The following content is
provided under a Creative Commons license. Your support will help
MIT OpenCourseWare continue to offer high quality
educational resources for free. To make a donation or
view additional materials from hundreds of MIT courses,
visit MIT OpenCourseWare at ocw.mit.edu. KENNETH ABBOTT: As I said,
my name is Ken Abbott. I'm the operating officer
for Firm Risk Management at Morgan Stanley, which means
I'm the everything else guy. I'm like the normal
stuff with a bar over it. The complement of normal--
I get all the odd stuff. I consider myself the
Harvey Keitel character. You know, the fixer? And so I get a lot of
interesting stuff to do. I've covered commodities,
I've covered fixed income, I've covered equities, I've
covered credit derivatives, I've covered mortgages. Now I'm also the
Chief Risk Officer for the buy side
of Morgan Stanley. The investment management
business and the private equity holdings that we have. And I look after
lot of that stuff and I sit on probably
40 different committees because it's become very,
very, very bureaucratic. But that's the way it goes. What I want to talk about today
is some of the core approaches we use to measure a risk
in a market risk setting. This is part of a larger course
I teach at a couple places. I'm a triple alum at
NYU-- no I'm a double alum and now I'm on their
faculty [INAUDIBLE]. I have a masters in economics
from their arts and sciences program. I have a masters in
statistics from Stern when Stern used to
have a stat program. And now I teach at Courant. I also teach at Claremont
and I teach at Baruch, part of that program. So I've been through
this material many times. So what I want to do is lay
the foundation for this notion that we call value at
risk, this idea of VaR. [INAUDIBLE] put this back on. Got it. I'll make it work. I'll talk about it from
a mathematical standpoint and from a statistical
standpoint, but also give you some of
the intuition behind what it is that we're trying to do
when we measure this thing. First, a couple words
about risk management. What is the risk do? 25 years ago, maybe three firms
had risk management groups. I was part of the
first risk management group at Bankers Trust in 1986. No one else had a risk
management group as far as I know. Market risk management really
came to be in the late '80s. Credit risk management
had obviously been around in large financial
institutions the whole time. So our job is to make
sure that management knows what's on the books. So step one is, what is the
risk profile of the firm? How do I make sure
that management is informed about this? So it requires two things. One, I have to know
what the risk profile is because I have to
know it in order to be able to communicate it. But the second thing, equally
important, particularly important for you
guys and girls, is that you need to
be able to express relatively complex
concepts in simple words and pretty pictures. All right? Chances are if you go
to work for big firm, your boss won't be a quant. My boss happens to have a
degree from Carnegie Mellon. He can count to 11
with his shoes on. His boss is a lawyer. His boss is the chairman. Commonly, the most senior people
are very, very intelligent, very, very articulate,
very, very learned. But not necessarily quants. Many of them have had a year
or two of calculus, maybe even linear algebra. You can't show them-- look,
when you and I chat and we talk about regression analysis, I
could say X transpose X inverse X transpose y. And those of you that have
taken a regression course think, ah, that's beta hat. And we can just stop it there. I can just put this form up
there and you may recognize it. I would have to spend 45
minutes explaining this to people on the top
floor because this is not what they're studying. So we can talk the
code amongst ourselves, but when we go outside our
little group-- getting bigger-- we have to make sure that we
can express ourselves clearly. That's done in clear,
effective prose, and in graphs. And I'll show you some of
that stuff as we go on. So step one, make
sure management knows what the risk profile is. Step two, protect the firm
against unacceptably large concentrations. This is the subjective part. I can know the risk,
but how big is big? How much is too much? How much is too concentrated? If I have $1 million of
sensitivity per basis point, that's a 1/100th of
1% move in a rate. Is that big? Is that small? How do I know how much? How much of a particular
stock issue should I own? How much of a bond issue? How much futures open interest? How big a limit should I
have on this type of risk? That's where intuition and
experience come into play. So that's the second
part of our job is to protect against
unacceptably large losses. So the third, no
surprises, you can liken the trading business--
it's taking calculated risks. Sometimes you're going to lose. Many times you're going to lose. In fact, if you win 51% of
the time, life is pretty good. So what you want to do
is make sure you have the right information so you
can estimate, if things get bad, how bad will they get? And to use that,
we leverage a lot of relatively simple notions
that we see in statistics. And so I should use a coloring
mask here, not a spotlight. We do a couple things. Just like the way when they talk
about the press in your course about journalism,
we can shine a light anywhere we want, and
we do all the time. You know what? I'm going to think about
this particular kind of risk. I'm going to point out that
this is really important. You need to pay attention to it. And then I could shade it. I can make it blue, I can make
a red, I can make it green. I'd say this is good, this
is bad, this is too big, this is too small,
this is perfectly fine. So that's just a little bit of
quick background on what we do. So I'm going to go through
as much of this as I can. I'm going to fly
through the first part and I want to hit these because
these are the ways that we actually estimate risk. Variance, covariance
as a quadratic form. Monte Carlo simulation,
the way I'll show you is based on a quadratic form. And historical simulation
is Monte Carlo simulation without the Monte Carlo part. It's using historical data. And I'll go through
that fairly quickly. Questions, comments? No? Excellent. Stop me-- look,
if any one of you doesn't understand something
I say, probably many of you don't understand it. I don't know you guys, so
I don't know what you know and what you don't know. So if there's a term that
comes up, you're not sure, just say, Ken, I
don't have a PhD. I work for a living. I make fun of academics. I know you work
for a living too. All right. There's a guy I tease
at Claremont [INAUDIBLE] in his class, I say, who is
this pointy-headed academic [INAUDIBLE]. Only kidding. All right, so I'm going to talk
about one-asset value at risk. First I'm going to introduce
the notion of value at risk. I'm going to talk
about one asset. I'm going to talk about
price-based instruments. We're going to go
into yield space, so we'll talk about
the conversions we have to do there. One thing I'll do after
this class is over, since I know I'm going to fly
through some of the material-- and since this is
MIT, I'm sure you're used to just flying
through material. And there's a lot of
this, the proof of which is left to the reader
as an exercise. I'm sure you get a
fair amount of that. I will give you papers. If you have questions, my
email is on the first page. I welcome your questions. I tell my students
that every year. I'm OK with you sending
me an email asking me for a reference, a
citation, something. I'm perfectly fine with that. Don't worry, oh, he's too busy. I'm fine. If you've got a question,
something is not clear, I've got access to
thousands of papers. And I've screened them. I've read thousands
of papers, I say this is a good one,
that's a waste of time. But I can give you
background material on regulation, on bond pricing,
