Lecture 60 - Coverage (Example in Multi Objective Decision Making)

IIT Kanpur July 201
20 followers
Jun 11, 2024
[Music]
welcome back my dear friends a very good
morning good afternoon and good evening
to all of you for whichever part of the
world you're listening to these lectures
and which what time frame it is or time
of the day when you are listening
as you know this is the last class for
multired decision making the course
under npil Muk uh it has been spread
over 12 weeks with 60 lectures and we
are in the 60th
lecture we have covered many different
topics in multicriteria decision making
first part main chunk was multiobjective
then utility Theory and then non
panametric
methods then
uh the last one would be an example this
would only give you a feel that how the
problems could be solved if you recap we
have discussed about linear programming
integer programming quadratic
programming then considered
multiobjective cases variance of them
utility Theory and all its different
shades very simple concept I didn't not
not goone into the proofs then you have
considered multivariate U Theory again
as I keep saying very briefly the
non-parametric
methods Electra method EP electr tops is
V or
hpda the last two lectures were in
decision tree diagram and only decision
tree diagram if you remember I did
mention that the diagrams was quite
detailed so I just covered it using the
word document not the PPD slides
obviously everything will be shared like
for all
lectures all the um slides have been
shared each week you get it and
obviously based on that and the lectures
which you hear you take the assignment
which is 10 questions for each week and
then the end s so this is the last
lecture of multired decision making and
we will be considering a problem here
and different without going to the
details I discuss the
problem details means not going to the
detail Solutions is example in
multiobjective decision making and the
background is about portfolio
optimization even though if you remember
we have discussed problems of of
choosing a house and apartment in
different cities depending on the price
the safety features uh the location near
to the bus stop or Transportation
facilities being nearby how close are
the school how close are the hospitals
how close of the market on the
markets whether it's a gated community
whether your relatives are there whether
your friends are
there what is the price how many bedroom
apartment is
there then we consider the concept of
idea of trying to choose an education
Institute based on rigor the placement
facilities The Internship facilities the
cost how far it is from the
home and all these
things then we considered the idea of
trying to buy a car depending on safety
features power cost
maintenance resale
value uh luggage
space the style
Factor all these
things and there were other different
problems
also but production problems were there
the problems of decision tree analysis
of trying to build up a factory to sell
a product
or build up a factory to sell a
moped here the problem is on from the
financial perspective so the problem is
that you have a port you want to
formulate a portfolio portfolio is a
basket of different
assets and we have initially a wealth
the W to invest his n number of
assets what is needed to be found out is
basically the total percentage of the
total amount which you can invest in
these n number of assets the problem
could also have been looked at from the
point of view of what is the total
Quantum of number of assets you're going
to py or what is the total value of the
money you going to invest in trying to
buy those number of assets the idea
Remains the Same why I'll just very
briefly mention it so
consider you have total total amount is
say for example
a and the assets are marked 1 to capital
N so the numbers you buy in these assets
are N1 to
NN the prices are S1 to
SN which means W1 till W which is the
proportions would be found out
accordingly consider a amount has been
totally invested in all these assets
starting from 1 to n no amount of money
is
left which means W1 which I'll mention
here would be equal to N1 into
S1 summation of Ni I SI which is equal
to a similarly you can find out W2 till
W capital N so w are the weights of the
proportions of the total wealth invested
in the asset V is the variance coent
Matrix for the returns for the an asset
what is return if you remember we have
found out the concept of Total return
and rate of return we are taking the
rate of return and we
discuss we did discuss about the ideas
that utility theory that if the
quadratic util function was applicable
then the rate of return would be normal
and vice
versa m is the vector of returns of n
assets n is the number of scripts or
assets we only considered the
scripts return star or R star risk star
or Sigma square star
are the respective predefined values of
the returns set by the investor from the
portfolio and the second value being
predefined value of the risk set by the
investor from the portfolio so
technically
rstar I will denote I'll use a different
color consider blue so I'll use RP
star is this value of return and this
value of risk would be given by Sigma
squ P start where p is not probability p
is the
portfolio so the objective can be Min
either minimize the risk or maximize the
return in that case it are single
objective functions now I combine them
when I combine them you minimize the
risk maximize the return subject to
conations being return for that
portfolio has to be greater than R star
because that's what the investor wants
and the second constraint being risk
being less than equal to Sigma Square P
star obviously that has to be return has
to be more than the stut value and risk
has to be less than the sttip value less
than equal to or greater than equal to
obviously sum of the weights is one
weights are between 0 and one here is
one important point to remember people
who do Finance remember there is the
concept of Short Selling in which case
weights can be negative also which you
are not considering here so here to go
into the detail of the nomenclature and
