 0.9 March May July Time September Data 3-2 Size premium ⇢ Banz (1981) and Reinganum (1981): the US small cap stocks outperformed large-cap stocks (in 1936-1975) ⇢ Fama, French (1992, 1993): a size premium of 0.27% per month in the US over the period 1963-1991 ⇢ Results are robust: I for stock price momentum by Jegadeesh , Titman (1993) and Carhart (1997) I for liquidity by Pastor, Stambaugh (2003) and Ibbotson, Hu (2011) I for industry factors, high leverage, low liquidity by Menchero et al. (2008) Comparison of Strategies 1.6 Cumulative Return 1.5 TEDAS - Tail Event Driven Asset allocation 1.4 1.3 1.2 1.1 1 0.9 March May July Time September Data 3-3 Why small and mid cap stocks? ⇢ Strong absolute returns ⇢ Diversification benefits (Eun, Huang, Lai (2006)) ⇢ High risk-adjusted returns Comparison of Strategies 1.6 Cumulative Return 1.5 TEDAS - Tail Event Driven Asset allocation 1.4 1.3 1.2 1.1 1 0.9 March May July Time September Data 3-4 Strong absolute returns 1.2 1.1 1 0.9 0.8 0.7 0.6 0.5 0.4 0.3 2007 2008 2009 2010 2011 2012 2013 2014 Figure 3: Cumulative index performance: MSCI World Large Cap, MSCI World Mid Cap, MSCI World Small Cap, MSCI World Small and Mid Cap Comparison of Strategies 1.6 Cumulative Return 1.5 TEDAS - Tail Event Driven Asset allocation 1.4 1.3 1.2 1.1 1 0.9 March May July Time September Data 3-5 Diversification 0.04 0.03 0.02 Return 0.01 0 −0.01 −0.02 −0.03 −0.04 −0.05 Jan Feb Mar Apr May Jun Jul Aug Time Figure 4: DAX and Hamborner REIT AG daily returns in 2013122020140831 Comparison of Strategies 1.6 Cumulative Return 1.5 TEDAS - Tail Event Driven Asset allocation 1.4 1.3 1.2 1.1 1 0.9 March May July Time September Data 3-6 German equity I 125 stocks - SDAX (48), MDAX (47) and TecDAX (50) as on 20140801 I DAX index 2 1.8 Cumulative Return ⇢ Frankfurt Stock Exchange (Xetra), weekly data 1.6 1.4 1.2 1 DAX MDAX 0.8 SDAX TecDAX 0.6 Feb11 Sep11 Apr12 Oct12 May13 Time Nov13 Jun14 Dec14 ⇢ Span: 20121221 - 20141127 (100 trading weeks) ⇢ Source: Datastream Comparison of Strategies 1.6 Cumulative Return 1.5 TEDAS - Tail Event Driven Asset allocation 1.4 1.3 1.2 1.1 1 0.9 March May July Time September Jul15 Data 3-7 Mutual Fund ⇢ Open-End: buy and sell the shares, meet the demand for customers ⇢ Unit Investment Trust: exchange-traded fund (ETF), Fixed/ unmanaged Portfolio ⇢ Closed-End: fixed number of shares, not redeemable by the fund, buy and sell on the exchange Comparison of Strategies 1.6 Cumulative Return 1.5 TEDAS - Tail Event Driven Asset allocation 1.4 1.3 1.2 1.1 1 0.9 March May July Time September Data 3-8 Mutual fund flowchart Comparison of Strategies 1.6 Cumulative Return 1.5 TEDAS - Tail Event Driven Asset allocation 1.4 1.3 1.2 1.1 1 0.9 March May July Time September Data 3-9 Why Mutual Funds? ⇢ Importance of MF I $30 trillion worldwide, 15 trillion in U.S in 2013 I 88% investment companies managed asset by holding MF ⇢ Big data: 76 200 MFs worldwide in 2013 ⇢ Diversification Structure of U.S. Mutual funds, by asset classes Comparison of Strategies 1.6 Cumulative Return 1.5 TEDAS - Tail Event Driven Asset allocation 1.4 1.3 1.2 1.1 1 0.9 March May July Time September Data 3-10 Dynamics of Mutual funds investment TNA, bln US$ Number of Mutual funds 80,000 70,000 60000 50,000 1998 2000 2002 2004 2006 2008 2010 2012 2014 40,000 30,000 20,000 10,000 0 1998 2000 2002 2004 2006 2008 2010 2012 2014 Year Figure 5: Worldwide Mutual Funds: total number and TNA Comparison of Strategies 1.6 Cumulative Return 1.5 TEDAS - Tail Event Driven Asset allocation 1.4 1.3 1.2 1.1 1 0.9 March May July Time September Data 3-11 Mutual Funds I Core asset (Y ): S&P500 I Satellite assets (X ): 583 Mutual funds ⇢ Span: 19980101 - 20131231 ⇢ Source: Datastream Sample Mutual Funds' Cumulative Gain 9 8 7 Stock Investment Value ⇢ Monthly data 6 5 4 3 2 1 0 1997 2000 2002 2005 2007 2010 2012 2015 Comparison of Strategies 1.6 Cumulative Return 1.5 TEDAS - Tail Event Driven Asset allocation 1.4 1.3 1.2 1.1 1 0.9 March May July Time September Data 3-12 Benchmark Strategies 1. RR: dynamic risk-return optimization Details 2. ERC: Risk-parity portfolio (equal risk contribution) 3. 