Portfolio Optimization and Risk Prediction with Non ...
Transcript of Portfolio Optimization and Risk Prediction with Non ...
1/52
Portfolio Optimization and Risk Prediction withNon-Gaussian COMFORT Models
Marc S. Paolellab,c Pawe l Polaka,c
aColumbia University bSwiss Finance Institute c University of Zurich
September 16, 2016
Marc S. Paolella & Pawe l Polak COMFORT: Portfolio Optimization, Risk Management
2/52
• VaR Performance • ES Performance • MLE > Ad-Hoc 2-Step • JRSS-A 2016
• MGHyp-CCC-GARCH
Framework • EM Algo. Estimation • Common Market
Factor • Option Pricing • JoE 2015
• Analytic ES • DCC Modeling • Portfolio Optimization • Risk Tracking Method • Sharpe Ratios > 1 • Submitted
Risk Modeling
Primary Model COMFORT
PSARM
• Leverage Investing • Risk Trackers • Tactical Asset
Allocation • Large Swiss Banks
and Pension Funds
Generalizing the MGHyp: • Multi-Factors for Full
Tail Heterogeneity • Lévy-Stable /
Tempered Stable
• FREE-COMFORT • Markov-Switching CCC • PCA Factor Structure • High-Freq. Strategy • Black-Litterman • Machine Learning
COMFORT Road Map Highway
Academic
Industry Consulting
Large-Scale Port. Opt.
30 min talk!!!
PSARM - Portfolio Selection with Active Risk Monitoring.
Marc S. Paolella & Pawe l Polak COMFORT: Portfolio Optimization, Risk Management
3/52
Normal Model vs. Stylized Facts of Financial Returns
Normal World vs. Real World
Returns DistributionMultivariate Normal (Gaussian) → Multivariate Fat-Tailed
Risk MeasureVolatility → VaR and Expected Shortfall
Measure of DependenceCorrelations → Correlations & Tail Dependence
Market DynamicsConstant Volatility & Correlations → Time Varying Volatility & Correlation
SymmetryAll Returns Are Symmetric → Asset Specific Skewness
... the biggest problems we now have with the whole evaluation of risk is the fat-tail problem, which is really creatingvery large conceptual difficulties. Because as we all know, the assumption of normality enables us to drop off a hugeamount of complexity of our equations very much to the right of the equal sign. Because once you start putting innon-normality assumptions, which unfortunately is what characterizes the real world, then the issues become extremelydifficult.
Greenspan, A. (1997)
Our models capture the stylized facts of financial returns from the RealWorld.
Marc S. Paolella & Pawe l Polak COMFORT: Portfolio Optimization, Risk Management
4/52
Pitfalls of Copula Models
Copula Models vs. COMFORT Models
AssumptionsAll Taken From The Real World → All Taken From The Real World
Model EstimationFast → Fast
Two Step Fixed Point Algorithm(I) Margins; (II) Dependency
Portfolio DistributionNo Closed Form - Needs Simulations → Closed Form
(slow) (fast)
Both methods allow for time-varying volatilities and non-ellipticity, and also are fast toestimate.
Simulation from the copula is required to generate the portfolio predictive density, and astail risk is often the concern, a very large number of replications will be required to achieveadequate accuracy, thus rendering the method, particularly for high dimensions, too slow forasset allocation purposes.
Our models maintain the flexibility of Copula models (real worldassumptions, estimation speed), but deliver a closed form expression for
the portfolio distribution.
Marc S. Paolella & Pawe l Polak COMFORT: Portfolio Optimization, Risk Management
5/52
Modern Portfolio Management
Markowitz Portfolio vs. COMFORT Portfolio
AssumptionsAll Taken From The Normal World → All Taken From The Real World
ErrorsHuge Modeling Error → Small Modeling Error
Portfolio DistributionClosed Form - Univariate Normal → Closed Form
Portfolio OptimizationFast → Fast
Quadratic Programming Problem Convex Optimization Problem
Our portfolio methods maintain the benefits of the iid MultivariateNormal framework (estimation speed, tractability of the portfoliodistribution), but have much smaller modeling error, and deliver a
superior forecast.
Marc S. Paolella & Pawe l Polak COMFORT: Portfolio Optimization, Risk Management
6/52
Table of Contents:
Multivariate Normal Mean-Variance Mixture Dist. and MGHyp case.
MGHyp in Finance.
New Class of COMFORT Models.
The Model and The Dynamics.
Fast VaR and ES Formulae, and Portfolio Optimization.
VaR backtesting, and Portfolio Performance.
Risk Fear Portfolio Strategy.
Equities Example.
Equities + Fixed Income Example.
COMFORT + Machine Learning for Improved PortfolioPerformance.
Marc S. Paolella & Pawe l Polak COMFORT: Portfolio Optimization, Risk Management
7/52
Multivariate Normal Mean-Variance Mixture Distribution
The random vector Y is said to have a multivariate normalmean-variance mixture distribution (MNMVM) if
Y = m (G ) + H1/2√GZ,
where
Z ∼ N (0, IK );
G ≥ 0 is a non-negative, univariate random variable which isindependent of Z;
H is a K × K symmetric and positive definite matrix; and
m : [ 0,∞ )→ Rd is a measurable function.
The name MNMVM comes from the fact that:
Y | (G = g) ∼ N (m (g) , gH) .
The multivariate generalized hyperbolic (MGHyp) distribution is aspecial case of MNMVM with:
m (G ) = µ+ γG and G ∼ GIG (λ, χ, ψ) .
Marc S. Paolella & Pawe l Polak COMFORT: Portfolio Optimization, Risk Management
8/52
MGHyp in Finance
Y = µ+ γG + H1/2√GZ, and G ∼ GIG (λ, χ, ψ) .
introduced by Barndorff-Nielsen (1977);
an inf. div. dist. used to construct Levy process (e.g.Ornstein-Uhlenbeck model driven by Levy & superpositions of them);
tractable portfolio dist.: w′Y ∼ GHyp (w′µ,w′γ,w′Hw, λ, χ, ψ);
used in: (i) portfolio, (ii) risk management, and (iii) option pricing;
special and limiting cases include:
multivariate normalmultivariate Laplacemultivariate normal inverse Gaussian (NIG)multivariate t-distributionvariance-gamma, Madan and Seneta (1990), Seneta (2004);
the tail behavior of the GHyp distribution spans a range fromexponential, to power tail.