total number of stocks 1 to n t the time
period why time period is important and
come to that W are the wids so in the
vector format it is W1 to WN each of
these stocks can be invested within a
particular bound the minimum bound is w
mean maximum is wi Max so if the minimum
and the maximum are between 0 and one
obviously all of the values of would be
between 0 and one if I say the second
stock cannot have an investment of more
than 30% in that case
W2 max value is 20% it's 02 similarly if
I say the fifth stock cannot have
anything less than
uh minimum value is required is 50% in
that case the value of the minimum wi
would be for that particular stock will
be 0.5 RI is the random variable
representing the rate of return for the
I stock which I have said and RP star is
return threshold required by the
investor RIT is the realization of
random variable RI during time period T
Sigma is the standard devision of the I
stock rig is the correlation Sigma
square star p is the threshold value
which I mentioned SD is the standard
deviation a is the Matrix of returns and
B is basically the minimum portfolio
which is to be required I'm not going
into the details of the these but I'll
just highlight the important points
remember that in these multiobjective
problems it can be that you want to buy
some stock at each equal interval of
time and sell also stocks so obviously
buying and selling will entail some cost
because you have to help get the help of
the broker consider that so obviously
the some percentage amount has to be
spent both for buying and selling
those values which I I considered is
basically w+ being an N Vector of stock
purchased and W minus w+ and W minus are
vector plus means purchased minus being
sold and W not again the vector is
basically the initial stock value Q
which is basically denoted as V in when
we are doing the actual understanding of
the General formulation the problem is
basically the variance covariance Matrix
why variance coari Matrix the principal
diagonal of the variances and of
diagonal element are the variance Co
varience values for
in now the first problem I'll only
discuss the problem again I said not
going into the solution technique only
discuss the results the first problem is
I'll use two different colors for the
minimization red maximization
green for the maximization problem it is
maximizing the
return return is already RI into wi
which I already mentioned RI is the
return for the I stock W is the weight
corresponding to the stock
minus P transpose P this is the because
you have to multiply uh row vectors and
column vectors and considering that
compatibility of row and column
multiplication is there in Matrix that
so that's why it is known as P prime or
P transpose P transpose into w+ which is
w+ is which are being purchased so
obviously there's a price for that and W
minus other those are sold that is also
price for that price means for the
broker similarly so you want to maximize
this and from the minimization one you
want to minimize the quadratic function
quadratic function is basically if you
if you go into it it is double
summation of w WJ Sigma i
j W and WJ are the weights Sigma J is
the covidence values for the I and J
what are the constraint so I'll let me
start from the the simplest one some of
the weights I should use a different
color so let me use
the uh valid color sum of the weights is
one we
know the weights are between minimum and
maximum we have
assumed the fourth one if I'm going from
uh from the last Assumption of the
constraint this is like a rate flow
equation
W basically which you started with w+
which you bought W minus you sold with a
plus and minus sign corresponding to the
inflow and outflow this basically is the
initial
value the variance coars Matrix is less
than equal to Sigma s star p as I have
mention based on the Investor's set risk
value and summation of R iwi being
greater than RP is also the return value
set by the investor so one is and the
first one is
basically the the constraint based on
the values of the individual weights so
here a is
basically uh a matrix based on the
weights parameters W's are
the weights corresponding to the amount
which you want to invest in the N number
of assets B is basically the right hand
side they can be formulated accordingly
so main being the second one the third
one the fifth one and fourth and first
have been added in order to bring more
realistic flavor to the problem now the
stocks taken are BC Dax Dow Jones Hong
Kong STI Singapore for three year time
period from 2012 to 2015 January of 2012
to 2015 January and we take Windows now
if you remember I mentioned the time
frame so what why we are taking Windows
is basically to bootstrap bootstrap
means that when you're tying your laces
of the boots so you basically keep tying
them from the lowest level and then
tighten up
so the the the shoe which you are or the
shoes which you wearing are really tight
around your cuff muscles and your
ankles so boot stepping is a method of
sampling which has different varieties
of way of trying to handle it I'm not
going to the details that would fall
under the Ambit of of statistics and
simulation so in order to do
bootstrapping I consider the total set
of uh length of the um time frame for
the stocks is 3 years and I break that
into windows so this is the total time
frame I use the violet color
here and each length of the window which
I take I'll use the blue
color and there is and I keep Shifting
the windows air is shifted but there's a
overlap this
overlap depending on the number of days
I here I'm marking as light blue so this
o is the overlap L is the length of the
window and total length of time is T So
generally the length is taken as about
one week or 2 is maximum when the
overlaps are taken to about 3 to 4 days
so you keep repeating the overlap and
based on the average value what you want
to use the standard deviation of those
windows you find out such number of
times of the windows have been
formulated and then bootstrap to find
out the average values so as I said
there's a whole bunch of studies here
the window size is
150 signifies a we have taken in 150