on derivative algorithms. Let me know. I'm happy to provide that
at any point in time. You get that free
with your tuition. A couple of key metrics. I don't want to spend
too much time on this. Interest rate
exposure, how sensitive am I to changes in interest
rates, equity exposure, commodity exposure,
credit spread exposure. We'll talk about linearity,
we won't talk too much about regularity of cash flow. We won't really
get into that here. And we need to know correlation
across different asset classes. And I'll show you
what that means. At the heart of this
notion of value at risk is this idea of a
statistical order statistic. Who here has heard
of order statistics? All right, I'm going
to give you 30 seconds. The best simple description
of an order statistic. PROFESSOR: The
maximum or the minimum of a set of observations. KENNETH ABBOTT: All right? When we talk about
value at risk, I want to know the worst
1% of the outcomes. And what's cool about
order statistics is they're well established
in the literature. Pretty well understood. And so people are
familiar with it. Once we put our toe
into the academic water and we start talking
about this notion, there's a vast
body of literature that says this is
how this thing is. This is how it pays. This is what the
distribution looks like. And so we can
estimate these things. And so what we're looking
at in value at risk, if my distribution of
returns, how much I make. In particular, if I
look historically, I have a position. How much would this position
have earned me over the last n days, n weeks, n months. If I look at a frequency
distribution of that, I'm likely-- don't have to-- I'm
likely to get something that's symmetric. I'm likely to get
something that's unimodal. It may or may not
have fat tails. We'll talk about
that a little later. If my return distribution
were beautifully symmetric and beautifully normal and
independent, then the risk-- I could measure this
1% order statistic. What's the 1% likely worst
case outcome tomorrow? I might do that by integrating
the normal function from negative infinity--
for all intents and purposes five or six standard deviations. Anyway, from negative
infinity to negative 2.33 standard deviations. Why? Because the area under
the curve, that's 0.01. Now this is a one-sided
confidence interval as opposed to a two-sided
confidence integral. And this is one of these
things that as an undergrad you learn two-sided, and
then the first time someone shows you one sided you're
like, wait a minute. What is this? Than you say, oh, I get it. You're just looking at the area. I could build a gazillion
two-sided confidence intervals. One sided, it's got
to stop at one place. All right so this set
of outcomes-- and this is standardized-- this is in
standard deviation space-- negative infinity to 2.33. If I want 95%, or 5% likely
loss, so I could say, tomorrow there's a
5% chance my loss is going to be x or
greater, I would go to 1.645 standard deviations. Because the integral from
negative infinity to 1.645 standard deviations
is about 0.05. It's not just a good
idea, it's the law. Does that make sense? And again, I'm going to
say assuming the normal. That's like the
old economist joke, assume a can opener when
he's on a desert island. You guys don't know that one. I got lots of economics jokes. I'll tell them later on
maybe-- or after class. If I'm assuming normal
distribution, and that's what I'm going to
do, what I want to do is I'm going to set this thing
up in a normal distribution framework. Now doing this approach and
assuming normal distributions, I liken it to using Latin. Nobody really uses it
anymore but everything we do is based upon it. So that's our starting point. And it's really easy
to teach it this way and then we relax
the assumptions like so many things in life. I teach you the
strict case then we relax the assumptions to get
to the way it's done now. So this makes sense? All right. So let's get there. This is way oversimplified--
but let's say I have something like this. Who has taken
intermediate statistics? We have the notion
of stationarity that we talk about all the time. The mean and variance
constant is one simplistic way of thinking about this. Do you have a better way
for me to put that to them? Because you know what
their background would be. PROFESSOR: No. KENNETH ABBOTT: All right. Just, mean and
variance are constant. When I look at the
time series itself, the time series mean and
the time series variance are not constant. And there also could be other
time series stuff going on. There could be seasonality,
there could be autocorrelation. This looks something
like a random walk but it's not stationary. It's hard for me to draw
inference by looking at that alone. So we want to try
to predict what's going to happen in the
future, it's kind of hard. And the game, here, that we're
playing, is we want to know how much money do I need to
hold to support that position? Now, who here has taken
an accounting course? All right, word to
the wise-- there's two things I tell students
in quant finance programs. First of all, I know you have to
take a time series course-- I'm sure-- this is MIT. If you don't get a
time series course, get your money back because
you've got to take time series. Accounting is important. Accounting is important
because so much of what we do, the way
we think about things is predicated on the dollars. And you need to know how
the dollars are recorded. Quick aside. Balance sheet. I'll give you a 30 second
accounting lecture. Assets, what we own. Everything we own-- we
have stuff, it's assets. We came to that stuff
one of two ways. We either pay for it out of our
pocket, or we borrowed money. There's no third way. So everything we
own, we either paid for out of our pocket
or borrowed money. The amount we paid for out
of our pocket is the equity. The ratio of this to
this is called leverage among other things. All right? If I'm this company. I have this much
stuff and I bought it with this much debt,
and this much equity. Again, that's a gross
oversimplification. When this gets down to
zero, it's game over. Belly up. All right? Does that make sense? Now you've taken a
semester of accounting. No, only kidding. But it's actually important to
have a grip on how that works. Because what we
need to make sure of is that if we're going to take
this position and hold it, we need to make sure that
with some level of certainty-- every time we lose
money this gets reduced. When this goes down to
zero, I go bankrupt. So that's what
we're trying to do. We need to protect
this, and we do it by knowing how much of
this could move against us. Everybody with me? Anybody not with me? It's OK to have
questions, it really is. Excellent. All right, so if I do a
frequency distribution of this time series, I just
say, show me the frequency with which this thing shows. I get this thing,
it's kind of trimodal. It's all over the place. It doesn't tell me anything. If I look at the levels--
the frequency distribution, the relative frequency
distribution of the levels themselves, I don't get
a whole lot of intuition. If I go into return
space, which is either looking at the log
differences from day to day, or the percentage
changes from day to day, or perhaps the absolute
changes from day to day-- it varies from market to market. Oh, look, now we're
in familiar territory. So what I'm doing
here-- and this is why I started out with
a normal distribution because this thing is unimodal. It's more or less symmetric. Right? Now is it a perfect measure? No, because it's
probably got fat tails. So it's a little
bit like looking for the glasses you lost up on
67th Street down on 59th street because there's
more light there. But it's a starting point. So what I'm saying to you is
once I difference it-- no, I won't talk about [INAUDIBLE]. Once I difference
the timeshares, once I take the timeshares and