Marc S. Paolella & Pawe l Polak COMFORT: Portfolio Optimization, Risk Management
9/52
New Class of MGHyp-CC and hybrid GARCH-SV Models
unconditional models conditional models
Yti.i.d.∼ N (µ,H) → univariate:
(t-GARCH) Yt | Ft−1 ∼t (µt , σt , ν)Yt | Ft−1 ∼GHyp (µ, γ, σt , λ, χ, ψ)
↓ multivariate (Gaussian):Yt | Ft−1 ∼ N (µt ,Ht) ; µt ,Ht ∈ Ft−1
BEKK, CCC, VC, DCC, ADCC, RSDC↓
Yti.i.d.∼ MGHyp (µ,γ,H, λ, χ, ψ) → MGHyp(µ,γ,Ht , λ, χ, ψ)-CC models
Ht = StΓtSt ; St ∈ Ft−1
Γt from CCC, VC, DCC, cDCC, RSDC,. . .
Notation: Ft = σ (εs : s ≤ t); ξ ∈ Ft−1 means that ξ is measurable with respect to Ft−1.
Marc S. Paolella & Pawe l Polak COMFORT: Portfolio Optimization, Risk Management
10/52
One Interesting Feature: The Common Market Factor G
-20
-10
0
10
02/01/01 07/10/03 06/07/06 06/04/09 30/12/11
MR
K
10
20
30
02/01/01 07/10/03 06/07/06 06/04/09 30/12/11
Gt
2
4
6
02/01/01 07/10/03 06/07/06 06/04/09 30/12/11
S t
2
4
6
02/01/01 07/10/03 06/07/06 06/04/09 30/12/11
vol t|t
-1
Marc S. Paolella & Pawe l Polak COMFORT: Portfolio Optimization, Risk Management
11/52
Financial Applications: Risk Measures and Portfolio
The MGHyp class, including all the limiting cases, is closed under linearoperations, so, the conditional distribution of the portfolio Pt = w′Yt isdirectly available and it is given by
Pt | Φt−1 ∼ GHyp (w′µ,w′γ,w′Htw, λ, χ, ψ) .
Given the density of the portfolio, VaR and ES can be numericallycomputed.
Standard methods to compute VaR and ES are time consuming, here wepropose faster solutions.
From the predictive portfolio distribution, one can do mean–variance ormean–ES portfolio optimization.
Marc S. Paolella & Pawe l Polak COMFORT: Portfolio Optimization, Risk Management
12/52
Fast VaR Formula for Risk Prediction
The corresponding portfolio cdf can be written as
Pr (Pt ≤ r) =
∫ ∞0
Φ
(r − µw − γwg
σt,w
)fGt (g ;λ, χ, ψ)dg ,
where Φ is the cdf of standard normal random variable.
Use of this representation reduces the computation time required for thenumerical inversion in the computation of the VaR values by a factor offour.
Marc S. Paolella & Pawe l Polak COMFORT: Portfolio Optimization, Risk Management
13/52
Fast ES Formula for Risk Prediction
Proposition
If P ∼ GHyp (µ, γ, h, λ, χ, ψ), then
ESα (P) = −µ− γ
αE [G ]FP∗ (−VaRα (P)) +
h
α
C√2π,
where FP∗ is the cdf of P∗ ∼ GHyp (µ, γ, h, λ+ 1, χ, ψ); and theconstant C is given by
C =χ−λ (√χψ)λ2Kλ
(√χψ) 2K
λ
(√χψ
)χ−λ
(√χψ
)λ exp
(−
(VaRα (P) + µ) γ
h2
),
with λ = λ+ 1/2, χ = χ+ (VaRα(P)+µ)2
h2 , and ψ = ψ + γ2
h2 .
Fast and accurate computation of ES for GHyp portfolio, e.g., in theelliptical case this is 100 times faster than standard ES computation!
Marc S. Paolella & Pawe l Polak COMFORT: Portfolio Optimization, Risk Management
14/52
Portfolio Optimization
The minimum-variance portfolio and the mean-variance portfolio are thesolutions of
minw∈W
σ2t,w, (1)
where
W ={
w ∈ RK :∑K
k=1 wk = 1, wk ≥ 0, k = 1, . . . ,K}
, and,
in case of the classical Markowitz (1952) mean-variance portfolio,µw + γwE [Gt | Φt−1] ≥ p, where p is a constant for the targetminimum return of the portfolio,
the minimum-ES (min-ES) portfolio and mean-ES portfolio replacethe conditional variance σ2
t,w by the conditional ES in (1),
the computational efficiency of min-ES portfolio optimization isgreatly increased by using the results from Rockafeller and Uryasev(2000).
Marc S. Paolella & Pawe l Polak COMFORT: Portfolio Optimization, Risk Management
15/52
mean-ES and min-ES Portfolio Optimization
Rockafeller and Uryasev (2000) introduce an auxiliary function
Fα (x ,w) = x +1
α
∫ ∞−∞
[p + x ]− fPt (p | µw, γw, σt,w, λ, χ, ψ) dp,
where [z ]− = −z if z ≤ 0, and [z ]− = 0 if z > 0, and
min(x,w)∈R×W
Fα (x ,w) = minw∈W
ESt|t−1α (Pt) , (2)
where the pair (x∗,w∗) achieves the first minimum if and only if w∗
achieves the second minimum and x∗ = VaRt|t−1α
(w∗′Yt
).
The first minimization in (2) is more efficient and provides globaloptimum because :
it completely avoids the calculation of the ES (and VaR) of thecandidate portfolios,
it is a convex programming problem.
Marc S. Paolella & Pawe l Polak COMFORT: Portfolio Optimization, Risk Management
16/52
Empirics
Data: 2, 767 daily vector returns of K = 30 components of the DowJones Industrial Index (DJ-30) from January 2nd, 2001, toDecember 30th, 2011 (from Wharton Data Base).
The returns for each asset are computed asyk,t = 100 log (pk,t/pk,t−1), where pk,t is the price of a share ofasset k at time t.