60/40 portfolio Details Details Comparison of Strategies 1.6 Cumulative Return 1.5 TEDAS - Tail Event Driven Asset allocation 1.4 1.3 1.2 1.1 1 0.9 March May July Time September Empirical Results 4-1 Return TEDAS approach:German stocks’ results 1.5 1.5 1.5 1.4 1.4 1.4 1.3 1.3 1.3 1.2 1.2 1.2 1.1 1.1 1.1 1 1 1 0.9 0.9 0.9 0.8 Jan14 Apr14 Aug14 Dec14 0.8 Jan14 Apr14 Aug14 Dec14 0.8 Jan14 Apr14 Aug14 Dec14 Figure 6: Strategies’ cumulative returns’ comparison: TEDAS Basic, TEDAS Naïve, TEDAS Hybrid, DAX30 TEDAS_strategies Comparison of Strategies 1.6 Cumulative Return 1.5 TEDAS - Tail Event Driven Asset allocation 1.4 1.3 1.2 1.1 1 0.9 March May July Time September Empirical Results 4-2 Return TEDAS approach:German stocks’ results 1.6 1.6 1.6 1.4 1.4 1.4 1.2 1.2 1.2 1 1 1 0.8 Jan14 Apr14 Aug14 0.8 Dec14 Jan14 Apr14 Aug14 0.8 Dec14 Jan14 Apr14 Aug14 Dec14 Figure 7: Strategies’ cumulative returns’ comparison (with transaction costs 1% of portfolio value): TEDAS Hybrid, 60/40, ERC, RR TEDAS_strategies Comparison of Strategies 1.6 Cumulative Return 1.5 TEDAS - Tail Event Driven Asset allocation 1.4 1.3 1.2 1.1 1 0.9 March May July Time September Empirical Results 4-3 Strategies’ performance: German stocks Strategy TEDAS Basic TEDAS Naïve TEDAS Hybrid RR ERC 60/40 DAX30 Cumulative return 144% 143% 143% 108% 129% 121% 103% Sharpe ratio 0.3792 0.3184 0.3079 0.0687 -0.0693 0.0306 0.0210 Maximum drawdown 0.1069 0.0564 0.1068 0.0934 0.1792 0.0718 0.1264 Comparison of Strategies 1.6 Cumulative Return 1.5 TEDAS - Tail Event Driven Asset allocation 1.4 1.3 1.2 1.1 1 0.9 March May July Time September Empirical Results 4-4 TEDAS approach:German stocks’ results τ = 0.05 τ = 0.15 10 10 8 8 6 6 4 4 2 0 2 0 2 4 6 8 0 10 0 2 4 τ = 0.25 10 10 8 8 6 6 4 4 2 2 0 0 2 4 6 8 10 6 8 10 τ = 0.35 6 8 10 0 0 2 4 Figure 8: Frequency of the number of selected variables for 4 different ⌧ Comparison of Strategies 1.6 Cumulative Return 1.5 TEDAS - Tail Event Driven Asset allocation 1.4 1.3 1.2 1.1 1 0.9 March May July Time September Empirical Results 4-5 TEDAS approach:German stocks results τ = 0.05 τ = 0.15 15 15 10 10 5 5 0 0 0 50 100 150 0 τ = 0.25 50 100 150 τ = 0.35 15 15 10 10 5 5 0 0 0 50 100 150 0 50 100 150 Figure 9: The frequency of stocks Comparison of Strategies 1.6 Cumulative Return 1.5 TEDAS - Tail Event Driven Asset allocation 1.4 1.3 1.2 1.1 1 0.9 March May July Time September Empirical Results 4-6 Selected Stocks Table 1: The selected German Stocks for ⌧ = 0.05 Top 5 influential Stocks Sartorius Aktiengesellschaft Frequency 12 Index TecDAX XING AG 8 TecDAX Surteco SE 7 SDAX Kabel Deutschland Holding AG 7 MDAX Biotest AG 6 MDAX Industry Provision of laboratory and process technologies and equipment Online business communication services Household Goods & Home Construction Cable-based telecommunication services Producing biological medications Comparison of Strategies 1.6 Cumulative Return 1.5 TEDAS - Tail Event Driven Asset allocation 1.4 1.3 1.2 1.1 1 0.9 March May July Time September Empirical Results 4-7 - b in each window, ⌧ = 0.05 b in application; ⌧ = 0.05 TEDAS - Tail Event Driven Asset allocation Selected Stocks Comparison of Strategies 1.6 1.5 Cumulative Return Figure 10: Different 1.4 1.3 1.2 1.1 1 0.9 March May July Time September Empirical Results 4-8 Return TEDAS approach:Mutual Funds results 5 5 5 4 4 4 3 3 3 2 2 2 1 1 1 0 2007 2009 2011 2013 2015 0 2007 2009 2011 2013 2015 0 2007 2009 2011 2013 2015 Figure 11: Strategies’ cumulative returns’ comparison: TEDAS Basic, TEDAS Naïve, TEDAS Hybrid, S&P500 TEDAS_strategies Comparison of Strategies 1.6 Cumulative Return 1.5 TEDAS - Tail Event Driven Asset allocation 1.4 1.3 1.2 1.1 1 0.9 