look at the percentage changes, and I look at the frequency
distribution of those changes, I get this which is
far more amenable. And I can draw
inference from that. I can say, ah, now if
this thing is normal, then I know that x%
of my observations will take place over here. Now I can start
drawing inferences. And a thing to
keep in mind here, one thing we do
constantly in statistics is we do parameter estimates. And remember, every time
you estimate something you estimate it with error. I think that maybe the
single most important thing I learned when I got
my statistics degree. Everything you estimate
you estimate with error. People do means,
they say, oh, it's x. No, that's the average and
that's an unbiased estimator, but guess what, there's
a huge amount of noise. And there's a
certain probability that you're wrong by x%. So every time we come up with a
number, when somebody tells me the risk is 10, that means
it's probably not 10,000, it's probably not zero. Just keep that in mind. Just sort of throw that in
on the side for nothing. All right, so when I take
the returns of this same time series, I get something
that's unimodal, symmetric, may or may not have fat tails. That has important
implications for whether or not my normal distribution
underestimates the amount of risk I'm taking. Everybody with me on
that more or less? Questions? Now would be the time. Good enough? He's lived this. All right. So once I have my time series
of returns, which I just plotted there, I can
gauge their dispersion with this measure
called variance. And you guys probably know this. Variance, the expected
value of x_i minus x bar-- I love these thick
chalks-- squared. And it's the sum of x_i minus
x bar squared over n minus 1. It's a measure of dispersion. Variance has its--
Now, I should say that this is sigma squared hat. Right? Estimate-- parameter estimate. Parameter. Parameter estimate. This is measured with error. Anybody here know what the
distribution of this is? Anyone? $5. Close. n chi-squared. Worth $2. Talk to me after class. It's a chi-squared distribution. What does that mean? That means that we know it
can't be 0 or less than 0. If you figure out a way to
get variances less than zero, let's talk. And it's got a long
right tail, but that's because this is squared. [INAUDIBLE] one
point can move it up. Anyway, once I
have my returns, I have a measure of the dispersion
of these returns called variance. I take the square
root of the variance, which is the standard
deviation, or the volatility. When I'm doing it
with a data set, I usually refer to it as
the standard deviation. When I'm referring to
the standard deviation of the distribution, I usually
call it the standard error. Is that a law or is that
just common parlance? PROFESSOR: Both. The standard error is
typically for something that's random, like an estimate. Whereas the standard deviation
is more like for sample-- KENNETH ABBOTT: Empirical. See, it's important because
when you first learn this, they don't tell you that. And they flip them
back and forth. And then when you take
the intermediate courses, they say, no, don't
use standard deviation when you mean standard error. And you'll get points off on
your exam for that, right? All right, so, the
standard deviation is the square root
of the variance, also called the volatility. In a normal distribution,
1% of the observations is outside of 2.33
standard deviations. For 95%, it's out past 1.64,
1.645 standard deviations. Now you're saying,
wait a minute, where did my 1.96 go that
I learned as an undergrad. Two-sided. So if I go from the mean
to 1.96 standard deviations on either side,
that encompasses 95% of the total area of the
integral from negative infinity to positive infinity. Everybody with me on that? Does that make sense? The two-sided versus one-sided. That's confused me. When I was your age,
it confused me a lot. But I got there. All right so this
is how we do it. Excel functions are VAR and--
you don't need to know that. All right, so in this case,
I estimating the variance of this particular time series. I took the standard deviation
by taking the square root of the variance. It's in percentages. When you do this, I tell
you, it's like physics, your units will screw
you up every time. What am I measuring? What are my units? I still make units mistakes. I want you to know that. And I'm in this
business 30 years. I still make units mistakes. Just like physics. I'm in percentage
change space, so I want to talk in terms
of percentage changes. The standard deviation is
1.8% of that time series I showed you. So 2.33 standard deviations
times the standard deviation is about 4.2%. What that says, given this data
set-- one time series-- I'm saying, I expect to
lose, on any given day, if I have that position,
99% of the time I'm going to lose 4.2% of it or less. Very important. Think about that. Is that clear? That's how I get there. I'm making a statement about
the probability of loss. I'm saying there's
a 1% probability, for that particular time
series-- which is-- all right? If this is my
historical data set and it's my only historical
data set, and I own this, tomorrow I may be 4.2%
lighter than I was today because the market
could move against me. And I'm 99% sure, if the
future's like the past, that my loss tomorrow is
going to be 4.2% or less. That's VaR. Simplest case, assuming
normal distribution, single asset, not fixed income. Yes, no? Questions, comments? AUDIENCE: Yes, [INAUDIBLE]
positive and [INAUDIBLE]. KENNETH ABBOTT: Yes, yes. Assuming my distribution
is symmetric. Now that's the right
assumption to point out. Because in the real world,
it may not be symmetric. And when we go into
historical simulation, we use empirical
distributions where we don't care if it's
symmetric because we're only looking at the downside. And whether I'm long
or short, I might care about the downside
or the putative upside. Because I'm short, and
I care about how much is going to move up. Make sense? That's the right
question to ask. Yes? AUDIENCE: [INAUDIBLE] if you're
doing it for upside as well? KENNETH ABBOTT: Yes. AUDIENCE: Could it
just be the same thing? KENNETH ABBOTT: Yes. In fact, in this case,
in what we're doing here of variance/covariance
or closed form VaR, it's for long or short. But getting your signs
right, I'm telling you, it's like physics. I still make that mistake. Yes? AUDIENCE: [INAUDIBLE] symmetric. Do you guys still use
this process to say, OK-- KENNETH ABBOTT: I use it
all the time as a heuristic. All right? Because let's say I've got-- and
that's a very good question-- let's say I've got five
years worth of data and I don't have time to
do an empirical estimate. It could be lopsided. If you tell me a two
standard deviation move is x, that means
something to me. Now, there's a
problem with that. And the problem is that
people extrapolate that. Sometimes people talk to me
and, oh, it's an eight standard deviation move. Eight standard deviation
moves don't happen. I don't think we've seen
an eight standard deviation move in the Cenozoic era. It just doesn't happen. Three standard deviation--
you will see a three standard deviation move once every
10,000 observations. Now, I learned this
the hard way by just, see how many times
do I have to do this? And then I looked it up in
the table, oh, I was right. When we oversimplify, and
start to talk about everything in terms of that
normal distribution, we really just lose
our grip on reality. But I use it as a
heuristic all the time. I'll do it even now,
and I know better. But I'll go, what's two
standard deviations? What's three
standard deviations? Because by and large-- and I
still do this, I get my data and I line it up and I do
frequency distributions. Hold on, I do this all
the time with my data. Is it symmetric? Is it fat tailed? Is it unimodal? So that's a very good question. Any other questions? AUDIENCE: [INAUDIBLE]
have we talked about the [? standard t ?]