Models compared:
Bollerslev (1990) MN-CCC; Tse and Tsui (2002) MN-VC; Engle(2002) MN-DCC; Aielli (2011) MN-cDCC;COMFORT models, non-elliptical (estimated via EM):
MALap, MNIG, MAt;CCC, VC, DCC, cDCC;GARCH(1, 1), GJR-GARCH(1, 1);
some elliptical models (estimated in 3-steps and via EM):
MLap, MNIG, Mt
Marc S. Paolella & Pawe l Polak COMFORT: Portfolio Optimization, Risk Management
17/52
Tails of Q-Q Plots for Residuals from MN-CCC vs.MN-cDCC GARCH(1,1)
−4 −3 −2 −1−12−10
−8−6−4−2
MN−CCC GARCH (MSFT)
1 2 3 4
2
4
6
MN−CCC GARCH (MSFT)
−4 −3 −2 −1−12−10
−8−6−4−2
MN−cDCC GARCH (MSFT)
1 2 3 4
2
4
6
MN−cDCC GARCH (MSFT)
−4 −3 −2 −1
−8
−6
−4
−2
MN−CCC GARCH (KO)
1 2 3 41
2
3
4
5MN−CCC GARCH (KO)
−4 −3 −2 −1
−8
−6
−4
−2
MN−cDCC GARCH (KO)
1 2 3 41
2
3
4
5MN−cDCC GARCH (KO)
−4 −3 −2 −1
−6
−4
−2
MN−CCC GARCH (DD)
1 2 3 4
2
4
6
MN−CCC GARCH (DD)
−4 −3 −2 −1
−6
−4
−2
MN−cDCC GARCH (DD)
1 2 3 4
2
4
6
MN−cDCC GARCH (DD)
−4 −3 −2 −1
−5
−4
−3
−2
−1MN−CCC GARCH (XOM)
1 2 3 41
2
3
4
5MN−CCC GARCH (XOM)
−4 −3 −2 −1
−5
−4
−3
−2
−1MN−cDCC GARCH (XOM)
1 2 3 41
2
3
4
5MN−cDCC GARCH (XOM)
H−1/2t
(Yt − µ
)Marc S. Paolella & Pawe l Polak COMFORT: Portfolio Optimization, Risk Management
18/52
Tails of Q-Q Plots for Residuals from MN-cDCC vs.MNIG-cDCC GARCH(1,1)-SV
−4 −3 −2 −1−12−10
−8−6−4−2
MN−cDCC GARCH (MSFT)
1 2 3 4
2
4
6
MN−cDCC GARCH (MSFT)
−4 −3 −2 −1−12−10
−8−6−4−2
MNIG−cDCC GARCH−SV (MSFT)
1 2 3 4
2
4
6
MNIG−cDCC GARCH−SV (MSFT)
−4 −3 −2 −1
−8
−6
−4
−2
MN−cDCC GARCH (KO)
1 2 3 41
2
3
4
5MN−cDCC GARCH (KO)
−4 −3 −2 −1
−8
−6
−4
−2
MNIG−cDCC GARCH−SV (KO)
1 2 3 41
2
3
4
5MNIG−cDCC GARCH−SV (KO)
−4 −3 −2 −1
−6
−4
−2
MN−cDCC GARCH (DD)
1 2 3 4
2
4
6
MN−cDCC GARCH (DD)
−4 −3 −2 −1
−6
−4
−2
MNIG−cDCC GARCH−SV (DD)
1 2 3 4
2
4
6
MNIG−cDCC GARCH−SV (DD)
−4 −3 −2 −1
−5
−4
−3
−2
−1MN−cDCC GARCH (XOM)
1 2 3 41
2
3
4
5MN−cDCC GARCH (XOM)
−4 −3 −2 −1
−5
−4
−3
−2
−1MNIG−cDCC GARCH−SV (XOM)
1 2 3 41
2
3
4
5MNIG−cDCC GARCH−SV (XOM)
H−1/2t
(Yt − µ
)and G
−1/2t H
−1/2t
(Yt − µ − γGt
)Marc S. Paolella & Pawe l Polak COMFORT: Portfolio Optimization, Risk Management
19/52
Correlation Dynamics
0.2
0.4
02/01/01 07/10/03 06/07/06 06/04/09 30/12/11
Corrt|t−1
(KO,CAT)
MN−cDCC GARCH(1,1) MNIG−cDCC hybrid GARCH(1,1)−SV MAt−DCC GARCH(1,1) MNIG−cDCC GARCH(1,1)
0.2
0.3
0.4
02/01/01 07/10/03 06/07/06 06/04/09 30/12/11
Corrt|t−1
(KO,CSCO)
0.55
0.6
0.65
02/01/01 07/10/03 06/07/06 06/04/09 30/12/11
Corrt|t−1
(GE,AXP)
0.4
0.5
0.6
02/01/01 07/10/03 06/07/06 06/04/09 30/12/11
Corrt|t−1
(IBM,CSCO)
0.4
0.5
0.6
02/01/01 07/10/03 06/07/06 06/04/09 30/12/11
Corrt|t−1
(IBM,INTC)
0.2
0.3
0.4
02/01/01 07/10/03 06/07/06 06/04/09 30/12/11
Corrt|t−1
(CAT,MRK)
0.65
0.7
0.75
02/01/01 07/10/03 06/07/06 06/04/09 30/12/11
Corrt|t−1
(BAC,C)
0.5
0.6
0.7
02/01/01 07/10/03 06/07/06 06/04/09 30/12/11
Corrt|t−1
(VZ,T)
KO - Coca-Cola; CSCO - Cisco; CAT - Caterpillar; IBM - Intl. Business Machines; INTC - Intel; AXP - American Express; GE - General
Electric; BAC - Bank of America; MRK - Merck & Co; C - CitygroupMarc S. Paolella & Pawe l Polak COMFORT: Portfolio Optimization, Risk Management
20/52
VaR Backtesting - Forecast Failure Rates
M VaR 1% VaR 5% VaR 10%