March May July Time September Empirical Results 4-9 Return TEDAS approach:Mutual Funds results 5 5 5 4 4 4 3 3 3 2 2 2 1 1 1 0 2007 2009 2011 2013 2015 0 2007 2009 2011 2013 2015 0 2007 2009 2011 2013 2015 Figure 12: Strategies’ cumulative returns’ comparison (with transaction costs 1% of portfolio value): TEDAS Hybrid, 60/40, ERC, RR TEDAS_strategies Comparison of Strategies 1.6 Cumulative Return 1.5 TEDAS - Tail Event Driven Asset allocation 1.4 1.3 1.2 1.1 1 0.9 March May July Time September Empirical Results 4-10 Strategies’ performance: Mutual funds Strategy TEDAS Basic TEDAS Naïve TEDAS Hybrid RR ERC 60/40 S&P500 Cumulative return 421% 454% 433% 116% 129% 121% 113% Sharpe ratio 0.6393 0.6974 0.6740 0.0214 0.0487 0.0252 0.0132 Maximum drawdown 0.0855 0.0583 0.0276 0.4772 0.4899 0.3473 0.5037 Comparison of Strategies 1.6 Cumulative Return 1.5 TEDAS - Tail Event Driven Asset allocation 1.4 1.3 1.2 1.1 1 0.9 March May July Time September Empirical Results 4-11 TEDAS approach:Mutual Funds results τ = 0.05 τ = 0.15 8 5 4 6 3 4 2 2 0 1 1 2 3 4 5 0 1 2 τ = 0.25 3 4 5 4 5 τ = 0.35 10 8 8 6 6 4 4 2 2 0 1 2 3 4 5 0 1 2 3 Figure 13: Frequency of the number of selected variables for 4 different ⌧ Comparison of Strategies 1.6 Cumulative Return 1.5 TEDAS - Tail Event Driven Asset allocation 1.4 1.3 1.2 1.1 1 0.9 March May July Time September Empirical Results 4-12 TEDAS approach:Mutual Funds results τ = 0.05 τ = 0.15 5 8 4 6 3 4 2 2 1 0 0 0 200 400 600 0 τ = 0.25 200 400 600 τ = 0.35 15 8 6 10 4 5 2 0 0 0 200 400 600 0 200 400 600 Figure 14: The frequency of mutual funds Comparison of Strategies 1.6 Cumulative Return 1.5 TEDAS - Tail Event Driven Asset allocation 1.4 1.3 1.2 1.1 1 0.9 March May July Time September Empirical Results 4-13 Selected Mutual Funds Table 2: The selected Mutual Funds for ⌧ = 0.05 Top 5 influential Stocks Blackrock Eurofund Class I Pimco Funds Long Term United States Government Institutional Shares Prudential International Value Fund Class Z Artisan International Fund Investor Shares American Century 2OTH Century International Growth Investor Class Frequency 12 8 Market U.S. U.S. 4 U.S. 3 U.S. 1 U.S. Comparison of Strategies 1.6 Cumulative Return 1.5 TEDAS - Tail Event Driven Asset allocation 1.4 1.3 1.2 1.1 1 0.9 March May July Time September Choice of ⌧ -spine 5-1 How to choose optimal ⌧ -spine? Comparison of Strategies 1.6 Cumulative Return 1.5 TEDAS - Tail Event Driven Asset allocation 1.4 1.3 1.2 1.1 1 0.9 March May July Time September Choice of ⌧ -spine 5-2 Generation of different ⌧ -spines Tau = 0.5*Xn 0.5 0.45 0.4 0.35 Tau 0.3 0.25 0.2 0.15 0.1 0.05 0 0 0.2 0.4 0.6 0.8 1 X Figure 15: Generation of 10 sets of ⌧ -spines Comparison of Strategies 1.6 Cumulative Return 1.5 TEDAS - Tail Event Driven Asset allocation 1.4 1.3 1.2 1.1 1 0.9 March May July Time September Choice of ⌧ -spine 5-3 TEDAS Basic with different ⌧ -spines Figure 16: Cumulative return for TEDAS Basic with various ⌧ -spines Comparison of Strategies 1.6 Cumulative Return 1.5 TEDAS - Tail Event Driven Asset allocation 1.4 1.3 1.2 1.1 1 0.9 March May July Time September Choice of ⌧ -spine 5-4 TEDAS Basic with different ⌧ -spines Cumulative return 2 1.8 1.6 1.4 1.2 1 Jan14 May14 Sep14 Jan15 Time Figure 17: TEDAS Basic cumulative returns’ for ⌧ -spines: ⌧j=1,...,5 = (0, 0.002, 0.0233, 0.1311, 0.5), ⌧j=1,...,5 = (0.05, 0.15, 0.25, 0.35, 0.5), ⌧j=1,...,50 = (0.01, 0.02, 0.03 . . . 