days which signifies 6 month seasonality
in the trend as I mentioned it was 7
days or 14 days here we have taken 150
we repeat 25 number of times and merge
it together in order to start the
bootstrap bootstrap numbers are taken as
500 1,000 2,000 and it can be
extended so when we do the rebalancing
with the transaction cost so the first
problem the risk return profile as it
should be so you remember there was a uh
two functions one was trying to maximize
the return based on the selling and
buying
cost and one was the risk variance so
when I plot it if you remember if you
have done simple Finance even
portfolio
analysis the graph the risk return is
like this efficient Frontier this
exactly what we are getting I'll use the
yellow color so not
to sub like make the blue dots vanish so
the set of colors obviously it should be
a if you do more number of series
simulation it will be a smoother curve
but considering the time frame for each
simulation we took number of simulations
accordingly and this is the portfolio
risk and return for BSC if you remember
we had taken Hong Kong Singapore Dow
Jones
accordingly the rebalancing part based
on the weights as they keep changing for
whole set of
data this is how the rebalancing is done
for the first problem if you see I don't
have the coloring scheme here but if you
see the dark bot color slowly V becomes
less and less as you increase the
combinations of risk and return and the
yellow one which is basically would
slowly start dominating as you increase
the risk and return Red also comes into
picture Brown is almost stable and so on
and so so if you had the coloring scheme
of the stocks you can easily find out
which stock is being slowly removed
which stock is coming to the picture
depending on on the combinations of the
risk and return for the first problem so
if I have the portfolio risk and return
for the rebalancing transaction cost so
this curve which you
look
actually if you remember the par optimal
Frontiers the one which you have done
and if you
remember highlight
highlighted that depending on
maximization I'll use four different
colors both uh okay by by the way this
is objective one o1 is objective one O2
is objective two I'm trying to plot this
is a two dimensional one it can be done
for three dimensional and so on and so
forth if both are being
maximized we'll try to basically move to
the top right hand corner in the first
quadrant top both are being
minimized you will move in the left
bottom corner in the third quadrant
similarly one being maximized another
being minimized the combinations of the
points should
point and this I should use the not the
green color my
apologies this we have already done so
you can
basically find it out now considering
the Maxum return minimum risk so
rebalancing is not there we consider the
return the second one mean return being
greater than RP star risk being less
than Sigma Square P some of the W is one
and here I want want the return average
return of the weights is
increasing return should be such it is
greater than
one so again without going to the
details this is the if I consider the
Hong Kong I'm not considering the BC so
I'm trying to basically
utilize the model for all the in scripts
but giving a of the diagram so then
return to risk again similarly like the
risk return
profile here is the asset weights
distribution it looks quite Jagger but
this is what we got so basically this is
a part where one of the stock really
dominates and gets overwhelmed with the
other one which is shown in a creamish
yellow then the risk return profile
considering the parent optimal Frontier
is exactly this if you see this is F1 F2
the two objective functions we have
considered now we come to the max return
over threshold value so the minimum so
all are by objective this is the
minimization problem I should Mark in
red this is the maximum problem I should
Mark in
green again the constraints are I'll use
a different color violet greater than RP
star so here risk is not being brought
into the picture less than equal to
Sigma square star p is not there in the
picture some of the weight is one weight
are between zero between minimum and
maximum and the returns considering is
increasing are such it's greater than
equal to one that means if RIS are
negative so it will be one minus that so
it may turn out to be on the
downward here I do it for the Dax with
the German again the risk return profile
with out the Scatter Plots on the Lower
Side the general Trend looks like thisis
return
profile the minimum threshold uh return
for the next one considering the the
weight distribution looks like this I'm
not explaining I'm just showing you the
results and again the F1 F2 which is the
objective function for the third problem
concerning the paror front is this
obviously you can have different paror
front but I'm not going into all those
details and and so with this problem we
had considered three different
problems with rebalancing simple problem
and in the max deviation
part so in the risk was always the same
the standard deviation Square in all the
three problems and we used all these um
stocks using BSC Dax hansing
Singapore Dow Jones German and we showed
they're working for a time frame of 3
years and how we took the the
overlapping time Windows use bootstrap
the results were
shown the next problem we considered is
basically an anat type of problem where
we are going to
consider the conditional value at
risk and not variance and here is one of
the problem formulation which you have
already discussed in m objective
case the nomenclature and the variables
Remains the Same rgi is the loss return
we are taking the loss return now what
is the loss return if you if you can
recollect RI which is
basically the in the numerator you have
the return you have the value of today
minus the value of one day before
divided by the value of of one day
before so that can be positive and
negative depending on the numerator
value
so return would be the loss distribution
will be the case if you take only the
negative return positive return would
give you the profit distribution so