distribution? PROFESSOR: We Introduced
it in the last lecture. And the problems set this
week does relate to that. KENNETH ABBOTT: All
right, perfect lead-in. So the statement I made,
it's 1% of the time I'd expect to lose more than
4.2 pesos on 100 peso position. That's my inferential statement. In fact, over the
same time period I lost 4.2% 1.5% of the time
instead of 1% of the time. What that tells me, what that
suggests to me, is my data set has fat tails. What that means is the
likelihood of a loss-- a simple way of thinking
about it [INAUDIBLE] care whether what that means
in a metaphysical sense, a way to interpret it. The likelihood of
a loss is greater than would be implied by
the normal distribution. All right? So when you hear
people say fat tails, generally, that's what
they're talking about. There are different ways you
could interpret that statement, but when somebody is talking
about a financial time series, it has fat tails. Roughly 3/4 of your
financial time series will have fat tails. They will also have
time series properties, they won't be true random walks. True random walks
says that I don't know whether it's going to go
up or down based on the data I have. The time series has no memory. When we start introducing
time series properties, which many financial time series
have, then there's seasonality, there's mean reversion,
there's all kinds of other stuff, other
ways that we have to think about modeling the data. Make sense? AUDIENCE: [INAUDIBLE] higher
standard deviation than [INAUDIBLE]. KENNETH ABBOTT:
Say it once again. AUDIENCE: Better
yield, does it mean that we have a higher standard
deviation than [INAUDIBLE]? KENNETH ABBOTT: No. The standard deviation is
the standard deviation. No matter what I do, this is
standard deviation, that's it. Don't have a higher
standard deviation. But the likelihood
of-- the put it this way-- the likelihood
of a move of 2.33 standard deviations
is more than 1%. That's the way I think of it. Make sense? AUDIENCE: Is there any way
for you to [INAUDIBLE] to-- KENNETH ABBOTT: What? AUDIENCE: Sorry,
is there any way to put into that graph what
a fatter tail looks like? KENNETH ABBOTT: Oh,
well, be patient. If we have time. In fact, we do
that all the time. And one of our
techniques doesn't care. It goes to the
empirical distribution. So it captures the
fat tails completely. In fact, the homework
assignment which I usually precede this lecture
by has people graphing all kinds
of distributions to see what these
things look like. We won't have time for that. But if you have questions,
send them to me. I'll send you some stuff
to read about this. All right, so now you
know one asset VaR, now you're qualified to
go work for a big bank. All right? Get your data,
calculate returns. Now I usually put in step 2b,
graph your data and look at it. All right? Because everybody's
data has dirt in it. Don't trust anyone else. If you're going
to get fired, get fired for being
incompetent, don't get fired for using
someone else's bad data. Don't trust anyone. My mother gives me data,
Mom, I'm graphing it. Because I think you let
some poop slip into my data. Mother Teresa could come to me
with a thumb drive: "Ken, S&P 500." Sorry, Mother Teresa. I'm graphing it before I use it. All right? So I don't want to say that
this is usually in here. We do extensive error testing. Because there could be bad data,
there could be missing data. And missing data is a whole
other lecture that I give. You might be shocked
at [INAUDIBLE]. So for one asset VaR, get my
data, create my return series. Percentage changes, log changes. Sometimes that's
absolute differences. Take the variance, take the
square root of the variance, multiply by 2.33. Done and dusted. Go home, take your
shoes off, relax. OK. Percentage changes
versus log changes. For all intents and purposes,
it doesn't really matter and I will often use
one or the other. The way I think about
this-- all right, there'll be a little
bit of bias at the ends. But for the overwhelming
bulk of the observations whether you use percentage
changes or log changes doesn't matter. Generally, even though I
know the data is closer to log-normally distributed
than normally distributed, I'll use percentage changes
just because it's easier. Why would we use
log-normal distribution? Well, when we're
doing simulation, the log-normal distribution
has this very nifty property of keeping your yields
from going negative. But, even that-- I
can call that into question because
there are instances of yields going negative. It's happened. Doesn't happen a
lot, but it happens. All right. So I talked about
bad data, talked about one-sided
versus two-sided. I'll talk about longs and shorts
a little bit later when we we're talking multi-asset. I'm going to cover a
fixed income piece. We use this thing called a PV01
because what I measure in fixed income markets isn't a price. I usually measure a yield. I have to get from a change
of yield to a change of price. Hmm, sounds like
a Jacobian, right? With kind of a poor
man's Jacobian. It's a measure that
captures the fact that my price-yield
relationship-- price, yield-- is non-linear. For any small approximation
I look at the tangent. And I use my PV01 which has
a similar notion to duration, but PV01 is a little
more practical. The slope of that
tells me how much my price will change for
a given change of yield. See, there it is. You knew you were going to
use the calculus, right? You're always
using the calculus. You can't escape it. But the price-yield
line is non-linear. But for all intents and
purposes, what I'm doing is I'm shifting the
price-yield relationship-- I'm shifting my yield
change into price change by multiplying my yield
change by my PV01 which is my price sensitivity to
1/100th percent move in yields. Think about that for a second. We don't have time
to-- I would love to spend an hour on this,
and about trading strategies, and about bull steepeners
and bear steepeners and barbell trades, but we
don't have time for that. Suffice to say if I'm
measuring yields the thing is going to trade as a 789
or a 622 or a 401 yield. How do I get that in
the change in price? Because I can't tell my
boss, hey, I had a good day. I bought it at 402
and sold it at 401. No, how much money did you make? Yield to coffee break
yield to lunch time, yield to go home
at the end of day. How do I get from change in
yield to change in price? Usually PV01. I could use duration. Bond traders who think in
terms of yield to coffee break, yield to lunch time, yield to
go home at the end of the day typically think
in terms of PV01. Do you agree with
that statement? AUDIENCE: [INAUDIBLE] KENNETH ABBOTT: How often
on the fixed income desk did you use duration measures? AUDIENCE: Well,
actually, [INAUDIBLE]. KENNETH ABBOTT: Because
of the investor horizon? OK, the insurance companies. Very important point I want to
reach here as a quick aside. You're going to hear
this notion of PV01, which is called PVBP or DV01. That's the price sensitivity
to a one basis point move. One basis point is 1/100th
of a percent in yield. Duration is the half life,
essentially, of my cash flow. What's the weighted expected
time to owe my cash flows? If my duration is 7.9 years,
my PV01 is probably about $790 per million. In terms of significant digits,
they're roughly the same but they have different meanings
and the units are different. Duration is measured in yield,
PV01 is measured in dollars. In bond space I
typically think in PV01. If I'm selling to
long term investors they have particular demands
because they've got cash flow payments they have to hedge. So they may think of it
in terms of duration. For our purposes, we're
talking DV01 or PV01 or PVBP, those three terms
more or less equal. Make sense? Yes? AUDIENCE: [INAUDIBLE] in
terms of [INAUDIBLE] versus [INAUDIBLE]? KENNETH ABBOTT: We could. In some instances, in
some areas and options we might look at
an overall 1% move. But we have to look at
what trades in the market. What trades in the
market is the yield. When we quote the
yield, I'm going to quote it going
from 702 to 701. I'm not going to have the
calculator handy to say, a 702 move to a 701. What's 702 minus
701 divided by 702? Make sense? It's the path of
least resistance. What's the difference between
a bond and a bond trader? A bond matures. A little fixed
income humor for you. Apparently very little. I don't want to spend
too much time on this because we just
don't have the time. I provide an example here. If you guys want
examples, contact me. I'll send you the spreadsheets
I use for other classes if you just want to
play around with it. When I talk about PV01,
when I talk about yields, I usually have some
kind of risk-free rate. Although this whole notion
of the risk-free rate, which is-- so much of modern
