why distributions matter ( 20 dec 2013 )
DESCRIPTION
Presentation on deficiencies of the Cornish Fisher modification to the Normal distribution and also on some Best Fit distributions including Gumbel, Johnson Family, Mixture of Normals, Skew-T, 3-Parameter Lognormal etc. Also includes bi-variate Best Fit Copula correlation with applications to Pairs Trading.TRANSCRIPT
Why Distributions
Matter
By Peter Urbani
16 Jan 2012
My ventures are not in one bottom trusted,Nor to one place; nor is my whole estate Upon the fortune of this present year;Therefore, my merchandise makes me not sad.
Spoken by Antonio in Act I, Scene I, Merchant of Venice, William Shakespeare, circa 1650
Why Distributions Matter
Rembrandt’s Christ in the storm on the lake of Galilee – the cover illustration of Peter Bernstein’s excellent Against the God’s – The Remarkable Story of Risk
Diversification
Diversification is the key principle upon which Modern Portfolio Theory (MPT) is built - although the concept of not putting all of one’s eggs into one basket dates all the way back to biblical times.
Central to this is the concept of Correlation as measure of dependence between assets
Key Assumptions
Correlation ( the standardised covariance between assets ) as the measure of dependence between assets
Normally distributed returns. The assumption that asset returns are normally distributed about their means.
They make an ASS out of U and ME
The Problem with Assumptions
Normality Testing
Normality Testing
ETF's
Normal
Not-Normal
Hedge Funds
Normal
Not-Normal
Assumed NormalAssumed Normal
Normal VaR
Normal CVaR
-40.00% -30.00% -20.00% -10.00% 0.00% 10.00% 20.00% 30.00% 40.00%
Assumed Normal FundPDF (Normal)
If not Normal then What ? - Modified
VaR ( Modified ) lower for small positive skew
Cornish Fisher - Modification
VaR ( Modified ) higher for small negative skew
Cornish Fisher - Modification
VaR ( Modified ) same as VaR ( Normal ) for no skew
Assumed NormalAssumed Normal
Normal VaR
Normal CVaR
-40.00% -30.00% -20.00% -10.00% 0.00% 10.00% 20.00% 30.00% 40.00%
Assumed Normal FundPDF (Normal)
Assumed ModifiedAssumed Normal and Modified Distributions
Normal VaR
Normal CVaR
Modified CVaR -
Modified VaR -
-40.00% -30.00% -20.00% -10.00% 0.00% 10.00% 20.00% 30.00% 40.00%
Assumed Normal FundPDF (Normal)
Assumed ModifiedNormal Fund PDF(Modified)
Cornish Fisher - Modification1.3.2 Properties
The qualitative properties of the Cornish-Fisher expansion are:
If is a sequence of distributions converging to the standard normal distribution , the Edgeworth- and Cornish-Fisher approximations present better approximations (asymptotically for ) than the normal approximation itself.
The approximated functions and are not necessarily monotone.
has the ``wrong tail behavior'', i.e., the Cornish-Fisher approximation for -quantiles becomes less and less reliable
for (or ).
The Edgeworth- and Cornish-Fisher approximations do not necessarily improve (converge) for a fixed and increasing order of approximation, .
For more on the qualitative properties of the Cornish-Fisher approximation see (Jaschke; 2001). It contains also an empirical analysis of the error of the Cornish-Fisher approximation to the 99%-VaR in real-world examples as well as its worst-case error on a certain class of one- and two-dimensional delta-gamma-normal models:
http://fedc.wiwi.hu-berlin.de/xplore/tutorials/xfghtmlnode8.html
Problems with Modified VaR – Not MonotoneModified VaR as a function of Skewness
-10.00%
-5.00%
0.00%
5.00%
10.00%
15.00%
20.00%
-6 -4 -2 0 2 4 6
Skewness
VaR
( M
od
ifie
d )
Modified VaR @ CL 99.00%Modified VaR @ CL 95.00%CurrentCurrent
http://discussions.ft.com/alchemy/forums/edhec-risk-forum/hedge-fund-risk-management-models-for-the-return-distribution/
0.00%
10.00%
20.00%
30.00%
40.00%
50.00%
60.00%
70.00%
80.00%
90.00%
100.00%
-0.5 -0.4 -0.3 -0.2 -0.1 0 0.1 0.2 0.3 0.4
Modified CDF
Normal CDF
0.00%
500.00%
1000.00%
1500.00%
2000.00%
2500.00%
-30.0% -20.0% -10.0% 0.0% 10.0% 20.0% 30.0%
Modified PDF
Normal PDF
-6
0
6
12
18
-6 -4 -2 0 2 4 6
↔Skew
↕ Kurt
S
GoodZ Z
S S
Z
Problems with Modified VaR – Bad Tail Behaviour
Problems with Modified VaR – Bad Tail Behaviour
0.00%
10.00%
20.00%
30.00%
40.00%
50.00%
60.00%
70.00%
80.00%
90.00%
100.00%
-0.5 -0.4 -0.3 -0.2 -0.1 0 0.1 0.2 0.3 0.4
Modified CDF
Normal CDF
0.00%
500.00%
1000.00%
1500.00%
2000.00%
2500.00%
-30.0% -20.0% -10.0% 0.0% 10.0% 20.0% 30.0%
Modified PDF
Normal PDF
-6
0
6
12
18
-6 -4 -2 0 2 4 6
↔Skew
↕ Kurt
S
GoodZ Z
S S
Z
How Prevalent is this problem ?VERY
ETF's
OK
WARNING: DegenerateCornish Fisher. CDFw ill turn in tails
WARNING: DegenerateCornish Fisher. CDFw ill turn in body
Hedge Funds
OK
WARNING: DegenerateCornish Fisher. CDFw ill turn in tails
WARNING: DegenerateCornish Fisher. CDFw ill turn in body
Public Function CFRegionWarning(ByVal Skew As Double, Kurt As Double) As String
Dim a As Double Dim b As Double Dim c As Double Dim Q As Double Dim R As Double Dim Denom As Double
Denom = 3 * Kurt - 4 * (Skew ^ 2)
If Denom > 0 Then a = 12 * Skew / Denom b = (10 * Skew ^ 2 - 9 * Kurt + 72) / Denom c = -12 * Skew / Denom Q = (a ^ 2 - 3 * b) / 9 R = (2 * (a ^ 3) - 9 * a * b + 27 * c) / 54 If R ^ 2 > Q ^ 3 Then CFRegionWarning = "" 'Its in Well Behaved Region Else CFRegionWarning = "WARNING: Degenerate Cornish Fisher. CDF will turn in body (S)" End If Else CFRegionWarning = "WARNING: Degenerate Cornish Fisher. CDF will turn in tails (Z)" End If
End Function
VBA code to check Cornish Fisher - Modification
If not Normal or Modified then What ?
ETF's
Gumbel (Max) 7.8%
Gumbel (Min) 8.2%
Johnson (Lognormal) 13.5%
Johnson (Unbounded) 0.6%
Mixture of Normals 20.1%
Modified Normal 35.8%
Normal 12.9%
Uniform 1.0%
Hedge Funds
Gumbel (Max) 23.3%
Gumbel (Min) 13.2%
Johnson (Lognormal) 13.2%
Johnson (Unbounded) 2.6%
Mixture of Normals 15.9%
Modified Normal 21.2%
Normal 6.9%
Uniform 3.7%
Best Fitting Distributions
Assumed NormalAssumed Normal
Normal VaR
Normal CVaR
-40.00% -30.00% -20.00% -10.00% 0.00% 10.00% 20.00% 30.00% 40.00%
Assumed Normal FundPDF (Normal)
Assumed ModifiedAssumed Normal and Modified Distributions
Normal VaR
Normal CVaR
Modified CVaR -
Modified VaR -
-40.00% -30.00% -20.00% -10.00% 0.00% 10.00% 20.00% 30.00% 40.00%
Assumed Normal FundPDF (Normal)
Assumed ModifiedNormal Fund PDF(Modified)
Best FittingBest Fit and Assumed Normal and Modified Distributions
Best Fit VaR -
Best Fit CVaR -
Normal VaR
Normal CVaR
Modified CVaR -
Modified VaR -
-40.00% -30.00% -20.00% -10.00% 0.00% 10.00% 20.00% 30.00% 40.00%
Best Fit Fund PDF(Gumbel (Min))
Assumed Normal FundPDF (Normal)
Assumed ModifiedNormal Fund PDF(Modified)