MNIG-CCC GARCH(1, 1)-SV 0.0096 0.0526 0.0905MNIG-CCC GJR-GARCH(1, 1)-SV 0.0084 0.0469 0.0837MNIG-CCC GJR-GARCH(1, 1) 0.0118 0.0481 0.0888MAt-CCC GJR-GARCH(1, 1) 0.0118 0.0605 0.0984MNIG-CCC GARCH(1, 1) 0.0130 0.0571 0.0962MNIG (EM-elliptical)-CCC GARCH(1, 1) 0.0130 0.0605 0.1018MALap-CCC GJR-GARCH(1, 1) 0.0135 0.0565 0.0945Mlap (EM-elliptical)-CCC GARCH(1, 1) 0.0135 0.0611 0.1013Mt (EM-elliptical)-CCC GARCH(1, 1) 0.0135 0.0639 0.1052MAt-CCC GARCH(1, 1) 0.0147 0.0633 0.1058MALap-CCC GARCH(1, 1) 0.0152 0.0616 0.1001MLap (3-step elliptical)-CCC-GARCH(1, 1) 0.0045 0.0255 0.0617MALap-CCC GJR-GARCH(1, 1)-SV 0.0158 0.0656 0.1041Mt (3-step elliptical)-CCC-GARCH(1, 1) 0.0040 0.0357 0.0775MNIG (3-step elliptical)-CCC-GARCH(1, 1) 0.0040 0.0272 0.0623MALap-CCC GARCH(1, 1)-SV 0.0175 0.0696 0.1052MALap-IID 0.0237 0.0628 0.1007MN-cDCC GARCH(1, 1) 0.0254 0.0718 0.1041MN-DCC GARCH(1, 1) 0.0265 0.0718 0.1041MN-VC GJR-GARCH(1, 1) 0.0271 0.0696 0.1013MN-CCC GJR-GARCH(1, 1) 0.0271 0.0701 0.1013MN-DCC GJR-GARCH(1, 1) 0.0271 0.0701 0.1007MN-cDCC GJR-GARCH(1, 1) 0.0271 0.0701 0.1007MN-VC GARCH(1, 1) 0.0271 0.0724 0.1041MN-CCC GARCH(1, 1) 0.0277 0.0730 0.1041
Marc S. Paolella & Pawe l Polak COMFORT: Portfolio Optimization, Risk Management
21/52
VaR Backtesting MN-GARCH vs.MALap-IID vs.COMFORT
−10
−5
0
5
10
28/12/04 28/09/06 01/07/08 05/04/10 30/12/11
1% VaR for 1/K portfolio
1/K portfolio returnsNIG−CCC GARCH(1,1)−SV1% VaR violationsMALap−IID1% VaR violationsMN−cDCC GARCH(1,1)1% VaR violations
Failure rates: NIG-CCC GARCH(1, 1)-SV 0.0102; MALap-IID 0.0238; MN-cDCC GARCH(1, 1) 0.0255.
Marc S. Paolella & Pawe l Polak COMFORT: Portfolio Optimization, Risk Management
22/52
The Risk Fear Portfolio: “Run For The Exit”
Define, respectively, σ# and ES# as a portfolio volatility and expectedshortfall thresholds above which the investor exits the market. Then themin-Var and min-ES optimal portfolios with risk fear are defined by
w∗VF
(σ#)
= w∗1(0,σ#) (σw∗) , w∗ESF
(ES#
)= w∗1(0,ES#) (ESw∗),
where σw∗ is the predicted conditional volatility for the min-Var optimalportfolio w∗, and ESw∗ is the predicted conditional expected shortfall forthe optimal portfolio w∗.
The strategy agrees with the stylized fact of “Runs for the Exit”, i.e.,stock market major sell-offs during the initial stage of market downturns.
Marc S. Paolella & Pawe l Polak COMFORT: Portfolio Optimization, Risk Management
23/52
The Risk Fear Portfolio: Risk Thresholds
We assume that the investor uses the following simple rules to set thethresholds
σ# (δ) = σw∗ + δ std (σw∗) , ES# (δ) = ESw∗ + δ std (ESw∗) ,
where
σw∗ and ESw∗ are sample moving averages, based on 1, 000observations, of the optimal portfolio volatility and expectedshortfall, respectively;
std (σw∗) and std (ESw∗) are the associated sample standarddeviations; and
δ is a parameter to be selected by the investor, e.g., δ = 1.
Marc S. Paolella & Pawe l Polak COMFORT: Portfolio Optimization, Risk Management
24/52
The Risk Fear Portfolio: Performance
90
100
110
120
130
140
150
160
170
180
28.12.2004 28.09.2006 01.07.2008 05.04.2010 30.12.2011
Cumulative Returns of Risk Fear Strategies
ES−Fear MNIG−CCC GARCH(1,1)σ−Fear MNIG−CCC GARCH(1,1)ES−Fear MN−DCC GARCH(1,1)σ−Fear MN−DCC GARCH(1,1)MNIG−CCC GARCH(1,1)
0.8 0.85 0.9 0.95 1 1.05 1.1 1.15 1.20.4
0.5
0.6
0.7
0.8
0.9
1
1.1SR of Risk Fear Strategies
δ0 0.2 0.4 0.6 0.8 1
0
0.2
0.4
0.6
0.8
1
1.2
1.4SR of min−Var vs. SR of min−ES (Risk Fear Strategies with δ = 1)
SR(min−Var)
SR(m
in−E
S)
The minimum ES portfolios tend to perform better than the minimumvariance portfolios.In result, the SR increases from 0.49 up to 1.1.