finance is predicated on this assumption that
there is a risk-free rate, which used to be
considered the US treasury. It used to be
considered risk-free. Well, there's a credit spread
out there for US Treasury. I don't mean to throw a
monkey wrench into the works. But there's no such thing. I'm not going to question 75
years of academic finance. But it's troublesome. Just like when I was taking
economics 30 years ago, inflation just mucked
with everything. All of the models fell apart. There were appendices
to every chapter on how you have to change this
model to address inflation. And then inflation went away
and everything was better. But this may not go away. I've got two components here. If the yield is 6%, I might
have a 450 treasury rate and 150 basis point credit spread. The credit spread reflects
the probability of default. And I don't want to get into
measures of risk neutrality here. But if I'm an issuer and I
have a chance of default, I have to pay my investors more. Usually when we
measure sensitivity we talk about that
credit spread sensitivity and the risk-free sensitivity. We say, well, how could
they possibly be different? And I don't want to
get into detail here, but the notion is, when credit
spreads start getting high, it implies a higher
probability of default. You have to think about credit
spread sensitivity a little differently. Because when you get
to 1,000 basis points, 1,500 basis points
credit spread, it's a high
probability of default. And your credit models
will think differently. Your credit models will
say, ah, that means I'm not going to get
my next three payments. There's an expected, there's
a probability of default, there's a loss given default,
and there's recovery. A bunch of other stochastic
measures come into play. I don't want to spend any more
time on it because it's just going to confuse you now. Suffice to say we have
these yields and yields are composed of risk-free
rates and credit spreads. And I apologize for
rushing through that, but we don't have time to do it. Typically you have
more than one asset. So in this framework where I
take 2.33 standard deviations times my dollar investment,
or my renminbi investment or my sterling investment. That example was with one asset. If I want to expand
this, I can expand this using this notion of
covariance and correlation. You guys covered correlation
and covariance at some point in your careers? Yes, no? All right? Both of them measure the way one
asset moves vis-à-vis another asset. Correlation is scaled between
negative 1 and positive 1. So I think of correlation
as an index of linearity. Covariance is not scaled. I'll give you an example of the
difference between covariance and correlation. What if I have 50 years
of data on crop yields and that same 50 years of data
on tons of fertilizer used? I would expect a
positive correlation between tons of fertilizer
used and crop yields. So the correlation would
exist between negative 1 and positive 1. The covariance
could be any number, and that covariance
will change depending on whether I measure
my fertilizer in tons, or in pounds, or
in ounces, or in kilos. The correlation will
always be exactly the same. The linear relationship is
captured by the correlation. But the units-- in
covariance, the units count. If I have covariance--
here it is. Covariance matrices
are symmetric. They have the variance
along the diagonal. And the covariance is
on the off-diagonal. Which is to say that the
variance is the covariance of an item with itself. The correlation
matrix, also symmetric, is the same thing scaled,
with correlations, where the diagonal is 1.0. If I have covariance--
because correlation is covariance--
covariance divided by the product of the
standard deviations gets me-- sorry--
correlation hat. This is like the
apostrophe in French. You forget it all the time. But the one time you really
need it, you won't do it and you'll be in trouble. If you have the covariances,
you can get to the correlations. If you have the
correlations, you can't get to the covariances
unless you know the variances. That's a classic
mid-term question. I give that almost-- not every
year, maybe every other year. Don't have time to spend
much more time on it. Suffice to say this
measure of covariance says when x is a certain
distance from its mean, how far is y from its mean
and in what direction? Yes? Now this is just a
little empirical stuff because I'm not as
clever as you guys. And I don't trust anyone. I read it in the textbook,
I don't trust anyone. a, b, here's a plus b. Variance of a plus b is
variance of a plus variance of b plus 2 times covariance a b. It's not just a good
idea, it's the law. I saw it in a thousand
statistics textbooks, I tested it anyway. Because if I want
to get fired, I'm going to get fired for
making my own mistake, not making someone
else's mistake. I do this all the time. And I just prove it
empirically here. The proof of which will be left
to the reader as an exercise. I hated when books said that. PROFESSOR: I actually
kind of think that's a proven
point, that you really should never trust output from
computer programs or packages-- KENNETH ABBOTT: Or your
mother, or Mother Teresa. PROFESSOR: It's
good to check them. Check all the calculations. KENNETH ABBOTT: Mother Teresa
will slip you some bad data if she can. I'm telling you, she will. She's tricky that way. Don't trust anyone. I've caught mistakes
in software, all right? I had a programmer-- it's
one of my favorite stories-- we're doing one of our first
Monte Carlo simulations, and we're factoring a matrix. If we have time, we'll get--
so I factor a covariance matrix into E transpose lambda E. It's
our friend the quadratic form. We're going to see this again. And this is a diagonal
matrix of eigenvalues. And I take the
square root of that. So I can say this is E transpose
lambda to the 1/2 lambda to the 1/2 E. And so my programmer had
gotten this, and I said, do me a favor. I said, take this, and transpose
and multiply by itself. So take the square
root and multiply it by the other square root, and
show me that you get this. Just show me. He said I got it. I said you got it? He said out to 16 decimals. I said stop. On my block, the square root
of 2 times the square root of 2 equals 2.0. All right? 2.0000000, what do you mean
out to 16 decimal places? What planet are you on? And I scratched the
surface, and I dug, and I asked a
bunch of questions. And it turned out
in this code he was passing a float to a fixed. All right? Don't trust anyone's software. Check it yourself. Someday when I'm dead and
you guys are in my position, you'll be thanking me for that. Put a stone on my
grave or something. All right so covariance. Covariance tells me
some measure of when x moves, how far does y move?
[? Or ?] for any other asset? Could I have a piece
of your cookie? I hardly had lunch. You want me to have a
piece of this, right? It's just looking
very good there. Thank you. It's foraging. I'm convinced 10
million years ago, my ape ancestors
were the first one at the dead antelope
on the planes. All right. So we're talking about
correlation, covariance. Covariance is not unit free. I can use either, but I have to
make sure I get my units right. Units screw me up every time. They still screw me up. That was a good cookie. All right. So more facts. Variance of xa
times yb; x squared variance a, y squared variance
b plus 2xy covariance ab. You guys seen this before? I assume you have. Now I can get pretty
silly with this if I want. x, a, y, b you get
the picture, right? But what you should
be thinking, this is a covariance matrix,
sigma squared, sigma squared, sigma squared. It's the sum of the variances
plus 2 times the sum of the covariances. So if I have one unit of every
asset, I've got n assets, all have to do to get the
portfolio variance is sum up the whole covariance matrix. Now, you never get only
one unit, but just saying. But you notice that this is
kind of a regular pattern that we see here. And so what I can
do is I can use a combination of my
correlation matrix and a little bit of linear
algebra legerdemain, to do some very
convenient calculations. And here I just give an
example of a covariance matrix and a correlation matrix. Note the correlation
matrices between negative 1 and positive 1. All right. Let me cut to the chase here. I'll draw it here
because I really want to get into some
of the other stuff. What this means, if I have a
covariance structure sigma. And I have a vector
of positions, x dollars in dollar/yen,
y dollars in gold, z dollars in oil. And let's say I've
got a position vector, x_1, x_2, x_3, x_n. If I have all my
positions recorded as a vector-- this is asset
one, asset two, and this is in dollars-- and I have
the covariance structure, the variance of