that's why the loss return distribution
rgi the loss distribution for the J
stock for the I day R GI is the nominal
value of the rust return Sigma Sigma
value I'm not reading the suffix is the
vector of variances xjs of the weights
which was w n is the number of stocks T
is the time period Alpha Beta 1 beta 2
are the corresponding values for the
value at risk which is a very useful
measure in in finance and beta 1 beta 2
are the corresponding constants
probability value we will
take gamma DL DL is basically the loss
distribution would be for the C value
and Lambda 1 Lambda 2 are the weights
which will assign to the C and RP if you
remember RP means the return of the
portfolio if you remember in one of the
multip subjective case we did consider
Lambda 1 Lambda 2 and based on that we
solve the
problem considering so these assum this
nomenclature from 11 to 19 which are
shown in slide are based on the problem
type solution methodology we take so
this is considered which is not part of
the course I'll just mention the name is
known as robust optimization I'm not
going to consider that for further
discussion so the first problem is here
we minimize two objectives take weights
of Lambda 1 to C which is one risk and
Lambda 2
to return now why there a minus sign
because we are minimizing so if you
maximize it will be plus and minus so we
are minimizing the risk and obviously
trying to consider the return also the
constraints are weights add up to one
weights are X's Lambda values add up to
one because there are two Lambda values
the weights can be between minimum
maximum but here they are
unbounded the first case if I consider
first case means the first con if I
consider without the probability that
means beta 1 is not there and second one
beta 2 is not there this means the
conditional value of risk is less than
some stipulated value and the second
part here which I'm highlighting using
the red color here is basically for the
case when the returns are considered so
returns were greater than some
stipulated value similar risk were less
than the stipulated value in the first
problem here risk is conditional value
risk so obviously risk has to be less
than the stipulate
value so these are the chance constraint
uncertain parameters so which I will
show but not discuss in details in the
problem so these are the chance
constraints the first and the second one
with the probabilities are and the
uncertain parameters are given what are
the un certain parameter under certain
parameters are the returns because they
are dependent on the past
data the data is taken from the
Australian Brazil France which is cak
Dax Germany foot England from Canada
Hansen Hong Kong Argentina Nik Japan NC
India NYC USA and Singapore we took it
from 2000 to 2010 I'm not going to the
details which is total of 2,840 days and
the summary statis statistics and the
most strapping concept were considered
these are the summary statistics which
means mean median standard division CES
qess medium minimum maximum value for
all the different indices and again the
values can be calculated very simply
these are the density
plots these are the extreme value
distributions depending of the loss
distribution which which was considered
here this is for the Dax only similarly
you can on the negative distribution
loss distribution for all the other
indices which I have not shown I'm just
showing one sample these are the data
description and the QQ plots theoretical
standard normal with the empirical
distribution being done and then again
the same concept of of Shifting window
so the windows which I'm marking in in
red are
L and The O Part which was overlapping
this is considered as blue which is s
total time frame is T same concept so we
take three window size one day shifting
two window size one day shifting one
window size one day shifting actually
there if you remember you have taken 150
days so it can vary but generally
depending on the problem we would
suggest for finance to be about 2 week
window with three or 4 days overlapping
this is Shifting given shifts these are
the data statistics for used in model
one and model two model one model two
which I
shown based on that for all the
different scripts and these are the 3D
Scatter Plots here why 3D because if
you're looking you have now v c value
which is a risk VAR value which is also
risk and return which I'll use a
different color green is positive so you
want to increase return decrease bar and
decrease C so if I look at from this
direction so obviously I would be
considering from this left hand side so
consider the room where I am giving the
lecture if somebody is seeing from the
left hand side so and the axis which is
coming towards me from your side is X
which is going onto the right is y and
the one which is going top is Zed if
somebody is looking from the left hand
side obviously it will be X and
Zed being understood because why you
cannot see it's going to the depth from
that left hand side if I'm looking from
this side I'll be only I would not be
noticing X I'll be noticing Y and uh Zed
and Y similarly in different direction
as you look you'll have the Feld
accordingly so these are the ca return
and V plots and I have taken different
levels of probabilities which are
mentioned on the top as I said I won't
go into the details depending on
different values of betas the values are
given accordingly
so I'll just go through it and they give
you a very good feel that how
multiobjective case can be analyzed for
different problems but this problem I
had solved so it is much easier and with
confidence I can share the results here
so with this I will end the uh the whole
set of series of 60 lectures for
multicrit decision making and I'm sure
it would be helpful for many of the
students for any queries for people who
are doing this course please get in
touch with the Tas and for any queries
if you have for further studies or any
other details you want to understand you
can most welcome write to me over my
email and you can find my email in my
homepage thank you very much and best of
luck to all of
[Music]
you
[Music]