0.49, 0.5) Comparison of Strategies 1.6 Cumulative Return 1.5 TEDAS - Tail Event Driven Asset allocation 1.4 1.3 1.2 1.1 1 0.9 March May July Time September Choice of ⌧ -spine 5-5 What is the best ⌧ -spine? Monte Carlo simulations ⇢ Yi =q̂⌧i ⌧j=1,2,3 = (0.05, 0.15, 0.35), n = 100, Yt ⇠ ALD(µ, , ⌧ ); Details ⇢ Xi ⇠ N(0, ⌦), n = 100 for every ⌧ , p = 150, = ( 5, q2, 1, 3, 1, 0.5, 0, ..., 0), "i ⇠ N(0, 2 ); p ˆinit k0 log(n _ p)(log n)0.1/2 , ! ˆ j = 1/| ˆinit | ^ n; n = 0.25 k j ˆinit ; j ⇢ ⌦i,j = 0.5|i j| , = 0.1, 0.5, 1 (three levels of noise); Comparison of Strategies 1.6 Cumulative Return 1.5 TEDAS - Tail Event Driven Asset allocation 1.4 1.3 1.2 1.1 1 0.9 March May July Time September Choice of ⌧ -spine 5-6 What is the best ⌧ -spine? ⇢ for ˆinit estimator ˆ⌧, ˆ from the model (2) is used, where ˆ is chosen according to the BIC criterion ( ) n X log(n) b def BIC n ,⌧ = log n 1 · ⇢⌧ (Yi Xi> ˆ⌧ ) + · df( n ) 2n i=1 ⇢ Apply one of TEDAS modification with different ⌧ -spines ⇢ Choose that ⌧ -spine, which gives the highest wealth Wi = d X w j xi , ⌧ , j=1 Comparison of Strategies 1.6 Cumulative Return 1.5 TEDAS - Tail Event Driven Asset allocation 1.4 1.3 1.2 1.1 1 0.9 March May July Time September Conclusions 6-1 beginframe ⇢ TEDAS solves TLND challenge ⇢ TEDAS approach performs better than traditional benchmark strategies ⇢ TEDAS outperforms for I different regions (global and Germany), I various assets I alternative time periods (daily, weekly and monthly), I big data and small data ⇢ Results for 3 gestalts of TEDAS are robust TEDAS - Tail Event Driven Asset allocation Comparison of Strategies 1.6 1.5 Cumulative Return ⇢ Discussion: 1.4 1.3 1.2 1.1 1 0.9 March May July Time September I How to choose optimal ⌧ -spine? Conclusions TEDAS 6-3 - Tail Event Driven ASset Allocation: equity and mutual funds’ markets Wolfgang Karl Härdle Sergey Nasekin 1.6 Cumulative Return Alla Petukhina 1.8 1.4 1.2 1 0.8 Jan Feb Mar Apr May Jun Jul Aug Sep Oct Nov Dec Time Ladislaus von Bortkiewicz Chair of Statistics C.A.S.E. – Center for Applied Statistics and Economics Humboldt–Universität zu Berlin http://lvb.wiwi.hu-berlin.de http://case.hu-berlin.de Comparison of Strategies 1.6 Cumulative Return 1.5 TEDAS - Tail Event Driven Asset allocation 1.4 1.3 1.2 1.1 1 0.9 March May July Time September Technical details 7-1 Notation def ⇢ q̂⌧ = Fbn 1 (⌧ ), with def Fbn (Yt ) = Z n 1X fˆn (u) du = H n 1 Yt i=1 P def where fˆnR(Yt ) = (1/nh) ni=1 K {(Yt x H(x) = 1 K (u)du, K (·) = '(·); Silverman (1986) rule-of-thumb: h = 1.06sn ⇢ ˆ⌧, 1/5 ✓ Yt Yi h ◆ (2) , Yi )/h}, , s sample standard deviation of Y are the estimated non-zero ALQR coefficients n Back to "TEDAS Step 1" Comparison of Strategies 1.6 Cumulative Return 1.5 TEDAS - Tail Event Driven Asset allocation 1.4 1.3 1.2 1.1 1 0.9 March May July Time September Technical details 7-2 Lasso Shrinkage Linear model: Y = X + "; Y 2 Rn , X 2 Rn⇥p , i.i.d., independent of {Xi ; i = 1, ..., n} 2 Rp , {"i }ni=1 The optimization problem for the lasso estimator: ˆlasso = arg min f ( ) p 2R subject to g( ) (3) 0 where 1 (y X )> (y 2 g ( ) = t k k1 f( )= where t is the size constraint on k k1 X ) Back to "Tail Events" Comparison of Strategies 1.6 Cumulative Return 1.5 TEDAS - Tail Event Driven Asset allocation 1.4 1.3 1.2 1.1 1 0.9 March May July Time September Technical details 7-3 Lasso Duality If (1) is convex programming problem, then the Lagrangian is L( , ) = f ( ) g ( ). and the primal-dual relationship is minimize sup L( , ) 0 | {z } primal maximize inf L( , ) 0 | {z } dual Then the dual function L⇤ ( ) = inf L( , ) is 1 L⇤ ( ) = y > y 2 with (y 1 ˆ> > ˆ X X 2 X ˆ)> X ˆ/k ˆk1 = t (y X ˆ)> X ˆ k ˆk1 Back to "Tail Events" Comparison of Strategies 1.6 Cumulative Return 1.5 TEDAS - Tail Event Driven Asset allocation 1.4 1.3 1.2 1.1 1 0.9 March May July Time September Technical details 7-4 Paths of Lasso Coefficients figures/lassop.pdf Figure 18: Lasso shrinkage of coefficients in the hedge funds dataset example (6 covariates were chosen for illustration); each curve represents a coefficient as a function of the scaled parameter ŝ = t/k k1 ; the dashed line represents the model selected by the BIC information criterion (ŝ = 3.7) Back to "Tail Events" Comparison of Strategies 1.6 Cumulative Return 1.5 TEDAS - Tail Event Driven Asset allocation 1.4 1.3 1.2 1.1 1 0.9 March May July Time September Technical details 7-5 8 9 10 11 