this portfolio that has these assets and this
covariance structure-- this is where the magic happens--
is x transpose sigma x equals sigma squared hat portfolio. Now you really could
go work for a bank. This is how portfolio
variance, using the variance/covariance
method, is done. In fact, when we were doing
it this way 20 years ago, spreadsheets only
have 256 columns. So we tried to simplify
everything into 256-- or sometimes you
had to sum it up using two different
spreadsheets. We didn't have
multitab spreadsheets. That was a dream,
multitab spreadsheets. This was Lotus 1-2-3 we're
talking about here, OK? You guys don't even know
what Lotus 1-2-3 is. It's like an abacus
but on the screen. Yes? AUDIENCE: What's
x again in this? KENNETH ABBOTT: Position vector. Let's say I tell you that you've
got dollar/yen, gold, and oil. You've got $100 of dollar/yen,
$50 of oil, and $25 of gold. It would be 100, 50, 25. Now, I should say
$100 of dollar/yen, your position vector
would actually show up as negative 100, 50, 25. Why is that? Because if I'm measuring
my dollar/yen-- and this is just a
little aside-- typically, I measure dollar/yen
in yen per dollar. So dollar/yen might be 95. If I own yen and I'm a dollar
investor and I own yen, and yen go from 95 per
dollar to 100 per dollar, do I make or lose money? I lose money. Negative 100. Just store that. You won't be tested on that,
but we think about that all the time. Same thing with yields. Typically, when I
record my PV01-- and I'll record some version,
something like my PV01 in that vector, my
interest rate sensitivity, I'm going to record
it as a negative. Because when yields go up and
I own the bond, I lose money. Signs, very important. And, again, we've
covered-- usually I do this in a two hour lecture. And we've covered it in less
than an hour, so pretty good. All right. I spent a lot more time
on the fixed income. [STUDENT COUGHING] Are you taking
something for that? That does not sound healthy. I don't mean to embarrass you. But I just want to make
sure that you're taking care of yourself because grad
students don't-- I was a grad student, I didn't take
care of myself very well. I worry. All right. Big picture,
variance/covariance. Collect data,
calculate returns, test the data, matrix construction,
get my position vector, multiply my matrices. All right? Quick and dirty,
that's how we do it. That's the simplified
approach to measuring this order statistic
called value at risk using this particular technique. Questions, comments? Anyone? Anything you think I need
to elucidate on that? And this is, in fact, how we
did this up until the late '90s. Firms used variance/covariance. I heard a statistic
in Europe in 1996 that 80% of the European banks
were using this technique to do their value at risk. It was no more
complicated than this. I use a little flow diagram. Get your data returns,
graph your data to make sure you
don't screw it up. Get your covariance matrix,
multiply your matrices out. x transpose sigma x. Using the position vectors and
then you can do your analysis. Normally I would
spend some more time on that bottom row and different
things you can do with it, but that will have
to suffice for now. A couple of points I want
to make before we move on about the assumptions. Actually, I'll fly through
this here so we can get into Monte Carlo simulation. Where am I going to get my data? Where do I get my data? I often get a lot of
my data from Bloomberg, I get it from public sources,
I get it from the internet. Especially when you
get it from-- look, if it says so on the
internet, it must be true. Right? Didn't Abe Lincoln say,
don't believe everything you read on the internet? That was a quote, I
saw that some place. You get data from
people, you check it. There's some sources
that are very reliable. If you're looking for yield
data or foreign exchange data, the Federal Reserve has it. And they have it back
20 years, daily data. It's the H.15 and the H.10. It's there, it's free,
it's easy to download, just be aware of it. Exchange-- PROFESSOR: [INAUDIBLE]
study posted on the website that goes
through computations for regression analysis
and asset pricing models and the data that's used
there is from the Federal Reserve for yields. KENNETH ABBOTT: It's
H.15 It's for yields, it's probably from the H.15. [INTERPOSING VOICES] PROFESSOR: Those files, you
can see how to actually get that data for yourselves. KENNETH ABBOTT: Now,
another great source of data is Bloomberg. Now the good thing
about Bloomberg data is everybody uses
it, so it's clean. Relatively clean. I still find errors in
it from time to time. But what happens is when you
find an error in your Bloomberg data, you get on the phone
to Bloomberg right away and say I found an
error in your data. They say, oh, what date? June 14, you know, 2012. And they'll say,
OK, we'll fix it. All right? So everybody does that, and
the data set is pretty clean. I found consistently
that Bloomberg data is the cleanest in my experience. How much data do we
use in doing this? I could use one year of data,
I can use two weeks of data. Now, times series, we usually
want 100 observations. That's always been
my rule of thumb. I can use one year of data. There are regulators
that require you to use at least a year of data. You could use two years of data. In fact, some firms
use one year of data. There's one firm that
uses five years of data. And there, we could say,
well, am I going to weight it. Am I going to weight my
more recent data heavily? I could do that with
exponential smoothing, which we won't have time to talk about. It's a technique I can use to
lend more credence to the more recent data. Now, I'm a relatively
simple guy. I tend to use
equally weighted data because I believe in
Occam's razor, which is, the simplest explanation
is usually the best. I think we get
too clever by half when we try to parameterize. How much more does
last week's data have an impact than from two
weeks ago, three weeks ago. I'm not saying that it
doesn't, what I am saying is, I'm not smart enough to know
exactly how much it does. And assuming that
everything's equally weighted throughout time is
just as strong an assumption. But it's a very simple
assumption, and I love simple. Yes? AUDIENCE: [INAUDIBLE]
calculate covariance matrix? KENNETH ABBOTT: Yes. All right, quickly. Actually I think I have
some slides on that. Let me just finish this
and I'll get to that. Gaps in data. Missing data is a problem. How do I fill in missing data? I can do a linear interpolation,
I can use the prior day's data. I can do a Brownian
bridge, which is I just do a Monte Carlo between them. I can do a regression
based, I can use regression to project changes from one
onto changes in another. That's usually a
whole other lecture I gave on how to
do missing data. Now you've got that
lecture for free. That's all you need to know. It's not only a lecture, it's a
very hard homework assignment. But how frequently
do I update my data? Some people update their
covariance structures daily. I think that's an overkill. We update our data set weekly. That's what we do. And I think that's overkill,
but tell that to my regulators. And we use daily data,
weekly data, monthly data. We typically use daily data. Some firms may do
it differently. All right. Here's your
exponential smoothing. Remember, I usually measure
covariance, sum of x_i minus x bar times y minus
y bar divided by n minus 1. What if I stuck
an omega in there? And I use this
calculation instead, where the denominator is the sum
of all the omegas-- you should be thinking finite series. You have to realize, I
was a decent math student, I wasn't a great math student. And what I found when I was
studying this, I was like, wow, all that stuff that I
learned, it actually-- finite series, who knew? Who knew that I'd
actually use it? So I take this, and let's say
I'm working backwards in time. So today's observations is t_0. Yesterday's observation
is t_1, t_2, t_3. So today's observation
would get-- and let's assume for the time
being that this omega is on the order 0.95. It could be anything. So today would be
0.95 to the 0 divided by the sum of all the omegas. Tomorrow it will be 0.95 divided
by the sum of the omegas. The next would be
0.95 squared divided by the sum of the omegas. 0.95 cubed and get
smaller and smaller. For example, if you use
0.94, 99% of your weight will be in the last 76 days. 76 observations, I
shouldn't say 76 days. 76 observations. So there's this notion that the
impact declines exponentially. Does that make sense? People use this pretty commonly,
but what scares me about it-- somebody stuck these
fancy transitions in between these slides. Anyway, is that here's
my standard deviation with a rolling six-month window. And here's my standard deviation
