non!linear!dependence!structures:!a!copulaopinion … awards 2013 - jean-da… · keating! and!...

11
Quant Awards 2013 – September 2013 1 Non Linear Dependence Structures: a Copula Opinion Approach in Portfolio Optimization JeanDamien Villiers ESSEC Business School Master of Sciences in Management – Grande Ecole September 2013

Upload: dangkhanh

Post on 30-May-2018

212 views

Category:

Documents


0 download

TRANSCRIPT

 Quant  Awards  2013  –  September  2013       1  

           

Non  Linear  Dependence  Structures:  a  Copula  Opinion  Approach  in  Portfolio  Optimization  

 

 Jean-­‐Damien  Villiers  

 

 

ESSEC  Business  School  

 Master  of  Sciences  in  Management  –  Grande  Ecole    

 

 

 

September  2013    

 

 

   

 Quant  Awards  2013  –  September  2013       2  

Non  Linear  Dependence  Structures:  a  Copula  Opinion  Approach  in  Portfolio  Optimization      

 Abstract  

 The  recent   financial   crisis  brought  out   the   “Risk-­‐on/Risk-­‐off”  notion,   characterized  by  periods  where   the  market   is  not   driven  any  more  by   the   fundamental   analysis.  Consequently,  a  question  about  the  effectiveness  of  the  classic  portfolio  optimization  models  has  risen:  the  strong  rise  of  cross  asset  correlations  suggests  integrating  new  market   anomalies   into   models,   such   as   the   non-­‐normality   of   returns   or   the   non-­‐linear  dependence  structures.  The  aim  of  this  paper  is  to  provide  a  methodology  to  implement   a   copula   approach   model,   derived   from   the   Black-­‐Littermann  methodology.   The   model   outperforms   the   Black-­‐Littermann   model,   and   reduces  sensitively  the  risk  of  our  portfolio.  

 

INTRODUCTION    

Markowitz  proved   in  1954   that  an  appropriate  pool  of  assets  could  reduce   the  systematic  risk  of  a  portfolio  without  impacting  its  return  expectancy.  However,  several  criticisms  have  been  made  on  this  model,  discussing  the  simplicity  of  the  optimization  based  on  only  mean  and  variance,  the  normality  of  returns  assumed,  and  its  operational  non-­‐stability  making  the  model   hardly   usable   in   practice.   Since   then,   several   models   have   emerged.   The   Black-­‐Littermann  model   is  one  of   them,  widely  used  by  asset  managers:  based  on  both  Bayesian  markets   forecasts   and   mean-­‐variance   optimization,   it   solves   the   issue   of   stability   on  expected   returns,   and   proposes   a   way   to   rely   both   on   fundamental   and   quantitative  analysis.    

 

Despite   the   democratization   of   this   methodology,   several   weaknesses   still   remain:   the  normality  of  asset  returns  and  their  linear  codependence  modeled  by  a  variance-­‐covariance  matrix   are   not  well   adapted   for   the   stressed   scenarios  we’ve   observed   over   the   previous  years.   Since   the   bankruptcy   of   Lehman   Brothers   a   new   constraint   must   be   taken   into  account:   we’ve   observed   a   sharp   rise   of   correlations   between   assets   historically  uncorrelated,   and   it   seems   necessary   to   widen   the   possibilities   of   the   Black-­‐Littermann  model  to  take  these  anomalies  into  account.  

 

In   this  research  paper,  we  propose  a  methodology  based  on  codependence  structures  and  marginal   distribution   functions   to  model   the  market,   keeping   the   strengths   of   the   Black-­‐Littermann  method.  More   particularly,  we  define  marginal   distribution   functions   for   each  type  of  asset,  and  quantify  the  correlations  between  them  through  copulas.  Finally,  we  make  our  asset  pooling  according   to   the  optimization  of  Omega  ratio,  a  risk  measure  depending  on  the  whole  distribution  of  returns.    

 

   

 Quant  Awards  2013  –  September  2013       3  

METHODOLOGY    

The   first   objective   of   the   model   is   to   simulate   the   expected   future   evolution   of   the   market  (“Posterior  Market  Distribution”),  mixing  the  results  of  a  quantitative  model  based  on  historical  data  (“Prior  Market  Distribution”)  and  the  forecasts  of  an  investor  (“Views  Distribution”).  The  cumulative   distribution   functions   are   then   averaged   according   to  weights   depending   on   the  confidence  levels  the  investor  has  on  his  views  (See  Appendix  1).  