Goodness of FitCummulative Probability Distributions
0%
10%
20%
30%
40%
50%
60%
70%
80%
90%
100%
-40.00% -30.00% -20.00% -10.00% 0.00% 10.00% 20.00% 30.00% 40.00%
Best Fit Fund CDF(Gumbel (Min))
Assumed Normal FundCDF (Normal)
Empirical CDF Fund
P-P Plot (Showing Goodness of Fit for Fund)
-25%
-20%
-15%
-10%
-5%
0%
5%
10%
15%
20%
-25% -20% -15% -10% -5% 0% 5% 10% 15%
Series1
Fund Best Fit (Gumbel (Min))
Fund Assumed (Normal)
Relationship Between Assets
-4.00%
-2.00%
0.00%
2.00%
4.00%
6.00%
8.00%
10.00%
-6.00% -4.00% -2.00% 0.00% 2.00% 4.00% 6.00% 8.00% 10.00%
FUND B
FUN
D A
Linear Regression
Relationship between Fund A and B
Assumed Bi-Variate Normal Copula
-4.00%
-2.00%
0.00%
2.00%
4.00%
6.00%
8.00%
10.00%
-6.00% -4.00% -2.00% 0.00% 2.00% 4.00% 6.00% 8.00% 10.00%
FUND B
FUN
D A
Assumed Bi-variateNormal Copula Lines
Linear Regression
Best Fit Bi-Variate Copula
-4.00%
-2.00%
0.00%
2.00%
4.00%
6.00%
8.00%
10.00%
-6.00% -4.00% -2.00% 0.00% 2.00% 4.00% 6.00% 8.00% 10.00%
FUND B
FUN
D A
Best Fit Bi-variateCopula Lines
Best Fit Regression
Linear Regression
-20.00%
-10.00%
0.00%
10.00%
20.00%
30.00%
40.00%
50.00%
-8.00% -6.00% -4.00% -2.00% 0.00% 2.00% 4.00% 6.00% 8.00% 10.00%
FUND D
FUN
D C
Best Fit Bi-variateCopula Lines
Best Fit Regression
Linear Regression
Example – Johnson Lognormal - Normal
Example – Normal - Normal
-15.00%
-10.00%
-5.00%
0.00%
5.00%
10.00%
15.00%
-6.00% -4.00% -2.00% 0.00% 2.00% 4.00% 6.00% 8.00%
FUND F
FUN
D E
Best Fit Bi-variateCopula Lines
Best Fit Regression
Linear Regression
Example – Modified Normal - Normal
-4.00%
-2.00%
0.00%
2.00%
4.00%
6.00%
8.00%
-15.00% -10.00% -5.00% 0.00% 5.00% 10.00% 15.00%
FUND H
FUN
D G
Best Fit Bi-variateCopula Lines
Best Fit Regression
Linear Regression
Example – Mix Normals – Modified Normal
-40.00%
-30.00%
-20.00%
-10.00%
0.00%
10.00%
20.00%
30.00%
-20.00% -15.00% -10.00% -5.00% 0.00% 5.00% 10.00% 15.00% 20.00% 25.00% 30.00%
FUND J
FUN
D I
Best Fit Bi-variateCopula Lines
Best Fit Regression
Linear Regression
Example – Mod Normal – Mod Normal
-40.00%
-30.00%
-20.00%
-10.00%
0.00%
10.00%
20.00%
30.00%
40.00%
-50.00% -40.00% -30.00% -20.00% -10.00% 0.00% 10.00% 20.00% 30.00% 40.00% 50.00%
FUND L
FUN
D K
Best Fit Bi-variateCopula Lines
Best Fit Regression
Linear Regression
Relationship Between Assets2011 Asset Class Correlations
-40.00%
-30.00%
-20.00%
-10.00%
0.00%
10.00%
20.00%
30.00%
40.00%
-1.00 -0.50 0.00 0.50 1.00
Correlation to S&P500
To
tal
Re
turn
s (
% )
Relationship Between Assets2011 Asset Class Correlations
Volatility
US Dollar Oil
Gold
US Real EstateUS High Yield Bonds
Emerging Markets Bonds
International Government Bonds
US Government Bonds
US Total Bond Market
US Bonds
Global Bond Index
Emerging Market Equities
World Equities
Cash US Equities
CTA'sHedge Funds
-20.00%
-15.00%
-10.00%
-5.00%
0.00%
5.00%
10.00%
15.00%
20.00%
-1.00 -0.50 0.00 0.50 1.00
Correlation to S&P500
To
tal R
etu
rns
( %
)
Correlations at multi-decade highs
Best Fit and Pearson Correl Pairs with CAGR > 7%
Base Fund Lowest Correlation via Best Fit Lowest Correlation via Pearson Correl
-40.00%
-30.00%
-20.00%
-10.00%
0.00%
10.00%
20.00%
30.00%
-25.00% -20.00% -15.00% -10.00% -5.00% 0.00% 5.00% 10.00% 15.00% 20.00% 25.00%
United States Oil Fund
Corn
Fund
Best Fit Bi-variateCopula Lines