Marc S. Paolella & Pawe l Polak COMFORT: Portfolio Optimization, Risk Management
25/52
Equities + Fixed Income Example: Three Different Portfolio Models withPortfolio Weights Constraints
(Case study done for a large Swiss Bank)
Marc S. Paolella & Pawe l Polak COMFORT: Portfolio Optimization, Risk Management
26/52
Prices, Returns and PSARM Forecasted Volatilities
21-May-2008 01-Feb-2010 12-Oct-2011 21-Jun-2013 03-Mar-201550
100
150Out-of-sample normalized prices
Fixed Income Index World Government Bond Index MSCI World Index Bond Index Rating AAA-BBB Swiss Performance Index Real Estate Funds
21-May-2008 01-Feb-2010 12-Oct-2011 21-Jun-2013 03-Mar-2015
-10
-5
0
5
10Out-of-sample returns
21-May-2008 01-Feb-2010 12-Oct-2011 21-Jun-2013 03-Mar-2015
1
2
3
4
5Forecasted volatilities
Marc S. Paolella & Pawe l Polak COMFORT: Portfolio Optimization, Risk Management
27/52
Portfolio Models, Constraints & Transaction Costs
Model 1
Asset Class Lower Limit BM Upper Limit TC - Buy TC - Sell
CASH 0.00% 0.00% 0.10% 0.00% 0.00%
Fixed Income Index 13.00% 15.00% 17.00% 0.12% 0.12%
World Government Bond Index 9.00% 10.00% 11.00% 0.09% 0.02%
MSCI World Index 24.00% 27.00% 30.00% 0.09% 0.07%
Bond Index Rating AAA - BBB 22.00% 25.00% 28.00% 0.65% 0.00%
Swiss Performance Index 16.00% 18.00% 20.00% 0.02% 0.02%
Real Estate Funds 4.00% 5.00% 6.00% 0.05% 0.05%
Model 2
Asset Class Lower Limit BM Upper Limit TC - Buy TC - Sell
CASH 0.00% 0.00% 10.00% 0.00% 0.00%
Fixed Income Index 5.00% 15.00% 25.00% 0.12% 0.12%
World Government Bond Index 0.00% 10.00% 20.00% 0.09% 0.02%
MSCI World Index 17.00% 27.00% 37.00% 0.09% 0.07%
Bond Index Rating AAA - BBB 15.00% 25.00% 35.00% 0.65% 0.00%
Swiss Performance Index 8.00% 18.00% 28.00% 0.02% 0.02%
Real Estate Funds 0.00% 5.00% 15.00% 0.05% 0.05%
Model 3
Asset Class Lower Limit BM Upper Limit TC - Buy TC - Sell
CASH 0.00% 0.00% 100.00% 0.00% 0.00%
Fixed Income Index 0.00% 15.00% 100.00% 0.12% 0.12%
World Government Bond Index 0.00% 10.00% 100.00% 0.09% 0.02%
MSCI World Index 0.00% 27.00% 100.00% 0.09% 0.07%
Bond Index Rating AAA - BBB 0.00% 25.00% 100.00% 0.65% 0.00%
Swiss Performance Index 0.00% 18.00% 100.00% 0.02% 0.02%
Real Estate Funds 0.00% 5.00% 100.00% 0.05% 0.05%
Marc S. Paolella & Pawe l Polak COMFORT: Portfolio Optimization, Risk Management
28/52
Out-Of-Sample Performance Comparison
Asset Class Correlations
Fixed Income Index 1.00 0.74 -0.28 0.46 -0.23 -0.05
World Government Bond Index 0.74 1.00 -0.38 0.55 -0.33 -0.09
MSCI World Index -0.28 -0.38 1.00 -0.22 0.74 0.17
Bond Index Rating AAA - BBB 0.46 0.55 -0.22 1.00 -0.23 -0.07
Swiss Performance Index -0.23 -0.33 0.74 -0.23 1.00 0.19
Real Estate Funds -0.05 -0.09 0.17 -0.07 0.19 1.00
Whole Sample Subperiod 1 Subperiod 2 Subperiod 3 Subperiod 4
Time Periods
21/05/2008-
03/03/2015
21/05/2008-
31/01/2010
01/02/2010-
11/10/2011
12/10/2011-
20/06/2013
21/06/2013-
03/03/2015
Asset Class \ Portfolio Strategy
Exp.
Ret. Vol. SR
Exp.
Ret. Vol. SR
Exp.
Ret. Vol. SR
Exp.
Ret. Vol. SR
Exp.
Ret. Vol. SR
Fixed Income Index 4.03% 5.46% 0.74 2.53% 6.92% 0.37 7.28% 5.19% 1.40 1.82% 4.92% 0.37 4.51% 4.48% 1.01
World Government Bond Index 3.89% 2.56% 1.52 4.55% 3.36% 1.36 3.46% 2.53% 1.36 2.35% 2.15% 1.09 5.22% 1.96% 2.66
MSCI World Index -0.28% 20.51% -0.01 -13.77% 30.12% -0.46 -7.39% 20.15% -0.37 9.90% 11.79% 0.84 10.23% 15.15% 0.67
Bond Index Rating AAA - BBB 4.04% 2.31% 1.75 6.60% 2.71% 2.44 3.39% 2.44% 1.39 1.08% 1.98% 0.55 5.09% 2.04% 2.49
Swiss Performance Index 3.58% 18.17% 0.20 -12.04% 26.01% -0.46 -0.01% 17.20% 0.00 14.47% 12.65% 1.14 12.02% 13.69% 0.88
Real Estate Funds 6.06% 6.85% 0.88 6.68% 6.65% 1.00 4.76% 6.76% 0.70 1.87% 6.40% 0.29 10.98% 7.57% 1.45
Benchmark 2.88% 7.98% 0.36 -3.07% 11.76% -0.26 0.53% 7.44% 0.07 6.15% 4.93% 1.25 7.95% 6.07% 1.31
PSARM Model 1 3.80% 7.67% 0.50 -0.97% 11.13% -0.09 1.71% 7.26% 0.23 6.46% 4.91% 1.32 8.02% 5.86% 1.37
PSARM Model 2 3.76% 4.94% 0.76 0.91% 7.22% 0.13 2.97% 4.49% 0.66 4.42% 3.35% 1.32 6.76% 3.78% 1.79
PSARM Model 3 3.69% 2.01% 1.84 3.94% 2.71% 1.46 2.43% 1.72% 1.42 2.60% 1.79% 1.45 5.80% 1.63% 3.55
PSARM outperforms the Benchmark in the whole sample, in every subperiod, and for everymodel.Relaxing the constraints on the portfolio weights improves the performance of PSARM.There is no need to invest into cash because there are highly negatively correlated assets inthe portfolio.