 

• The  Prior  Market  Distribution  is  the  result  of  Monte-­‐Carlo  simulations  based  on  independent  marginal  distributions  for  each  asset,  and  copulas  to  model  the  interdependency.  As  we  intend  to  quantify  non-­‐linear  dependencies  between  assets   in   cases  of  highly   stressed  markets,   it   is  necessary  to  use  marginal  distributions  adapted  to  leptokurtic  returns.  We  assume  log  normal  distributions   for   FX   and   Credit   products,   and   use   the   Heston   model   for   Commodities   and  Equities   products,   a   stochastic   volatility   model   widely   used   for   its   calibration   easy   to  implement,  defined  by  the  following  stochastic  differential  equations:    !!! = !!!  !" +   !!!!!!!

!    !!! = !(! − !!)  !" +  ! !!!!!

!  

where   !!  is   the   stock   price,   !!  the   instantaneous   variance,   !!!!"#  !!

! are  Wiener  processes  with   correlation   ρ,   μ   is   the   rate   of   return   of   the   asset,   θ   is   the  long   term   expected  variance,  κ  the  rate  at  which  Vt  reverts  to  θ,  and  ξ  is  the  volatility  of  volatility.  

The  dependence   structure  between   assets   is   then  modeled  by   the   Student   copula   (t-­‐copula),  extracted   from   the   bivariate   Student   distribution,   according   to   the   Sklar’s   theorem   (see  Appendix  2):  

! !!, !!; !, ! = !!,!(!!!! !! ,!!!! !! )  

with  ρ  the  correlation  coefficient  and  !!,!  the  bivariate  Student  distribution  with  a  correlation  matrix  depending  on    ρ  and  k,  the  degrees  of  freedom.  

The  choice  of  this  copula  is  justified  by  its  cumulative  distribution  function  easy  to  implement  and   the   presence   of   lower   and   upper   tails   dependence,   enabling   to   model   non   linear  dependence  of  rare  events  (see  Appendix  3).  

Both  copula  and  marginal  distributions  are  then  calibrated  on  historical  data  with  a  maximum  likelihood   algorithm,   and   the  Prior  Market  Distribution   can   finally   be   built  with  Monte-­‐Carlo  simulations,  and  stored  in  a  matrix  ! ∈ℳ!"#  where  J  is  the  number  of  simulations  and  N  the  number  of  assets.    

 Quant  Awards  2013  –  September  2013       4  

• The  Views  Distribution  is  another  input  of  the  model,  expressed  as  a  combination  of  two  matrix:    -­‐ Two   vectors   contain   the   lower   and   upper   bounds   of   the   forecasts   on   returns.   For  example,  if  we  consider  the  vectors  lb=(0%,  1%)  et  ub=(2%,  4%),  the  investor  expresses  2  forecasts,   in  the  ranges  [0%: 2%]  and    [1%: 4%].  The  assets  are  assumed  to  follow  uniform  distributions   between   these   bounds.   In   this   research   paper,   the   forecasts   are   voluntarily  subjective  views  on  the  market  and  are  not  the  result  of  any  model.      -­‐ A   “Pick”   matrix  ! ∈ℳ!"#  where   K   is   the   number   of   views   and   N   the   number   of  assets,   giving   the  weights  on   the  assets  affected  by   the  kth   view  (lower  and  upper  bounds  vectors  above)  

 

• The  Posterior  Market  Distribution:  we  finally  define  a  vector  ! ∈ℳ!"!  with  weights  between   0%   and   100%,  whether  we   trust   or   not   our   forecasts.   0%  would  mean   a  model  fully  relying  on  the  Views  Distribution.    We   estimate   this   vector   using   confidence   intervals.   More   precisely,   we   deduce   the  confidence   level  of  each   forecast   from  its  variance,  which   leads  to  build  a  volatility  model.  We  choose  the  heteroscedastic  model  E-­‐Garch(1,1),  described  by  the  following  equations:    !! = ! + !!  !ù  !! = !!!!    

log !!! = ! + ! log !!!!! + ![!!!!!!!!

− !!!!!!!!!

] + !!!!!!!!!

 

where  !!  and  !!!  are  the  return  and  the  variance  of  the  considered  asset.  Once   we   have   calibrated   each   view,   we   estimate   the   variance   of   the   residuals  !!!which  follow  a  Gaussian  distribution,  and  the  confidence  level  from  the  confidence  interval:    ∀! ∈ 1,!  !! = !"#$(!! − ! ≦ !! ≦ !! + !)  where  !!  and  !!  are  respectively  the  return  and  the  forecast  for  each  asset.    We  finally  apply  this  for  a  Gaussian  distribution:    !! = !!,!!! !! + ! − !!,!!! !! − !  

where  !!,!!!  is  the  Gaussian  cumulative  distribution  function,  with  parameters  (0  ,  !!!),  and  

T  a  defined  threshold.  