using different weights. The point I want to
make here, and it's an important point, my
assumption about my weighting coefficient has
a material impact on the size of my
measured volatility. Now when I see this,
and this is just me. There's no finance
or statistics theory behind this, any time the
choice-- any time an assumption has this material an impact,
bells and whistles go off and sirens. All right, and red lights flash. Be very, very careful. Now, lies, damn
lies, and statistics. You tell me the
outcome you want, and I'll tell you what
statistics to use. That's where this
could be abused. Oh, you want to show
high volatility? Well let's use this. You want to show low
volatility, let's use this? See, I choose to just take
the simplest approach. And that's me. That's not a terribly
scientific opinion, but that's what I think. Daily versus weekly,
percentage changes log changes. Units. Just like dollar/yen,
interest rates. Am I long or am I short? If I'm long gold, I show
it as a positive number. And if I'm short gold,
in my position vector, I show it as a negative number. If I'm long yen, and yen is
measured in yen per dollar, then I show it as
a negative number. If I'm long yen, but my
covariance matrix measures yen as dollars per yen--
0.000094, whatever-- then I show it as
a positive number. It's just like
physics only worse because it'll cost
you real-- no, I guess physics would
be worse because if you get the units wrong,
you blow up, right? This will just cost you money. I've made this mistake. I've made the units mistake. All right, we talked
about fixed income. So that's what I want to
cover from the bare bones setup for VaR. Now I'm going to skip
the historical simulation and go right to the
Monte Carlo because I want to show you another way we
can use covariance structures. [POWERPOINT SOUND EFFECT] That's going to happen
two or three more times. Somebody did this, somebody
made my presentation cute some years ago. And I just-- I apologize. All right, see, there's a lot
to meat in this presentation that we don't have
time to get to. Another approach to
doing value at risk is rather than use this
parametric approach, is to simulate the outcomes. Simulate the outcomes 100 times,
1,000 times, 10,000 times, a million times, and say, these
are all the possible outcomes based on my simulation
assumptions. And let's say I
simulate 10,000 times, and I have 10,000 possible
outcomes for tomorrow. And I wanted to measure my value
at risk at the 1% significance level. All I would do is take
my 10,000 outcomes and I would sort them and
take my hundredth worst. Put it in your pocket, go home. That's it. This is a different
way of getting to that order statistic. Lends a lot more flexibility. So I can go and I can tweak
the way I do that simulation, I can relax my
assumptions of normality. I don't have to use
normal distribution, I could use a t distribution,
I could do lots, I could tweak my distribution,
I could customize it. I could put mean
reversion in there, I could do all kinds of stuff. So another way we do
value at risk is we simulate possible outcomes. We rank the outcomes,
and we just count them. If I've got the
10,000 observations and I want my 5%
order statistic, well I just take my 500th. Make sense? It's that simple. Well, I don't want
to make it seem like it's that simple
because it actually gets a little messy in here. But when we do Monte
Carlo simulation, we're simulating what we
think is going to happen all subject to our assumptions. And we run through this
Monte Carlo simulation. Simulation of method using
sequences of random numbers. Coined during the
Manhattan Project, similar to games of chance. You need to describe your system
in terms of probability density functions. What type of distribution? Is this normal? Is it t? Is it chi squared? Is it F? All right? That's the way we do it. So quickly, how do I do that? I have to have random numbers. Now they're truly
random numbers. Somewhere at MIT you could
buy-- I used to say tape, but people don't use tape. They'll give you a website where
you can get the atomic decay. That's random. All right? Anything else is pseudo-random. What you see when
you go into MATLAB, you have a random number
generator, it's an algorithm. It probably takes some number
and takes the square root of that number and then goes
54 decimal places to the right and takes the 55 decimal
places to the right, multiplies those
two numbers together and then takes the fifth root,
and then goes 16 decimal places to the right to get that--
it's some algorithm. True story, before I came to
appreciate that these were all highly algorithmically
driven, I was in my 20's, I was taking a computer
class, I saw two computers, they were both running
random number of generators and they were generating
the same random numbers. And I thought I was
at the event horizon. I thought that light was
bending and the world was coming to an end, all right? Because this this stuff
can't happen, all right? It was happening
right in front of me. It was a pseudo-random
number generator. I didn't know, I was 24. Anyway. quasi-random numbers, it's sort
of a way of imposing some order on your random numbers. You random numbers, one
particular set of draws may not have enough draws
in a particular area to give you the
numbers you want. I can impose some
conditions upon that. I don't want to get into a
discussion of random numbers. How do I get from
random uniform-- most random number generators
give you random uniform number between 0 and 1. What you'll typically do is
you'll take that random uniform number, you'll map it over
to the cumulative density function, and map it down. So this gets you from
random uniform space into standard deviation space. We used to worry
about how we did this, now your software
does it for you. I've gotten comfortable
enough, truth be told. I usually trust my random number
generators in Excel, in MATLAB. So I kind of violate my
own rules, I don't check. But I think most of your
standard random number of generators are
decent enough now. And you can go
straight to normal, you don't have to do
random uniform and back into random normal. You can get it distributed
in any way you want. What I do when I do a Monte
Carlo simulation-- and this is going to be rushed because
we've only got like 20 minutes. If I take a covariance
matrix-- you're going to have to
trust me on this because again, I'm covering
like eight hours of lecture in an hour and a half. You guys go to MIT so
I have no doubt you're going to be all over this. Let's take this out
of here for a second. I can factor my
covariance structure. I can factor my covariance
structure like this. And this is the
transpose of this. I didn't realize that the first
time we did this commercially I saw this instead of
this and I thought we had sent bad data to the customer. I got physically sick. And then I remembered
AB transpose equals B transpose A--
these things keep happening. My high school math
keeps coming back to me. But I had forgotten this
and I got physically sick because I thought we'd
sent bad data because I was looking at this when it's
just the transpose of this. Anyway, I can factor
this into this where this is a matrix of eigenvectors. This is a diagonal matrix
of the eigenvalues. All right? This is the vaunted
Gaussian copula. This is it. Most people view
it as a black box. If you've had any more than
introductory statistics, this should be a
glass box to you. That's why I wanted to go
through this even though I'd love to spend another hour and
a half and do about 50 examples. Because this is
how I learned this, I didn't learn it from looking
at this equation and saying, oh, I get it. I learned it from actually
doing it about 1,000 times in a spreadsheet, and sunk
in like water into a stone. So I factor this
matrix, and then I take this, which is the
square root matrix, which is my transpose of
my eigenvector matrix and diagonal matrix contain the
square root of my eigenvalues. Now, could this ever
be negative and take me into imaginary root land? Well, if my variances
are positive or zero, then that will be a problem. So here we get into
this-- remember you guys studied positive
semidefinite, positive definite. Once again, it's another one of
these high school math things. Like, here it is. I had to know this. Suddenly I care whether
it's positive semidefinite. Covariance structures have
to be positive semidefinite. If you don't have a
complete data set, let's say you've got 100
observations, 100 observations, 100 observations, 25
observations, 100 observations, you may have a
negative eigenvalue. If you just measure the
covariance with the amount of data that you have. My intuition-- and I doubt
this is the [INAUDIBLE]-- is that you're measuring
with error and you have fewer observations you