12 Example of Lasso Geometry βOLS 4 −4 −3 −2 −1 0 1 2 3 β2 5 6 7 ● −4 −3 −2 −1 0 1 2 3 4 5 6 7 8 9 10 11 12 β1 Figure 19: Contour plot of the residual sum of squares objective function centered at the OLS estimate bols = (6, 7) and the constraint region P | j| t MVAlassocontour.xpl Comparison of Strategies 1.6 Cumulative Return 1.5 TEDAS - Tail Event Driven Asset allocation 1.4 1.3 1.2 1.1 1 0.9 March May July Time September Technical details 7-6 Quantile Regression The loss ⇢⌧ (u) = u{⌧ def Fy |x1 (⌧ ) = I(u < 0)} gives the (conditional) quantiles q⌧ (x). Minimize ˆ⌧ = arg min p 2R n X ⇢⌧ (Yi Xi> ). i=1 Re-write: minimize (⇠,⇣)2R2n + n ⌧ 1> n ⇠ + (1 ⌧ )1> n ⇣|X + ⇠ with ⇠, ⇣ are vectors of "slack" variables ⇣=Y o Back to "Tail Events" Comparison of Strategies 1.6 Cumulative Return 1.5 TEDAS - Tail Event Driven Asset allocation 1.4 1.3 1.2 1.1 1 0.9 March May July Time September Technical details 7-7 Non-Positive (NP) Lasso-Penalized QR The lasso-penalized QR problem with an additional non-positivity constraint takes the following form: minimize (⇠,⇣,⌘, ˜)2R2n+p ⇥Rp + subject to ⌧ 1> n ⇠ + (1 > ⌧ )1> n ⇣ + 1n ⌘ ⇠ ⇣ = Y + X ˜, ⇠ 0, ⇣ 0, ˜, ⌘ ˜, ⌘ ˜ (4) 0, ˜ def = Back to "Tail Events" Comparison of Strategies 1.6 Cumulative Return 1.5 TEDAS - Tail Event Driven Asset allocation 1.4 1.3 1.2 1.1 1 0.9 March May July Time September Technical details 7-8 Solution Transform into matrix (Ip is p ⇥ p identity matrix; Ep⇥n = minimize c >x ✓ ◆ In 0 ): subject to Ax = b, Bx 0 where A = In In 0 X , b = Y, x = 0 1 ⌧ 1n B B (1 ⌧ )1n C B C, B = B c=B B @ A 1p @ 01p 0 Ep⇥n 0 0 0 0 Back to "Tail Events" ⇠ ⇣ ⌘ 0 Ep⇥n 0 0 0 0 0 Ip Ip 0 > 0 0 Ip Ip Ip , 1 C C C C A Comparison of Strategies 1.6 Cumulative Return 1.5 TEDAS - Tail Event Driven Asset allocation 1.4 1.3 1.2 1.1 1 0.9 March May July Time September Technical details 7-9 Solution - Continued The previous problem may be reformulated into standard form minimize c >x subject to Cx = d, x + s = u, x 0, s 0 and the dual problem is: maximize d >y subject to C > y u>w w + z = c, z 0, w 0 Back to "Tail Events" Comparison of Strategies 1.6 Cumulative Return 1.5 TEDAS - Tail Event Driven Asset allocation 1.4 1.3 1.2 1.1 1 0.9 March May July Time September Technical details 7-10 Solution - Continued The KKT conditions for this linear program are 8 9 Cx d > > > > > > > > x +s u < = > C y w +z c F (x, y , z, s, w ) = = 0, > > > > x z > > > > : ; s w with y 0, z variables. 0 dual slacks, s 0 primal slacks, w 0 dual This can be solved by a primal-dual path following algorithm based on the Newton method Back to "Tail Events" Comparison of Strategies 1.6 Cumulative Return 1.5 TEDAS - Tail Event Driven Asset allocation 1.4 1.3 1.2 1.1 1 0.9 March May July Time September Technical details 7-11 Adaptive Lasso Procedure Lasso estimates ˆ can be inconsistent (Zou, 2006) in some scenarios. Lasso soft-threshold function gives biased results figures/thresh-eps-converted-to.pdf Figure 20: Threshold functions for simple and adaptive Lasso Comparison of Strategies 1.6 Cumulative Return 1.5 TEDAS - Tail Event Driven Asset allocation 1.4 1.3 1.2 1.1 1 0.9 March May July Time September Technical details 7-12 Adaptive Lasso Procedure The adaptive Lasso (Zou, 2006) yields a sparser solution and is less biased. L1 - penalty replaced by a re-weighted version; ! ˆ j = 1/| ˆjinit | , = 1, ˆinit is from (3) The adaptive lasso estimates are given by: ˆadapt = arg min p 2R n X (Yi i=1 Xi> )2 + kˆ ! > k1 init adapt (Bühlmann, van de Geer, 2011): ˆj = 0, then ˆj =0 Back to "Tail Events" Comparison of Strategies 1.6 Cumulative Return 1.5 TEDAS - Tail Event Driven Asset allocation 1.4 1.3 1.2 1.1 1 0.9 March May July Time September Technical details 7-13 Simple and Adaptive Lasso Penalized QR Simple lasso-penalized QR optimization problem is: ˆ⌧, = arg min p 2R n X ⇢⌧ (Yi i=1 Xi> ) + k k1 (5) Adaptive lasso-penalized QR model uses the re-weighted penalty: ˆadapt = arg min ⌧, p 2R n X ⇢⌧ (Yi i=1 Xi> ) + kˆ ! > k1 (6) Adaptive lasso-penalized QR procedure can ensure oracle properties for the estimator Details Back to "Tail Events" Comparison of Strategies 1.6 Cumulative Return 1.5 TEDAS - Tail Event Driven Asset allocation 1.4 1.3 1.2 1.1 1 0.9 March May July Time September Technical details 7-14 Algorithm for Adaptive Lasso Penalized QR The optimization for the adaptive lasso quantile regression can be re-formulated as a lasso problem: ⇢ the covariates are rescaled: X̃ = (X1 ⇢ the lasso problem (5) is solved: ˆinit , . . . , Xp 1 ˆpinit ); n ˆ˜ = arg min X ⇢ (Y ⌧ i ⌧, p 2R i=1 X̃i> ) + k k1 adapt ˜⌧, ⇢ the coefficients are re-weighted as ˆ⌧, = ˆ ˆinit Back to "Tail Events" Comparison of Strategies 1.6 Cumulative Return 1.5 TEDAS - Tail Event Driven Asset allocation 1.4 1.3 1.2 1.1 1 0.9 March May July Time September Technical details 7-15 Oracle Properties of an Estimator An estimator has oracle properties if (Zheng et al., 2013): ⇢ it selects the correct model with probability converging to 1; ⇢ the model estimates are consistent with an appropriate convergence rate; ⇢ estimates are asymptotically normal with the same asymptotic variance as that knowing the true model Back to "Simple and Adaptive Lasso Penalized QR" Comparison of Strategies 1.6 Cumulative Return 1.5 TEDAS - Tail Event Driven Asset allocation 1.4 1.3 1.2 1.1 1 0.9 March May July Time September Technical details 7-16 Oracle Properties for Adaptive Lasso QR In the linear model, let Y = X + " = X 1 1 + X 2 2 + ", where X = (X 1 , X 2 ), X 1 2 Rn⇥q , X 2 2 Rn⇥(p q) ; q1 are true nonzero coefficients, p2 q = 0 are noise coefficients; q = k k0 . p p Also assume that q/ n ! 0 and /{ q log(n _ p)} ! 1 and certain regularity conditions are satisfied Details Back Comparison of Strategies 1.6 Cumulative Return 1.5 TEDAS - Tail Event Driven Asset allocation 1.4 1.3 1.2 1.1 1 0.9 March May July Time September Technical details 7-17 Oracle Properties for Adaptive Lasso QR Then the adaptive L1 QR estimator has the oracle properties (Zheng et al., 2013): 1. Variable selection consistency: P( 2 = 0) 1 6 exp ⇢ log(n _ p) 4 ˆk = Op ( 2. Estimation consistency: k def . p q/n) 3. Asymptotic normality: uq2 = ↵T ⌃11 ↵, 8↵ 2 Rq , k↵k < 1, ⇢ (1 ⌧ )⌧ 1/2 1 T 1 1 L ˆ n uq ↵ ( ) ! N 0, 2 ⇤ f ( ) where ⇤ is the ⌧ th quantile and f is the pdf of " Back Comparison of Strategies 1.6 Cumulative Return 1.5 TEDAS - Tail Event Driven Asset allocation 1.4 1.3 1.2 1.1 1 0.9 March May July Time September Technical details 7-18 Risk-Return Asset Allocation Log returns Xt 2 Rp : minp wt 2R def 2 P,t (wt ) = wt> ⌃t wt s.t. µP,t (wt ) = rT , (7) wt> 1p = 1, wi,t def where rT "target" return, ⌃t = modeled with a GARCH model 0 t 1 {(Xt Details µ)(Xt µ)> }, ⌃t is Back to "Benchmark Strategies" Return to "TEDAS Gestalten" Comparison of Strategies 1.6 Cumulative Return 1.5 TEDAS - Tail Event Driven Asset allocation 1.4 1.3 1.2 1.1 1 0.9 March May July Time September Technical details 7-19 Dynamic Conditional Correlations Model Assume: rt |Ft 1 ⇠ N(0, Dt Rt Dt ), "t = Dt 1 rt , Dt2 = diag(!i ) + diag(↵i ) Qt = S (11> A Rt = {diag(Qt )} 1 rt > 1 rt 1 B) + A {Pt Qt {diag(Qt )} where rt 2 Rp , Dt = diag ( def it ) + diag( i ) 1 Dt2 1 , > 1 "t 1 "t 1 P t 1 } +B Qt 1, 2 Rp⇥p , "t 2 Rp standardized def returns with "it = rit it 1 , 1 vector of ones; Pt 1 = {diag(Qt )}1/2 , !i , ↵i , i , A, B coefficients, Hadamard (elementwise) product Return to "TEDAS Gestalten" Comparison of Strategies 1.6 Cumulative Return 1.5 TEDAS - Tail Event Driven Asset allocation 1.4 1.3 1.2 1.1 1 0.9 March May July Time September Technical details 7-20 The DCC Model - Continued ⇢ correlation targeting: S = (1/T ) PT > t=1 "t "t ⇢ Q0 = " 0 " > 0 positive definite, each subsequent Qt also positive definite ⇢ consistent but inefficient estimates: the log-likelihood function o 1 Xn 1 n log(2⇡) + 2 log |Dt | + log |Rt | + "> R " t , t t 2 T L(✓, ) = t=1 where ✓ parameters in D and parameters in R Back additional correlation Comparison of Strategies 1.6 Cumulative Return 1.5 TEDAS - Tail Event Driven Asset allocation 1.4 1.3 1.2 1.1 1 0.9 March May July Time September Technical details 7-21 The DCC Model - Continued Re-write: L(✓, ) = LV (✓) + LC (✓, ), with volatility part LV (✓) and correlation part LC (✓, ), o 1 Xn n log(2⇡) + log |Dt |2 + rt> Dt 2 rt 2 t=1 T d ⇢ r2 1 XX log(2⇡) + log( it2 ) + it2 , 2 it T LV (✓) = = t=1 i=1 1 Xn 1 log |Rt | + "> t R t "t 2 T LC (✓, ) = t=1 o "> " t t . Back Comparison of Strategies 1.6 Cumulative Return 1.5 TEDAS - Tail Event Driven Asset allocation 1.4 1.3 1.2 1.1 1 0.9 March May July Time September Technical details 7-22 Cornish-Fisher VaR Optimization Log returns Xt 2 Rp : minimize w 2Rd Wt { q↵ (wt ) · p (wt )} subject to wt> µ = µp , wt> 1 = 1, wt,i 0 Q def here Wt = W0 · tj=11 wt> j (1 + Xt j ), w̃ , W0 initial wealth, def 2 p (w ) = wt> ⌃t wt , Sp (w ) Kp (w ) Sp (w )2 +(z↵3 3z↵ ) (2z↵3 5z↵ ) , 6 24 36 here Sp (w ) skewness, Kp (w ) kurtosis, z↵ is N(0, 1) ↵-quantile If Sp (w ), Kp (w ) zero, then obtain Markowitz allocation def q↵ (w ) = z↵ +(z↵2 1) Return to "TEDAS Gestalten" Comparison of Strategies 1.6 Cumulative Return 1.5 TEDAS - Tail Event Driven Asset allocation 1.4 1.3 1.2 1.1 1 0.9 March May July Time September Technical details 7-23 Risk Parity (Equal risk contribution) Let (w ) = p w > ⌃w . Euler decomposition: def (w ) = n X i (w ) = i=1 n X i=1 wi (w ) @wi (w ) (w ) where @w is the marginal risk contribution and i (w ) = wi @w i i the risk contribution of i-th asset. The idea of ERC strategy is to find risk balanced porfolio, such that: i (w ) = j (w ) i.e. the risk contribution is the same for all assets of the portfolio Return to "Benchmark Strategies" Comparison of Strategies 1.6 Cumulative Return 1.5 TEDAS - Tail Event Driven Asset allocation 1.4 1.3 1.2 1.1 1 0.9 March May July Time September Technical details 7-24 60/40 allocation strategy 60/40 portfolio allocation strategy implies the investing of 60% of the portfolio value in stocks (often via a broad index such as S&P500) and 40% in government or other high-quality bonds, with regular rebalancing to keep proportions steady. Return to "Benchmark Strategies" Comparison of Strategies 1.6 Cumulative Return 1.5 TEDAS - Tail Event Driven Asset allocation 1.4 1.3 1.2 1.1 1 0.9 March May July Time September Technical details 7-25 Portfolio Skewness and Kurtosis Skewness SP and excess kurtosis KP are given by moment expressions 1 SP (w ) = 3 (m3 3m2 m1 + 2m13 ) (w ) P 1 KP (w ) = 4 (m4 4m3 m1 + 6m2 m12 + 3m14 ) 3 (w ) P where portfolio non-central moments also depend on w : def m1 = µP (w ) = w > µ m2 = m3 = 2 2 P + m1 d X d X d X wi wj wk Sijk i=1 j=1 k=1 Comparison of Strategies 1.6 Cumulative Return 1.5 TEDAS - Tail Event Driven Asset allocation 1.4 1.3 1.2 1.1 1 0.9 March May July Time September Technical details 7-26 Portfolio Skewness and Kurtosis - Continued m4 = d X d X d X d X wi wj wk wl Kijkl , i=1 j=1 k=1 l=1 2 P (w ) = E(ri ⇥ rj w > ⌃w where and Sijk = E(ri ⇥ rj ⇥ rk ), Kijkl = ⇥ rk ⇥ rl ) can be computed via sample averages from returns data. Sijk , Kijkl determine the d-dimensional portfolio co-skewness and co-kurtosis tensors def S = {Sijk }i,j,k=1,...,d 2 Rd⇥d⇥d def K = {Kijkl }i,j,k,l=1,...,d 2 Rd⇥d⇥d⇥d . Back Comparison of Strategies 1.6 Cumulative Return 1.5 TEDAS - Tail Event Driven Asset allocation 1.4 1.3 1.2 1.1 1 0.9 March May July Time September Technical details 7-27 Regularity Conditions for Adaptive Lasso QR A1 Sampling and smoothness: 8x in the support of Xi , 8y 2 R, 0 fYi |Xi (y |x), f 2 C k (R), |fYi |Xi (y |x)| < f , |fYi |Xi (y |x)| < f 0 ; 9f , such that fYi |Xi (x > ⌧ |x) > f > 0 A2 Restricted identifiability and nonlinearity: let 2 Rp , T ⇢ {0, 1, ..., p}, T such that Tj = j if j 2 T , Tj = 0 if j 62 T ; T = {0, 1, ..., s}, T ( , m) ⇢ {0, 1, ..., p}\T , then 9m 0, c 0 such that inf T E(X > i Xi ) 2A, 6=0 k T [T ( ,m) def k2 > 0, where A = { 2 Rp : k T c k1 3f 3/2 8f 0 ck inf 2A, E[|XiT |2 ]3/2 > 0, 6=0 E[|X T |3 ] i T k1 , k T c k0 n} Back Comparison of Strategies 1.6 Cumulative Return 1.5 TEDAS - Tail Event Driven Asset allocation 1.4 1.3 1.2 1.1 1 0.9 March May July Time September Technical details 7-28 Regularity Conditions - Continued A3 Growth rate of covariates: q 3 {log(n _ p)}2+⌘ ! 