Marc S. Paolella & Pawe l Polak COMFORT: Portfolio Optimization, Risk Management
29/52
PSARM Cumulative Returns After Transaction Costs
21-May-2008 01-Feb-2010 12-Oct-2011 21-Jun-2013 03-Mar-2015
80
90
100
110
120
Model 1: BM vs. PSARM ( SR(BM) = 0.36, SR(PSARM) = 0.5 )
PSARMBenchmark
21-May-2008 01-Feb-2010 12-Oct-2011 21-Jun-2013 03-Mar-2015
80
90
100
110
120
Model 2: BM vs. PSARM ( SR(BM) = 0.36, SR(PSARM) = 0.76 )
PSARMBenchmark
21-May-2008 01-Feb-2010 12-Oct-2011 21-Jun-2013 03-Mar-2015
80
90
100
110
120
Model 3: BM vs. PSARM ( SR(BM) = 0.36, SR(PSARM) = 1.84 )
PSARMBenchmark
PSARM increases Sharpe ratios, from 0.36 for the Benchmark, to 0.5, 0.76, and1.84, for Model 1, Model 2, and Model 3, respectively (after transaction costs).
Very tight constraints on the portfolio weights do not allow for largeimprovements in Model 1.
Marc S. Paolella & Pawe l Polak COMFORT: Portfolio Optimization, Risk Management
30/52
COMFORT
We propose a unified framework for large-scale portfolio optimizationwith:
1. A new class of models for multivariate asset returns which:a hybrid of GARCH and SV (but without the estimation problemsassociated with the latter);can be applied to a portfolio of hundreds of assets;are simple and fast to estimate (allow for parallel estimation);support various forms of the dynamics in the dependency structure(VC, DCC, cDCC, Markov-switching);are based on a non-elliptical distribution with a tractable portfoliodistribution;nest several models previously proposed in the literature and massivelyoutperform them in terms of in-sample fit, out-of-sample densityforecasting, risk prediction, and portfolio performance.
2. A new risk fear portfolio strategy, which combines portfoliooptimization with active risk monitoring. It:
outperforms all types of optimal portfolios, even in the presence ofconservative transaction costs and frequent rebalancing;avoids all the losses during the 2008 financial crisis, and;profits from the subsequent market recovery.
Marc S. Paolella & Pawe l Polak COMFORT: Portfolio Optimization, Risk Management
31/52
Extensions of the Model
FREE-COMFORT: Fast Reduced Estimation COMFORT(Ongoing Research, joint work with Joris van der Aa from ETH)
(High Frequency Portfolio Optimization On the Components of the DowJones Industrial Index (DJ-30) from February 1st, 2010 to February 5th,
2016.)
Marc S. Paolella & Pawe l Polak COMFORT: Portfolio Optimization, Risk Management
32/52
GARCH(1, 1) Estimates Across Models
5 10 15 20 25 30-0.2
0
0.2
k for k=1,...,30
5 10 15 20 25 30-0.2
0
0.2
k for k=1,...,30
5 10 15 20 25 30-0.05
0
0.05
0.1
k for k=1,...,30
5 10 15 20 25 30-0.2
0
0.2
k for k=1,...,30
5 10 15 20 25 30-0.05
0
0.05
0.1
k for k=1,...,30
Marc S. Paolella & Pawe l Polak COMFORT: Portfolio Optimization, Risk Management
33/52
GARCH(1, 1) Estimates Across Models
MN-CCC GARCH(1, 1) MALap-CCC GARCH(1, 1) MALap-CCC GARCH(1, 1)-SV
5 10 15 20 25 300
0.1
0.2
k for k=1,...,30
5 10 15 20 25 300
0.1
0.2
k for k=1,...,30
5 10 15 20 25 300.8
0.9
1
k for k=1,...,30
5 10 15 20 25 300.9
0.95
1
k +
k for k=1,...,30
5 10 15 20 25 300
0.01
0.02
k for k=1,...,30
5 10 15 20 25 300
0.1
0.2
k for k=1,...,30
5 10 15 20 25 300.8
0.9
1
k for k=1,...,30
5 10 15 20 25 300.9
0.95
1
k +
k for k=1,...,30
5 10 15 20 25 300
0.01
0.02
k for k=1,...,30
5 10 15 20 25 300
0.1
0.2
k for k=1,...,30
5 10 15 20 25 300.8
0.9
1
k for k=1,...,30
5 10 15 20 25 300.9
0.95
1
k +
k for k=1,...,30
Marc S. Paolella & Pawe l Polak COMFORT: Portfolio Optimization, Risk Management
34/52
Data
To demonstrate the performance of the HF min fear portfolio, we use a data set consistingof 115’654 intra-day returns on the 5-minute level and 1501 overnight returns. Ourcombined data set consists of 117’155 returns for K = 30 components of the Dow JonesIndustrial Index (DJ-30) from February 1st, 2010 to February 5th, 2016.
We take all DJ-30 components that are part of the composition as of 01.12.2015.
Each trading day consists of 77 equally spaced time points on which it is possible to trade.
The first possible trading moment of the day is at 9:31AM and the last trading moment isat 3:54PM.
At 3:54PM, we take on the optimal portfolio w∗ for 3:59PM and hold this positionovernight.
For our trading strategy, it is not possible to rebalance, close or open any positions before9.31AM or after 15.54PM.
Marc S. Paolella & Pawe l Polak COMFORT: Portfolio Optimization, Risk Management
35/52
Entry & Exit Thresholds
We set the thresholds for entering and exiting the market using the following two rules
ESδexit = ESw∗ + δexitstd(ESw∗ ), (3)
ESδenter = ESw∗ + δenterstd(ESw∗ ), (4)
where ESw∗ is the sample moving average, based on the last 1000 observations of the optimalportfolio expected shortfall; std(ESw∗ ) is the associated sample standard deviation; and δenter andδexit are parameters to be selected by the investor with δexit > δenter.We term our trading strategy the High Frequency Risk Fear Portfolio Strategy (hereafter HF minfear).