measure with more error. So it's possible if some
of your covariance measures have 25 observations
and some of them have 100 observations that
there's more error in some than in others. And so there's the
theoretical possibility for negative variance. True story, we didn't
know this in the '90s. I took this problem to the
chairman of the statistics department at NYU said, I'm
getting negative eigenvalues. And he didn't know. He had no idea,
he's a smart guy. You have to fill in
your missing data. You have to fill in
your missing data. If you've got 1,000
observations, 1,000 observations, 1,000
observations, 200 observations, and you want to make
sure you won't have a negative
eigenvalue, you've got to fill in those observations. Which is why missing data
is a whole other thing we talk about. Again, I could spend
a lot of time on that. And I learned that the hard way. But anyway, so I take
this square root matrix, if I pre-multiply that square
root matrix by row after row of normals, I will
get out an array that has the same
covariance structure as that with which I started. Another story here,
I've been using the same eigenvalue-- I
believe in full attribution, I'm not a clever guy. I have not an original
thought in my head. And whenever I use
someone else's stuff, I give them credit for it. And the guy who
wrote the code that did the eigenvalue
decomposition-- this is something that was
translated from Fortran IV. It wasn't even
[INAUDIBLE], there's a dichotomy in the world. There are people that
have written Fortran, and people that haven't. I'm guessing that there are two
people in this room that have ever written a line of Fortran. Anyone here? Just saying. Yeah, with cards
or without cards? PROFESSOR: [INAUDIBLE]. KENNETH ABBOTT: I
didn't use cards. See, you're an old-timer
because you used cards. The punch line is, I've
been using this guy's code. And I could show you the code. It's like the Lone
Ranger, I didn't even get a chance to thank him. Because he didn't put
his name on the code. On the internet now, if
you do something clever on the quant
newsgroups, you're going to post your name all over it. I've been wanting to thank
this guy for like 20 years and I haven't been able to. Anyway, eigenvalue code
that's been translated. Let me show you what this means. Here's some source data. Here's some percentage changes. Just like we talked about. Here is the
empirical correlation of those percentage changes. So the correlation of my
government 10 year to my AAA 10 year is 0.83. To my AA, 0.84. All right, you see this. And I have this
covariance matrix which is the-- the correlation
matrix is a scaled version of the covariance matrix. And I do a little bit of
statistical legerdemain. Eigenvalues and eigenvectors. Take the square root of that. And again, I'd love to spend
a lot more time on this, but we just don't--
suffice to say, I call this a transformation
matrix, that's my term. This matrix here is this. If we had another
hour and a half I'd take the step by
step to get you there. The proof of which is left
to the reader as an exercise. I'll leave this spreadsheet
for you, I'll send it to you. I have this matrix. This matrix is like a prism. I'm going to pass
white light through it, I'm going to get a
beautiful rainbow. Let me show you what I mean. So remember that matrix,
this matrix I'm calling t. Remember my matrix is 10 by 10. One, two, three, four, five,
six, seven, eight, nine, ten. 10 columns of data. 10 by 10 correlation matrix. Let's check. Now I've got row vectors
of sorry-- uncorrelated random normals. So what I'm doing then
is I'm pre-multiplying that transformation matrix
row by row by each row of uncorrelated random normals. And what I get is
correlated random normals. So what I'm telling
you here is this array happens to be 10
wide and 1,000 long. And I'm telling
you that I started with my historical data-- let
me see how much data have there. A couple hundred observations
of historical data. And what I've done is once I
have that covariance structure, I can create a
data set here which has the same statistical
properties as this. Not quite the same. It can have the same means
and the same variances. This is what Monte Carlo
simulation is about. I wish we had another hour
because I'd like to spend time and-- this is one
of these things, and again, when I first saw
this, I was like, oh my god. I felt like I got the
keys to the kingdom. And I did, this is manually,
did it all on a spreadsheet. Didn't believe
anyone else's code, did it all on a spreadsheet. But what that means-- quickly,
let me just go back over here for a second. I happen to have about
800 observations here. Historical observations. What I did was I happened to
generate 1,000 samples here. But I could generate
10,000 or 100,000, or a million or 10
million or a billion just by doing more
random normals. I could generate--
in effect, what I'm generating here is
synthetic time series that have properties similar
to my underlying data. That's what Monte Carlo
simulation is about. The means and the variances and
the covariances of this data set are just like that. Now, again, true story, when
somebody first showed me this I did not believe them. So I developed a
bunch of little tests. And I said, let me just look
at the correlation of my Monte Carlo data versus my
original correlation matrix. So 0.83, 0.84, 0.85,
0.85, 0.67, 0.81. You look at the corresponding
ones of the random numbers I just generated, 0.81, 0.82,
0.84, 0.84, 0.64, 0.52. 0.54 versus 0.52. 0.18 Versus 0.12. 0.51 versus 0.47. Somebody want to tell me
why they're not spot on? Sampling error. The more data I use the
closer it will get to that. If I do 1 million, I'd better
get right on top of that. Does that make sense? So what I'm telling
you here is that I can generate synthetic time series. Now, why would I
generate so many? Well because,
remember, I care what's going on out in that tail. If I only have 100 observations
and I'm looking empirically at my tail, I've only got one
observation out in the 1% tail. And that doesn't tell me a
whole lot about what's going on. If I can simulate that
distribution exactly, I can say, you know what, I
want a billion observations in that tail. Now we can look at that tail. If I have 1 billion
observations, let's say I'm looking at some
kind of normal distribution. I'm circling it out
here, I'm seeing-- I can really dig in and see what
the properties of this thing are. In fact, this can really
only take two distributions, and really, it's only one. But that's another story. So what I do in Monte
Carlo simulations, I'm simulating these outcomes
so we can get a lot more meat in this tail to understand
what's happening out there. Does it drop off quickly? Does it not drop off quickly? That's kind of what it's about. So we're about out of time. We just covered like four
weeks of material, all right? But you guys are from MIT. I have complete
confidence in you. I say that to the
people who work for me. I have complete
confidence in your ability to get that done by
tomorrow morning. Questions or comments? I know you're sipping
from the fire hose here. I fully appreciate that. So those are examples. When I do this with
historical simulation I won't generate these
Monte Carlo trials, I'll just use historical data. And my fat tails
are built into it. But what I've shown
you today is what we developed a
one-asset VaR model, then we developed a multi-asset
variance/covariance model. And then I showed you
quickly, and in far less time than I would like
to have shown you, is how I can use another
statistical technique, which is called the Gaussian copula,
to generate synthetic data sets that will have the
same properties as my source historical data. All right? There you have it. [APPLAUSE] Oh you don't have to--
please, please, please. And I'll tell you, for me,
one of the coolest things was actually being
able to apply so much of the math I learned in
high school and in college and never thought
I'd apply again. One of my best
moments was actually finding a use for trigonometry. If you're not an engineer,
where are you going to use it? Where do you use it? Seasonals. You do seasonal estimation. And what you do is you do
fast Fourier transform. Because I can describe
any seasonal pattern with a linear combination of
sine and cosine functions. And it actually works. I have my students do it
as an exercise every year. I say, go get New York
city temperature data. And show me some
linear combination of sine and cosine
functions that will show me the seasonal
pattern of temperature data. And when I first realized I
could use trigonometry, yes! It wasn't a waste of time. I still-- polar
coordinates, I still haven't found a
use for that one. But it's there. I know it's there. All right? Go home.
