Jun Pan MIT Sloan School of Management 253-3083

junpan@mit.edu 15.433-15.4331 Financial Markets, Fall 2015 E62-624

Class 22: Fixed Income, Term Structure Models

This Version: November 23, 20161

1 Term Structure Models

• The Challenge from the Data: In the fixed income market, term structure models
are used to model interest rates. The challenge from the data has two dimensions.
First, it should take into account of how the interest rates move over time. Second, for
a given time, it should be able to model the yield curve, also called the term structure
of interest rates. Figure 1 is a good summary of these two challenges from the data:
a good term structure model should be able to capture the dynamic variations of the
level of interest rates and the shape of the yield curve.
These two demands from the data are very similar to those in the equity market. A
good model for stock market returns should be able to take into account of how stock
returns vary over time, as well as how, for a give time, the cross-section of stocks
are priced in relation to one another. In the equity market, an i.i.d. model for stock
returns is a reasonable approximation. As such, the dynamics for stock returns are
really simple: constant expected return μ, constant volatility σ, and unpredictable
random shocks t+1 . Cross-sectionally, the expected stock returns are linked to one
another through their exposures (i.e., betas) to risk factors in a model such as the
CAPM. As such, the CAPM model is a static model with constant expected returns
and constant beta.
The need for a dynamic model shows up when we investigated the time-varying volatil-
ity in our volatility class and stochastic volatility in our options class. Here in this class,
we have a chance to take a closer look at these dynamic models.

• Term Structure Models, Historical Development: Term structure models were

developed in the mid-1970s by Cox, Ingersoll and Ross (1985) and Vasicek (1997). You
1
A small correction of Figure 3, which was missing a portion of the Fed target rate.

1
 U.S. Treasury Constant Maturity Yield (in percent)
18
3-month (avg= 4.10%)
2-year (avg= 4.47%)
16
5-year (avg= 5.46%)
10-year (avg= 5.96%)
14 30-year (avg= 6.58%)

averages reported for 1982-2016

12
Yield (in percent)

10

0
1960 1965 1970 1975 1980 1985 1990 1995 2000 2005 2010 2015

Figure 1: Time-Series of Treasury Constant Maturity Yields.

2
might notice that the CIR paper was published in 1985, but it was really a product
of the mid-1970s. These term structure models were a continuation of the work done
by Black, Merton, and Scholes, who popularized the application of continuous-time
models in Finance. Like the Black-Scholes model before them, these term-structure
models use the stochastic processes studied by mathematicians and physicists. For
example, the CIR model builds on the Feller process and the Vasicek model builds on
the Ornstein-Uhlenbeck process. In both cases, the starting point is the instantaneous
short-rate rt , which is modeled by a stochastic process (OU or Feller). The entire yield
curve is priced using the dynamics of this one short rate. As such, the CIR and Vasicek
models are one-factor short-rate models.
The second wave of term structure models came in the 1990s. When I entered the
Stanford PhD program in 1995, I was just in time to catch the excitement surrounding
term structure models. Relative to the original models of CIR and Vasicek, the effort
of the new generation of term structure models is to be empirically relevant. From the
work of Litterman and Scheinkman (1991), it became clear that a one-factor model
will not be able to capture the entire shape of the yield curve. Unlike the stock market,
where you can dismiss the risk uncaptured by the model as idiosyncratic risk, we do
not have the luxury of dismissing common risk factors in the fixed income market (e.g.,
the slope factor).
These multifactor models quickly found their way into the “real” world. It is my un-
derstanding that each investment bank has its own proprietary term-structure model.
And I was told by some practitioners that the industry has the best and most so-
phisticated term structure models. And they use these models to manage and hedge
interest-rate risk (level, slope, convexity, volatility, etc) as well as to price interest-rate
derivatives and other rate-sensitive instruments such as MBS. Looking back, I can now
understand why during the mid-1990s, the Wall Street hired so many physicists and
mathematicians. Most of my classmates in Physics ended up on Wall Street. I can also
understand the sudden demand for more sophisticated term structure models in the
1990s. The fixed income desks were very profitable and the range and trading volume
of their fixed income products were also expanding very rapidly during that time.
By now, the excitement surrounding term structure models has all but fizzled out. As
a PhD student at Stanford, I spent much more time learning and working on term-
structure models than anything else I did there. Since coming to Sloan in 2000, I have
not made much use of that part of my training. Nevertheless, I am very grateful to
my advisers at Stanford for having trained me in this area. As I wrote earlier in my