Marc S. Paolella & Pawe l Polak COMFORT: Portfolio Optimization, Risk Management
36/52
Gross Performance
19-Feb-2010 12-Nov-2010 11-Aug-2011 09-May-2012 11-Feb-2013 06-Nov-2013 08-Aug-2014 08-May-2015 05-Feb-201605-Feb-20160
100
200
300
400
500
600Gross Cumulative Returns of HF min Fear Strategy vs. Other Models
HF min Fearmin RiskNaive (rebal)Gaussian IID
19-Feb-2010 29-Sep-2010 11-May-2011 20-Dec-2011 02-Aug-2012 20-Mar-2013 29-Oct-2013 13-Jun-2014 05-Feb-2015-1
0
1
2
3
4Sharpe Ratio as a Function of the Entry Date
HF min Fearmin RiskNaive (rebal)Gaussian IID
19-Feb-2010 29-Sep-2010 11-May-2011 20-Dec-2011 02-Aug-2012 20-Mar-2013 29-Oct-2013 13-Jun-2014 05-Feb-2015-1
0
1
2
3
4Sortino Ratio as a Function of the Entry Date
HF min Fearmin RiskNaive (rebal)Gaussian IID
Marc S. Paolella & Pawe l Polak COMFORT: Portfolio Optimization, Risk Management
37/52
Parameters Calibration
Marc S. Paolella & Pawe l Polak COMFORT: Portfolio Optimization, Risk Management
38/52
Analysis of Net Performance for Optimal Parameters
16-Apr-2010 04-Jan-2011 22-Sep-2011 14-Jun-2012 12-Mar-2013 27-Nov-2013 22-Aug-2014 15-May-2015 05-Feb-201605-Feb-201650
100
150
200
250Cumulative Returns of HF min Fear Strategy vs. Other Models for κ = 0.0005
HF min Fearmin RiskNaive (rebal)Gaussian IID
16-Apr-2010 16-Nov-2010 22-Jun-2011 26-Jan-2012 29-Aug-2012 10-Apr-2013 11-Nov-2013 19-Jun-2014 05-Feb-2015-0.5
0
0.5
1
1.5
2Sharpe Ratio as a Function of the Entry Date
HF min Fearmin RiskNaive (rebal)Gaussian IID
16-Apr-2010 16-Nov-2010 22-Jun-2011 26-Jan-2012 29-Aug-2012 10-Apr-2013 11-Nov-2013 19-Jun-2014 05-Feb-2015-0.5
0
0.5
1
1.5
2Sortino Ratio as a Function of the Entry Date
HF min Fearmin RiskNaive (rebal)Gaussian IID
Marc S. Paolella & Pawe l Polak COMFORT: Portfolio Optimization, Risk Management
39/52
COMFORT & Machine Learning forImproved Portfolio Performance
Marc S. Paolella & Pawe l Polak COMFORT: Portfolio Optimization, Risk Management
40/52
Trend Chasing Product For Improved PortfolioPerformance
We combine statistical and machine learning techniques topredict future returns of the components of a largeportfolio of assets.
Advanced Multivariate Times Series ModelingCaptures The Dynamics of Asset Returns.
Machine Learning Techniques Allow to Predict FutureReturns.
Marc S. Paolella & Pawe l Polak COMFORT: Portfolio Optimization, Risk Management
41/52
Portfolio Optimization with TrendPrediction
We backtest our Portfolio Strategy with Trend Prediction on 4 setsof stock returns:
Long History of the returns: daily returns of the 30 components ofthe Dow Jones Index, from 04.01.1993 until 31.12.2012.
The most recent data: daily returns of the 30 components of theDow Jones Index, more recent data and components from01.06.1999 until 31.12.2014.
Large Set of Assets: daily returns of the 459 components of the S&P500 Index, from 04.01.2009 until 15.05.2014.
High Frequency: 5 minutes returns of the 30 components of the DowJones Index, from 02.01.2009, 10.00.00 until 31.12.2009, 15.25.00.
In the backtesting procedure we used a rolling window of 250observations. First out-of-sample date is 29.12.1993, 24.05.2000,05.01.2010, and 07.01.2009, 14.25.00, respectively.
Marc S. Paolella & Pawe l Polak COMFORT: Portfolio Optimization, Risk Management
42/52
Portfolio Optimization with TrendPrediction
The strategy uses only information on past returns, does not usefuture returns, nor any exogenous variables.
It is a long-only portfolio (for negative predicted means, the portfolioweights are set to zero).
The Trend Prediction Portfolio Strategy can be integrated with theexisting portfolio strategy in a form of views on the marketanalogously to the views of the investor in the Black-Littermanmodel.
Marc S. Paolella & Pawe l Polak COMFORT: Portfolio Optimization, Risk Management
43/52
Equities Example: Components of DJ-30 Index(1993-2012)
Marc S. Paolella & Pawe l Polak COMFORT: Portfolio Optimization, Risk Management
44/52
Trend Prediction: DJ-30 Components(1993-2012)
100
200
300
400
500
600
700
29.12.1993 19.09.1995 09.06.1997 02.03.1999 16.11.2000 16.08.2002 10.05.2004 31.01.2006 23.10.2007 16.07.2009 06.04.2011 28.12.2012
Cumulative Returns of Trend Chasing Portfolio vs. 1/N Portfolio
Trend Chasing Portfolio: E(R) = 31%; vol(R) = 26%; SR(R) = 1.211/N Portfolio: E(R) = 13%; vol(R) = 20%; SR(R) = 0.64
The cumulative returns of the portfolio are plotted over time, and compared with theequally weighted portfolio (1/N portfolio).
For the Trend Chasing Portfolio the average annual return, volatility, and Sharpe ratio aregiven by: 31%, 26%, 1.21, respectively.
While the equally weighted portfolio (proxy for the market) has the average annual return,volatility, and Sharpe ratio equal to 13%, 20%, 0.64, respectively.
Marc S. Paolella & Pawe l Polak COMFORT: Portfolio Optimization, Risk Management
45/52
Equities Example: Components of DJ-30 Index(2000-2014)
Marc S. Paolella & Pawe l Polak COMFORT: Portfolio Optimization, Risk Management
46/52
Trend Prediction: DJ-30 Components(2000-2014)
50
100
150
200
250
300
350
400
24.05.2000 14.01.2002 27.08.2003 12.04.2005 21.11.2006 09.07.2008 22.02.2010 03.10.2011 20.05.2013 31.12.2014
Cumulative Returns: Trend Chasing vs. 1/N Portfolio (Benchmark), daily DJ−30 Components Data
1/N Portfolio E(R) = 7.5%; vol(R) = 20%; SR(R) = 0.38Trend Chasing E(R) = 20%; vol(R) = 25%; SR(R) = 0.79
The cumulative returns of the portfolio are plotted over time, and compared with theequally weighted portfolio (1/N portfolio).
For the Trend Chasing Portfolio the average annual return, volatility, and Sharpe ratio aregiven by: 20%, 25%, 0.79, respectively.