   We   finally  build   the   final  Posterior  Market  Distribution,  which   is  a  weighted  average  of   the  ‘Prior  Market  Distribution’  and  ‘Views  Distribution’,  according  to  the  following  formula:    !!,! ≡ !!!! !!,! + (1 − !!)!!(!!,!)    

where  !!  is   Prior   cumulative   distribution   function,  !!  the   Views   cumulative   distribution  function,  !!  the  confidence  in  the  kth  view,  and  W  the  sorted  returns  of    the  J  simulations.  

 

 

 

 Quant  Awards  2013  –  September  2013       5  

OPTIMIZATION  CRITERIA:  THE  OMEGA  RATIO    

Once  we’ve  computed  the  Posterior  Market  Distribution,  we  have  J  simulations  of  the  possible  evolution   of   the   market,   according   to   both   quantitative   (‘Prior’)   and   fundamental   (‘Views’)  analysis.  We  choose  the  parametric  ratio  Omega  as  criteria  to  optimize  our  portfolio  and  select  allocation  weights:  

 

Ω H =1 − F x dx!!

!

F x dx!!!

 

 

Where  F  is  the  CDF  of  the  portfolio  returns,  and  H  the  targeted  return.  

Example  of  Cumulative  Distribution  Function:  

 

 

 

 

 

 

 

 

 

 

 

The   Omega   ratio,   introduced   by   Keating   and   Shadwick,   is   interpreted   as   the   ratio   of  performances   over   the   threshold   H   divided   by   the   performances   under   the   threshold.   If   the  average   return   of   the   portfolio   is   higher   than   the   defined   threshold,   the   ratio   will   tend   to  increase.  Consequently,  we  optimize  our  portfolio  increasing  this  ratio  as  much  as  possible.  

 

The   strength   of   this   ratio   is   to   take   into   account  more   than   two  moments   of   the   distribution,  since  it  is  derived  from  the  CDF.  This  risk  measure  is  consequently  more  precise  than  the  Sharpe  Ratio.  

 

From   an   algorithm   point   of   view,   we   are   facing   a   multiple   parameter   optimization   problem,  which  is  solved  by  a  simplex  method:  the  Nelder-­‐Mead  optimization  technique.  

 

   

 Quant  Awards  2013  –  September  2013       6  

INPUTS  OF  THE  MODEL    

• To  build   the  Prior  Market  Distribution  we  use   the  daily   closing  prices  of   cross  asset   single  stocks  and  index,  between  14/11/2007  and  12/11/2011.      

   

• The   Views   Distribution   is   chosen   voluntarily   subjective   but   could   be   also   the   result   of  econometric  forecasts,  or  any  external  model.  The  lower  and  upper  bounds  are  the  same  for  each  forecast  but  we  consider  6  different  economic  scenarios,  implying  different  weights  on  the  11  assets:    

   Pick  Matrix:  

 

!""#$ %#&'#" ()$)"$&#)*+,'-.#&!"#$ %&'()*+),-)./%0++)1*)234)/).&.)56.,57)078 *+,-1*.2504.09- %":!";).:<"#=:>)0?@A"BCD?)9:DB<EB<F)-<E=:CBC<# +G.507',-)HI>: %&'()*+),-).5-J)K/HIJ)/).&.)56.,57)078 *+,-KJ+2504LDAF LDAF)%=AACD? L&+8%+70?M<#BN<?B)O:"F<)E:<FCB %":!";),-)!D:;)C?MP)L:"F< +G!!&592074QHIIGCO$)JC<AF)!:<FCB %":!";)LADR"A)GJ +G*LGJ82074QHII6*)8<RB)2ADE"A)'S4 T9*)6+*0Q T9*9!&32504

6*)8<RB)2$":F)'S4 T9*)6*%0Q T9*9.&.2504

LADR"A).CB"?# 8T)LADR"A).CB"?# 8T.0.(725046N<:OC?O)*":U<B)6V=CBC<# *-!0)6N<:OC?O)*":U<B# *-6*W'32504!DNNDFCBC<# 8T),%-)X)!DNNDFCBC<#).5 8T,%-.5

!"#$% &#'()*+#,-. /00()*+#,-.