3
 lecture notes, not everything we do in life is of practical use. Still, they are useful and
meaningful in our growth process.
For our class, however, I don’t want to emphasize too much on the modeling part,
because it takes quite a bit of mathematical skills. Instead, I would like to use the
term structure models as a way for us to understand conceptually how the various
parts of the yield curve are connected through a pricing model and the role of the risk
factors in generating the pricing results.

• Bond Pricing in Continuous-Time: Let rt be the time-t instantaneous short rate.

Let today be time 0, and let P0 be the present value of a dollar to be paid in T years.
Discounting this future dollar all the way from T to today using the short rate, we
have:  T 
P0 = E e− 0 rt dt (1)

Let me explain this expression in sequence:

T
– The reason why we need to do 0 rt dt is because we have to add up all of the
future short rates along the path from 0 to T . Take the extreme example of a
T
constant short rate r. We have 0 rt dt = r T and P0 = e−r T .
T T
– We put 0 rt dt onto e− 0 rt dt because the rates are continuously compounded.
(You will find that working with ex and ln(x) typically gives us a lot of tractability
in Finance.)
– Later on, we will see how rt is going to be driven by a random risk factor. Because
of this, there could be many paths of rt , depending on the random outcomes of
the risk factor. And the present value of a future dollar to be paid in year T is
an expectation, E (·), taken over all potential random paths of rt with t running
from 0 to T .

• Relating back to Option Pricing: The calculation in Equation (1) is similar to the
 
calculation of E Q e−rT (K − ST ) 1ST <K in option pricing. The difference is that we
do not have to deal with the random variation in ST . But we have to deal with the
random variation in the riskfree r, which turns out to be more difficult to deal with.
Instead of fixing a maturity date for this interest rate r (as in yields to maturity), we
choose to work with the “short rate” so that this one rate can be used to discount
T
future cashflows over any horizon. We just need to add them up via 0 rt dt.
A by-product of this modeling choice is that we now have to keep track of the entire path
T
of rt from 0 to T in order to calculate 0 rt dt. Remember that when you performed

4
 option pricing via simulation in your Assignment 3, you didn’t have to keep track the
path of ST from 0 to T . You only needed to know the values of ST . So in order to
have one million scenarios of ST , you needed to simulate one million random variables.
To price bonds, however, you need to simulate the entire path of rt from 0 to T .
Suppose we decide to discretize the time interval from 0 to T into monthly intervals,
then pricing a one-year bond with one million scenarios would involve simulating 12 ×
one million random variables; pricing a 10-year bond would involve simulation 120 ×
one million random variables. In short, pricing bond is generally more involving than
pricing equity options and pricing bond derivatives would be even more challenging.
That is why models with closed-form solutions are very useful. Otherwise, we will have
to resort to either simulations or solving partial differential equations.
Also notice that to be precise, I should take the expectation in Equation (1) under the
risk-neutral measure. For this class, however, let me not make a distinction between
the two, just to keep things simple.

• The Vasicek Model: In the Vasicek model, the short rate rt follows

drt = κ (r̄ − rt ) dt + σ dBt , (2)

where, as in the Black-Scholes model, σ dBt is the diffusion component with B as a

Brownian motion. This model has three parameters:

– r̄: The long-run mean of the interest rate, r̄ = E(rt ).

– κ: The rate of mean reversion. When rt is above its long-run mean r̄, r̄ − rt is
negative, exerting a negative pull on rt to make it closer to r̄. A larger κ amplifies
this pull of mean reversion and a smaller κ dampens it. Conversely, when rt is
below its long-run mean r̄, r̄ − rt is positive, exerting a positive pull on rt , again
to make it closer to its long-run mean r̄.
– σ: controls the volatility of the interest rate.