0, ⌘ > 0 n A4 Moments of covariates: Cramér condition E[|xij |k ] 0.5Cm M k for some constants Cm , M, 8k 2 k! 2, j = 1, ..., p A5 Well-separated regression coefficients: 9b0 > 0, such that 8j q, | ˆj | > b0 Comparison of Strategies 1.6 Cumulative Return 1.5 TEDAS - Tail Event Driven Asset allocation 1.4 1.3 1.2 1.1 1 0.9 March May July Time September Technical details 7-29 Proof of the CF-CVaR Expansion 1 Define the Cornish-Fisher expansion: q1 ↵ def = z1 2 ↵ + (z1 ↵ def def 1)s + (z13 ↵ 3z1 ↵ )k (2z13 ↵ 5z1 ↵ )s 2 , where s = S/6, k = K /24, S and K are skewness and excess kurtosis, respectively; z1 ↵ def 1 (1 = ↵). Re-write: q1 where a0 = ↵ = a0 + a1 z1 s, a1 = 1 ↵ + a2 z12 ↵ + a3 z13 3k + 5s 2 , a2 = s, a3 = k (8) ↵, 2s 2 Comparison of Strategies 1.6 Cumulative Return 1.5 TEDAS - Tail Event Driven Asset allocation 1.4 1.3 1.2 1.1 1 0.9 March May July Time September Technical details 7-30 Proof of the CF-CVaR Expansion 2 Define the conditional Value-at-Risk (CVaR) or expected shortfall (ES): Z 1 1 def CVaR↵ = VaRq d q, 1 ↵ ↵ def where VaRq = 1 (↵) p T Observe: CVaR↵ = = = 1 1 1 1 p p ↵ T ↵ T ↵ Z 1 ↵ Z 1 Z 1 1 (↵) p Tdq (9) (↵)d q (10) u'(u)d u, (11) ↵ z↵ 1 where (11) follows from the change of variable: u = zq = 1 (q) Comparison of Strategies 1.6 Cumulative Return 1.5 TEDAS - Tail Event Driven Asset allocation 1.4 1.3 1.2 1.1 1 0.9 March May July Time September Technical details 7-31 Proof of the CF-CVaR Expansion 3 Substitute (8) into (11): CVaR↵ = p 1 T ↵ Z z↵ a0 + a1 z + a2 z 2 + a3 z 3 '(z)d z 1 = a0 A0 + a1 A1 + a2 A2 + a3 A3 , A0 = A1 = A2 = A3 = p 1 p T ↵ Z p z↵ '(z)d z = 1 T, p z↵ T '(z↵ ), ↵ p Z z ✓ ◆ ↵ p '(z↵ )z↵ T z 2 '(z)d z = T +1 , 1 ↵ 1 ↵ 1 p Z z p ↵ T T 2 z 3 '(z)d z = (z↵ + 2)'(z↵ ). 1 ↵ 1 ↵ 1 1 T ↵ Z z'(z)d z = 1 1 Collecting terms and simplifying gives the desired result. Return to "Conditional VaR Optimization" Comparison of Strategies 1.6 Cumulative Return 1.5 TEDAS - Tail Event Driven Asset allocation 1.4 1.3 1.2 1.1 1 0.9 March May July Time September Technical details 7-32 density 0.00 0.10 0.20 0.30 Asymetric Laplace Distribution −4 −2 0 2 4 X Figure 21: Standart ALD: ⌧ = 0.3,⌧ = 0.5,⌧ = 0.7, ⌧ = 0.1 Comparison of Strategies 1.6 Cumulative Return 1.5 TEDAS - Tail Event Driven Asset allocation 1.4 1.3 1.2 1.1 1 0.9 March May July Time September Technical details 7-33 Quantile regression using ALD ⇢ Yu & Moyeed(2001) Yi ⇠ ALD(µ, , ⌧ ), if its pdf is given by n f (y |µ, , ⌧ ) = ⌧ (1 ⌧ ) exp ⇢⌧ (y µ) o where µ is location, - scale and ⌧ -skewness parameters, and loss function ⇢⌧ (u) = u{⌧ I(u < 0)} ⇢ Sanches et. al (2013) yi = xi> Re-write: here ✓ = ⌧ + "i , i = 1, . . . , n Yi |Ui = ui ⇠ N(xi ⌧ + ✓ui , p⌧2 ui ) Ui ⇠ Exp( ), i = 1, . . . , n 1 2⌧ ⌧ (1 ⌧ ) and p⌧2 = 2 ⌧ (1 ⌧ ) Back to "Choice of ⌧ -spine" Comparison of Strategies 1.6 Cumulative Return 1.5 TEDAS - Tail Event Driven Asset allocation 1.4 1.3 1.2 1.1 1 0.9 March May July Time September References 8-1 References W. K. Härdle, S. Nasekin, D. K. C. Lee , K. F. Phoon TEDAS - Tail Event Driven Asset Allocation SFB 649 Discussion Paper 2014-032: Jun, 2014 L. B. Sanchez V. H. Lachos, F. V. Labra Likelihood Based Inference for Quantile Regression Using the Asymmetric Laplace Distribution Discussion Paper, Universidade Estadual de Campinas: Sep, 2013 Comparison of Strategies 1.6 Cumulative Return 1.5 TEDAS - Tail Event Driven Asset allocation 1.4 1.3 1.2 1.1 1 0.9 March May July Time September