While the equally weighted portfolio (proxy for the market) has the average annual return,volatility, and Sharpe ratio equal to 7.5%, 20%, 0.38, respectively.
Marc S. Paolella & Pawe l Polak COMFORT: Portfolio Optimization, Risk Management
47/52
DJ-30 Components (2000-2014) WithReinvestment
100
200
300
400
500
600
700
800
900
1000
1100
24.05.2000 14.01.2002 27.08.2003 12.04.2005 21.11.2006 09.07.2008 22.02.2010 03.10.2011 20.05.2013 31.12.2014
Cumulative Returns: Trend Chasing vs. 1/N Portfolio (Benchmark), daily DJ−30 Components Data
1/N Portfolio E(R) = 7.5%; vol(R) = 20%; SR(R) = 0.38Trend Chasing E(R) = 20%; vol(R) = 25%; SR(R) = 0.791/N Portfolio (with reinvestment)Trend Chasing (with reinvestment)
The cumulative returns of the portfolio are plotted over time, and compared with theequally weighted portfolio (1/N portfolio).
Two types of portfolio with and without reinvestment.
For the Trend Chasing Portfolio the average annual return, volatility, and Sharpe ratio aregiven by: 20%, 25%, 0.79, respectively.
While the equally weighted portfolio (proxy for the market) has the average annual return,volatility, and Sharpe ratio equal to 7.5%, 20%, 0.38, respectively.
Marc S. Paolella & Pawe l Polak COMFORT: Portfolio Optimization, Risk Management
48/52
Equities Example: Components of SP-500 Index(2010-2014)
Marc S. Paolella & Pawe l Polak COMFORT: Portfolio Optimization, Risk Management
49/52
Portfolio with Trend Prediction & Proportional TransactionCosts
100
120
140
160
180
200
220
240
260
280
05/01/2010 18/08/2010 04/04/2011 15/11/2011 29/06/2012 15/02/2013 30/09/2013 15/05/2014
Cumulative Returns: Trend Chasing vs. 1/N Portfolio (Benchmark), 459 stocks from the SP 500 Index, daily data
Trend Chasing, E(R)=42.78%; vol(R)=34.09%; SR(R)=1.25Trend Chasing TC=5bp (standard) E(R)=40.11%; vol(R)=34.08%; SR(R)=1.18Trend Chasing TC=10bp (conservative) E(R)=37.44%; vol(R)=34.08%; SR(R)=1.11/N Portfolio E(R)=12.85%; vol(R)=18.31%; SR(R)=0.7
The cumulative returns of the portfolio are plotted over time, and compared with theequally weighted portfolio (1/N portfolio).
The average annual return, volatility, and Sharpe ratio for the Trend Chasing Portfolio:(i) TC = 0: 42.78%, 34.09%, 1.25;
(ii) TC = 5bp (standard): 40.11%, 34.08%, 1.18; and(iii) TC = 10bp (conservative): 37.44%, 34.08%, 1.1, respectively.
While the equally weighted portfolio (proxy for the market) has the average annual return,volatility, and Sharpe ratio equal to 12.85%, 18.31%, 0.7, respectively.Marc S. Paolella & Pawe l Polak COMFORT: Portfolio Optimization, Risk Management
50/52
Equities Example: Components of DJ-30 Index
High Frequency Portfolio Strategy: 5 Minutes Data(07.01.2009 - 31.12.2009)
Marc S. Paolella & Pawe l Polak COMFORT: Portfolio Optimization, Risk Management
51/52
Portfolio with Trend Prediction & Proportional TransactionCosts: HF
80
100
120
140
160
180
07JAN09 14:25:00 02MAR09 10:45:00 21APR09 12:30:00 10JUN09 14:15:00 31JUL09 10:35:00 21SEP09 12:20:00 09NOV09 14:05:00 31DEC09 15:20:00
Cumulative Returns: Trend Chasing vs. 1/N Portfolio (Benchmark), 5 minutes DJ 30 components data (02.01.2009 10:00:00 − 31.12.2009 15:25:00)
Trend Chasing, E(R)=78.96%; vol(R)=41.22%; SR(R)=1.92
Trend Chasing TC=5bp (standard) E(R)=59.99%; vol(R)=41.22%; SR(R)=1.46
Trend Chasing TC=10bp (conservative) E(R)=41.01%; vol(R)=41.22%; SR(R)=0.99
1/N Portfolio E(R)=17.06%; vol(R)=25.76%; SR(R)=0.66
The cumulative returns of the portfolio are plotted over time, and compared with theequally weighted portfolio (1/N portfolio).
The average annual return, volatility, and Sharpe ratio for the Trend Chasing Portfolio:(i) TC = 0: 78.96%, 41.22%, 1.92;
(ii) TC = 5bp (standard): 59.99%, 41.22%, 1.46; and(iii) TC = 10bp (conservative): 41.01%, 41.22%, 0.99, respectively.
While the equally weighted portfolio (proxy for the market) has the average annual return,volatility, and Sharpe ratio equal to 17.06%, 25.76%, 0.66, respectively.
Marc S. Paolella & Pawe l Polak COMFORT: Portfolio Optimization, Risk Management
52/52
• VaR Performance • ES Performance • MLE > Ad-Hoc 2-Step • JRSS-A 2016
• MGHyp-CCC-GARCH
Framework • EM Algo. Estimation • Common Market
Factor • Option Pricing • JoE 2015
• Analytic ES • DCC Modeling • Portfolio Optimization • Risk Tracking Method • Sharpe Ratios > 1 • Submitted
Risk Modeling
Primary Model COMFORT
PSARM
• Leverage Investing • Risk Trackers • Tactical Asset
Allocation • Large Swiss Banks
and Pension Funds
Generalizing the MGHyp: • Multi-Factors for Full
Tail Heterogeneity • Lévy-Stable /
Tempered Stable
• FREE-COMFORT • Markov-Switching CCC • PCA Factor Structure • High-Freq. Strategy • Black-Litterman • Machine Learning
COMFORT Road Map Highway
Academic
Industry Consulting
Large-Scale Port. Opt.
30 min talk!!!
Thank you for your attention!Pawe l Polak
Marc S. Paolella & Pawe l Polak COMFORT: Portfolio Optimization, Risk Management