• Bond Pricing under Vasicek: Bond pricing under the Vasicek model turns out of
be very simple. Let today be time t and let rt be today’s short rate, then the time-t
value of a dollar to be paid T years later at time t + T is

Pt = eA+B rt ,

5
 where

e−κT − 1
B=
 κ −κT   
1−e σ 2 1 − e−2κT 1 − e−κT
A = r̄ −T + 2 −2 +T
κ 2κ 2κ κ

2 Calibrating the Model to the Data

• The Vasicek Model: As usual, we work with models in order to understand, at a
conceptual level, the key drivers in the pricing of a security. Applying the model to
the data, we further understand quantitatively how well the model works and what’s
missing in the model.
For a one-factor model such as the Vasicek model, we know its limitation even before
applying it to the data. In the fixed income market, the level of the interest rates is
the number one risk factor in terms of its importance, but it is not the only risk factor.
In Assignment 4, I ask you to work with a discrete-time version of the Vasicek model
by first estimating the model parameters, r̄, κ, and σ, using the time-series data of
3-month Tbill rates. Basically, I am asking you to calibrate the model only to the
time-series information of the short-end of the yield curve, without allowing you to
take into account of the information contained in the other parts of the yield curve.
Then I ask you to price the entire yield curve. Not surprising, you will find that the
calibrated model does not work very well to accommodate the different shapes of the
yield curve.
An alternative approach is to calibrate the model using the yield curve. For example,
on any given day, we estimate the model parameters, r̄, κ, and σ, so that the pricing
errors between the model yields and the market yields are minimized. By doing so,
the model will do a much better job in matching the market observed yield curve, but
it will miss the time-series information. Moreover, you will have one set of parameters
per day, which is inconsistent with the assumption that these parameters are constant.
The better solution is to introduce more factors to the model. For example, instead
of forcing the long-run mean r̄ to be a constant, we can allow it to vary over time by
modeling it as a stochastic process. Instead of forcing the volatility coefficient σ to be
a constant, we can allow it to vary as another stochastic process. There, you have a
three-factor model. The pricing will be more complicated and so will be the estimation.
Working with these multi-factor models requires some patience, perseverance, and the

6
 love for the subject matter. Indeed, it is not for everybody.

• Curve Fitting: On a topic related to model calibration is yield curve fitting. In this
approach, there is no consideration along the time-series dimension. The zero rate
r(τ ) of maturity τ is modeled as a parametric function, which is then used to price
all market traded coupon-bearing bonds. On any given day, the parameters in that
parametric function will be chosen so that the pricing errors between the model yields
and the market yields are minimized. This exercise of yield curve fitting is repeated
daily and the model parameters are updated daily as well.
Figure 2 plots the yield curve during the depth of the 2008 crisis. It uses the Svensson
model for curve fitting. The parameters in the Svensson model are first optimized so
that the model can price all of the market-traded bonds on December 11, 2008 with
minimum pricing errors. Using these parameters, the black line is the corresponding
par coupon curve. The blue or purple dots are the market yields for the market-
traded bonds. For each dot, there is a companion red “+”, which is the model yield
for the corresponding bond. In a fast decreasing interest rate environment such as
December 2008, most of the existing bonds are premium bonds. As we discussed
earlier, with an upward sloping term structure, the yields of these bonds are lower
than the corresponding par-coupon yields of the same maturity. That is why most of
the red “+”s are below the par coupon curve. If there are many discount bonds being
traded at the time, then you will see some red “+”s above the par coupon curve.
This curve fitting exercise is useful in connecting the yields of different maturities
through a parametric function of zero rates. For example, there is quite a bit of
overlap in discount rates between a ten-year yield and a ten-year minus one-month
yield. The presence of a parametric function of zero rates acknowledges the overlap
(ten years minus one month) and the pricing difference between these two yields will be
sensitive only to the one-month gap. But the usefulness of a curve fitting exercise stops
at this level. If you would like to use a model to help you with derivatives pricing on
the yield curve (e.g., Bond options, swaptions, caps/floors, etc), a curve-fitting model
will not be helpful at all because it does not take into consideration of how yields vary
over time. For derivatives pricing on the yield curve, you need to use dynamic models.
The usual approach is to use multi-factor versions of CIR or Vasicek models. Affine
models are examples of these multi-factor versions of CIR and Vasicek.