1234 !" #!"/!5 !" $!"567! !" #$"681 !" %!"9:1 !" &!";<*=(>"*?42).@ !" &!";<*=(>"*?A#$2A@ !" &!"=<*;B,C"C(3 !" %!";<*;B,C"C(3 !" %!"1#DD#.C"C(3 !" &!"8#A. !" &!"

!"#$%&'()!"#$%&'("%)*+,)(-&. /0 /0 /0 /0 /0 /0 /0 1/0 2/0 1/0 /0,($3'$(456 /0 /0 /0 /0 /0 /0 /0 2/0 2/0 7/0 /08)9"3 /0 /0 /0 /0 7/0 /0 /0 2/0 2/0 /0 /0:&%;"%&'(" /0 /0 /0 1/0 1/0 /0 /0 /0 2/0 /0 /0:&%;#$%&'(" <=0 /0 7=0 /0 /0 /0 1/0 /0 /0 /0 1/0>94966'(" /0 =/0 7/0 /0 /0 /0 /0 /0 /0 /0 1/0

 Quant  Awards  2013  –  September  2013       7  

 

 • Finally,   the   confidence   vector   is   the   output   of   the   E-­‐Garch(1,1)   model,   using   confidence  intervals  at  99%:    

 

 

RESULTS    

To  measure  the  improvement  of  the  copula  approach,  we  define  as  benchmark  the  output  of  the  Black-­‐Littermann  model,  assuming  similar  views  and  economic  scenarios.  

The  model  is  run  for  100,000  simulations  and  the  ratio  Omega  is  optimized  with  a  threshold  of  1%.  We  finally  get  the  following  basket  allocation:  

 

 

The  backtesting  of  the  model  is  made  out  of  the  calibration  window,  between  December  2011  and  September  2012.  The  portfolio  weights  are  kept  unchanged  during  this  period.  We  get  a  higher  return  from  5,43%  to  8,39%  whilst  the  volatility  and  the  Omega  ratio  of  our  portfolio  strongly  decrease.  

 

 

!"#$#%&"'(")$*+&# !"#$%&'#('!"#$%&'("%)*+,)(-&. /0123,($4'$(567 80/239):"4 /0;23<&%="%&'(" 10>23<&%=#$%&'(" ?10?23@:5:77'(" 10A23

!""#$%&'#( )"%$*+,'&&-./%((+)-($0/%.* 1#23"%+42'('#(+!22.#%$0 52.-%61%70 !" #" !"859 $" !#" %!#"9:;5 &" ##" %'":<+1=>?:9 $" (" %("@A+1=>?:9 '$" &" !!">B+,41!, !)" '" !*">B+@!=? !(" #" !&"?B+>C8:9:>5 $" (" %(">B+>C8:9:>5 #" #(" %#)"14BB4 !" +" %*"<4,? $" #," %#,"9#&%" #$$" #$$"

!""#$%&'#()*#+," -,&./()#()&0,)1,/'#+

!((.%"'2,+)3#"%&'"'&4 5*,6%)-%&'#) 7#8,9&)/,&./( :'60,/)/,&./(

!"#$%&'())*+,#--&!*-$.,#+% /0123 40513 60227 860253 602439:;<"#&=;(-(:-&>;;+:#$. 702?3 20463 6066/ 8504@3 50713

 Quant  Awards  2013  –  September  2013       8  

CONCLUSION    

This   research   paper   examines   whether   the   modeling   of   non-­‐linear   term   structure  dependences  between  assets  can  reduce  the  systematic  risk  of  a  portfolio,  without  affecting  its   return.   The   model   presented   outperforms   the   Black-­‐Littermann   benchmark,   both   on  return  and  risk  measures.  

Compared  to  the  Black-­‐Littermann  model,  we  mainly  highlight  three  improvements:  

-­‐ The  use  of   copulas   enables   to  model   the  Risk-­‐on  Risk-­‐off   phenomenon:   in   case  of   high  stressed  scenarios,  the  fundamental  stock  analysis  is  less  effective  and  a  copula  opinion  pooling  would  be  more  reliable.    

-­‐ The  simulation  of  assets  based  on  marginal  distributions  combined  with  the  Monte-­‐Carlo  method  widens   the  scope,   letting   the  possibility   to   choose  a  different  diffusion  process  for  each  asset.    

-­‐ The   confidence   matrix   is   one   of   the   hardest   parameter   to   estimate   in   the   Black  Littermann  model.  We  propose  a  way  to  compute  it,  based  on  a  volatility  model,  letting  to  the  investor  the  forecasts  on  returns  as  only  subjective  inputs.    