Best Fit Regression
Linear Regression
-40.00%
-30.00%
-20.00%
-10.00%
0.00%
10.00%
20.00%
30.00%
-25.00% -20.00% -15.00% -10.00% -5.00% 0.00% 5.00% 10.00% 15.00% 20.00%
DB Crude Oil Long ETN
Corn
Fund
Best Fit Bi-variateCopula Lines
Best Fit Regression
Linear Regression
Best Fit Pair CORN - OLO Pearson Pair CORN - USO
-6.00%
-4.00%
-2.00%
0.00%
2.00%
4.00%
6.00%
8.00%
10.00%
12.00%
-20.00% -15.00% -10.00% -5.00% 0.00% 5.00% 10.00% 15.00%
SPECTRUM Lg Cap U.S. Sector ETN
Barc
lays
ETN
+ S
&P
VEQ
TOR E
TN
Best Fit Bi-variateCopula Lines
Best Fit Regression
Linear Regression
-6.00%
-4.00%
-2.00%
0.00%
2.00%
4.00%
6.00%
8.00%
10.00%
12.00%
-15.00% -10.00% -5.00% 0.00% 5.00% 10.00% 15.00% 20.00% 25.00% 30.00% 35.00%
Ultra Consumer Services
Barc
lays
ETN
+ S
&P
VEQ
TOR E
TN
Best Fit Bi-variateCopula Lines
Best Fit Regression
Linear Regression
Best Fit Pair VQT - UCC Pearson Pair VQT - EEH
-10.00%
-5.00%
0.00%
5.00%
10.00%
15.00%
20.00%
25.00%
-10.00% -5.00% 0.00% 5.00% 10.00% 15.00%
SPDR Barclays Capital High Yield Bond ETF
US
Treasu
ry L
ong B
ond B
ull
ETN
Best Fit Bi-variateCopula Lines
Best Fit Regression
Linear Regression
-10.00%
-5.00%
0.00%
5.00%
10.00%
15.00%
20.00%
25.00%
-8.00% -6.00% -4.00% -2.00% 0.00% 2.00% 4.00% 6.00% 8.00% 10.00% 12.00%
iBoxx $ HY Corp Bond Fund
US
Treasu
ry L
ong B
ond B
ull
ETN
Best Fit Bi-variateCopula Lines
Best Fit Regression
Linear Regression
Best Fit Pair DLBL - HYG Pearson Pair DLBL - JNK
-15.00%
-10.00%
-5.00%
0.00%
5.00%
10.00%
15.00%
20.00%
-20.00% -15.00% -10.00% -5.00% 0.00% 5.00% 10.00% 15.00% 20.00%
SPDR DJ Wilshire Global Real Estate ETF
Wils
hire
US
REI
T ET
F
Best Fit Bi-variateCopula Lines
Best Fit Regression
Linear Regression
-30.00%
-20.00%
-10.00%
0.00%
10.00%
20.00%
30.00%
40.00%
-15.00% -10.00% -5.00% 0.00% 5.00% 10.00% 15.00%
FTSE NAREIT Real Estate 50 Index Fund
Wils
hire
US
REI
T ET
F
Best Fit Bi-variateCopula Lines
Best Fit Regression
Linear Regression
Best Fit Pair WREI - FTY Pearson Pair WREI - RWO
Other Measures of Dependence
• Spearman’s Rank Order Correlation
• Lin’s Concordance measure
• Copula methods
• Distance measures
• Mutual Information and other entropy based measures
Conclusions• Distributions do differ from Normal at least 15 – 20% of the time and up to
30 – 40% of the time depending on the data set being used – Test them
• The Cornish Fisher modification is not strictly monotone and should probably not be used at confidence levels above 95%
• The Cornish Fisher modification has poor tail behaviour almost half of the time – CHECK
• Correlation is a limited and linear measure of dependence only
• Non-Linear Copula based methods offer significant promise in helping to find better diversification and pairs trading opportunities
Pietro (‘Peter’) Urbani (45)• Chief Investment Officer (CIO) – Infiniti Capital $3bn Fund of Hedge Funds Group
Head of Quantitative Research – Infiniti Capital
• CEO – KnowRisk Consulting – Asset Consulting
• Head of Investment Strategy – Fairheads Asset Managers HNW Trust and Investment BoutiqueHead of Research – Fairheads Asset Managers
• Head of Portfolio Management – Nexus Securities
• Senior Portfolio Manager – Commercial Union – Superfund
• Equities Dealer – Junior Portfolio Manager – Mathison & Hollidge Stockbrokers
http://nz.linkedin.com/in/peterurbani