7
Figure 2: Treasury Yield Curve on December 11, 2008.

8
3 Relative Value Trading with a Term Structure Model
In March 2011, Chifu Huang (a former MIT Sloan Finance professor) came to Prof. Merton’s
class to give a guest lecture. I found his talk to be very informative and the following is
based on one portion of his talk.

• How to Use a Term-Structure Model to Identify Trading Opportunity: Rel-

ative value trading in the fixed income market does not make a judgment on the level
of interest rates or the slope of the curve. It assumes that a few points on the yield
curve are always fair. For example, the time-series data on the 10yr, 2yr, and 1-month
rates can be used to estimate a three-factor term structure model.
Recall that in the Vasicek model, the short rate is the only risk factor (i.e., state
variable). That is why in your Assignment 4, I ask you to estimate the model using
only the 3M Tbill rates. With a three-factor model, we have three risk factors (i.e.,
state variables) and we need three points on the yield curve to help us estimate the
model. Intuitively, the 10yr gives us information about the level of long-term interest
rates; the 2yr together with the 10yr informs us about the slope of the curve; and
the 1-month Tbill rate captures the short-term interest rate (including expectations
on monetary policy in the near term).
Once you have the model estimated by the time-series data (which is a non-trivial
task if you would like to do it properly), this model is going to give you predictions
about the level of interest rates across the entire yield curve. You can then compare
the model price with the market price to judge for yourself whether or not a market
price is cheap or expensive. Once you convince yourself that your model helps you pick
up a trading opportunity, you would structure a trade around it. You can buy cheap
maturities and sell expensive maturities, and, at the same time, hedge your portfolio
so that it is insensitive to the changes of the level or the slope of the yield curve.
The main judgment call is to understand why your model identifies some maturities
as cheap or expensive. If it is due to institutional reasons (which does not show up in
your model but does show up in the data), then you can make judgment as to whether
or not such institutional reasons will dissipate over time (and how fast).

• An Example:
One example was given by Chifu. In August 1998, Russian defaulted on its local
currency debt, and the effect lingered well into September and was later known as
the LTCM crisis. As shown in Figure 3, in September 1998, bond markets rallied in

9
 Figure 3: Fed Target and Treasury Yields in 1998.

Figure 4: Cheapness and Richness of US 30-Year Swap Rate Based on a Two-Factor Model.

10
anticipation of a rate cut. On September 29, the Fed cut the fed funds target rate by
25 bps.
Figure 4 is a slide presented by Chifu in his talk. In September 1998, his two-factor
model picks up a trading opportunity regarding the 30yr bond. According to the
model, the market price for the 30yr bond is cheap relative to the model price. The
deviation between the data and the model was at the range of 10 to 20 bps. The
30-year rate was around 5.5% at that time, implying a modified duration of about 15
years. So a 10 bps price deviation in 30yr would translate to 10 bps × 15 = 150 bps
in bond return. And a 20 bps deviation will translate to 3% in bond return.
So what are the reasons for this cheapening of 30yr? It is because residing over the
30yr region are pension funds and life insurance companies who are either inactive
“portfolio rebalancers” or rate-targeted buyers. As a result, the rally that happened
in the rest of the yield curve didn’t find its way to the 30yr region. There is a lag in
how information (regarding an impending rate cut) gets transmitted to this region. As
you can see from Figure 3 and 4, it was only after the Fed’s rate cut on September 28
when the 30yr yield was brought back in alignment with the rest of the yield curve.

11