The  model  presented  enables  to  capture  different  Market  phenomenon  as  the  leptokurticity  of   returns,   the   non-­‐linearity   of   correlations,   the   heteroscedasticity   and   asymmetry   of   the  volatility,   whilst   we   let   the   possibility   to   take   into   account   any   fundamental   analysis   on  stocks.   For   practical   reasons,   we’ve   chosen   to   apply   the   model   in   a   very   specific   case  (Student  Copula,  Heston  and  EGARCH  diffusion  process…),  but   the  model  can  be  seen  as  a  generic  methodology  with  further  possible  improvements  (implementation  of  Jump  process,  calibration   on   Market   prices,   discussion   about   the   copula…).   In   any   case,   we   suggest  measuring  the  effectiveness  of  a  model  with  a  risk  ratio  depending  on  the  whole  distribution  of  returns,  such  as  the  Omega  ratio  we’ve  presented.  

 

 

   

 Quant  Awards  2013  –  September  2013       9  

APPENDIX  

 

Appendix  1:  

 

The  cumulative  distribution   function  of   the  “Posterior  Market  Distribution”   is   the  weighted  average  between  the  “Prior  Market  Distribution”  and  the  “View  Distribution”.    

 

 

 

 

 

 

 

 

 

 

 

 

 

 

Appendix  2:  

 

A  copula  is  a  multivariate  cumulative  distribution  function  whose  marginal  distributions  are  uniform   random   variables   on   [0,1].   If   C   is   a   copula   on  ℝ!,   we   can   find   a   random   vector  (U!,… ,U!)  such  as:  

 

P U! ≤ u!,… ,U! ≤ u! = C(u!,… , u!)  

P U! ≤ u = u  for  all  i ∈ [0,1]  

 

A   copula   can  be  determined  by   the  previous  definition,   or   using  pre-­‐existing  multivariate  distribution.   In  this   latter  case,  we  use  the  Skal’s  theorem,  explaining  the  link  between  the  copula   C   and   the   multivariate   distribution   F,   depending   on   its   marginal   univariate  distributions  !!  and  !!.  

 

Sklar’s  Theorem:  

 

 Quant  Awards  2013  –  September  2013       10  

Let   F   be   a   bivariate   distribution   with   marginal   distribution  !!  and  !! .   The   copula   C  associated  is  defined  by:  

C u!, u!  

= C F! x! , F! x!  

= F F!!! u! , F!!! u!  

= F(x!, x!)  

C  is  unique  when  the  margins  !!  and  !!  are  continuous.  

 

Appendix  3:  

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

Source  :   «  Nonlinear   Term   Structure   Dependence:   Copula   Functions,   Empirics,   and   Risk  Implications”  Markus  Junker,  Alex  Szimayer,  Niklas  Wagner    

 

Lower  Tail  dependence:    

A  copula  C  has  a  lower  tail  dependence  if  :  !! = lim!→!!!(!,!)!

 exists  and  !! ∈]0,1]    

Upper  Tail  dependence:  

A  copula  C  has  a  upper  tail  dependence  if  :  !! = lim!→!!!!!!!!(!,!)

!!!  exists  and  !! ∈]0,1]    

 

 Quant  Awards  2013  –  September  2013       11  

REFERENCES    

• Markus  Junker,  Alex  Szimayer,  Niklas  Wagner,  2004,    “Nonlinear  Term  Structure  

Dependence:  Copula  Functions,  Empirics,  and  Risk  Implications”  

• Attilio  Meucci,  2005,  “Beyond  Black-­‐Litterman:  Views  on  Non-­‐Normal  Markets”  

• Attilio  Meucci,  2006,  “Beyond  Black-­‐Litterman  in  Practice:  a  Five-­‐Step  Recipe  to  Input  

Views  on  non-­‐Normal  Markets”  

• Michael  Stein,  2008,  Copula  Opinion  Pooling  in  Asset  Allocation  

• Attilio  Meucci,  2011,  A  New  Breed  of  Copulas  for  Risk  and  Portfolio  Management  

• Arthur  Charpentier,  2010,  Copules  et  risques  corrélés,  Journées  d’Études  Statistique  

• Fischer   Black   and   Rober   Litterman,   1992,   “Global   Portfolio   Optimization”,   Financial  

Analysts  Journal  

• Con  Keating  and  William  F.  Shadwick,  2002,  An  Introduction  to  Omega  

• S.J.  Kane  and  M.C.  Bartholomew-­‐Biggs,  M.  Cross  and  M.  Dewar,  “OPTIMIZING  OMEGA”  

• Joanna  Gatz,  2007,  Properties  and  Applications  of  the  Student  T  Copula