“Raisuli, what shall we do? We have lost everything …”
“Sherif, was there never anything in your life for which it was worth losing everything for?”
—From The Wind and The Lion, MGM Films, Columbia Pictures, 1975.
This is probably the most important chapter in the book. Interest rates and the term structure of rates are the very foundation of modern principles of finance, yet this topic is not one that most senior bank executives are most au fait with. No doubt many of them would suggest that it is not an area they need to be technically expert in, but the fact remains that an understanding of the yield curve is fundamental to an understanding of the principles of finance.
The extracts in this chapter cover every aspect of interest rates, their term structure, spot and forward rates, dynamics of asset prices, fitting the curve, the secured curve, the multicurrency curve – everything but the kitchen sink, as they might say in 1940s Britain.
The new material in this book follows the book extracts, and ties in curve interpolation methodology with issues of relevance for the bank's internal funding or “funds transfer pricing” curve.
This extract from The Principles of Banking (2012)
This extract from Analysing and Interpreting the Yield Curve (2004)
The pricing of derivative instruments such as options is a function of the movement in the price of the underlying asset over the lifetime of the option, and valuation models describe an environment where the price of an option is related to the behavior process of the variables that drive asset prices. This process is described as a stochastic process, and pricing models describe the stochastic dynamics of asset price changes, whether this is a change in share prices, interest rates, foreign exchange rates or bond prices. To understand the mechanics of interest rate modeling therefore, we must familiarize ourselves with the behavior of functions of stochastic variables. The concept of a stochastic process is a vital concept in finance theory. It describes random phenomena that evolve over time, and these include asset prices. For this reason an alternative title for this chapter could be An Introduction to Stochastic Processes.
This is a text on bonds after all, not mathematics, and it is outside the scope of this book comprehensively to derive and prove the main components of dynamic asset pricing theory. There are a number of excellent textbooks that the reader is encouraged to read which provide the necessary detail, in particular Ingersoll (1987), Baxter and Rennie (1996), Neftci (1996) and James and Webber (2000). Another recommended text that deals with probability models in general, as well as their application in derivatives pricing, is Ross (2000). In this chapter, we review the basic principles of the dynamics of asset prices, which are then put into context in the following chapters, which look at term structure models.
The first property that asset prices, which can be taken to include interest rates, are assumed to follow is that they are part of a continuous process. This means that the value of any asset can and does change at any time and from one point in time to another, and can assume any fraction of a unit of measurement. It is also assumed to pass through every value as it changes, so for example if the price of a bond moves from 92.00 to 94.00, it must also have passed through every point in between. This feature means that the asset price does not exhibit jumps, which is not the case in many markets, where price processes do exhibit jump behavior. For now however, we may assume that the price process is continuous.
Models that seek to value options or describe a yield curve also describe the dynamics of asset price changes. The same process is said to apply to changes in share prices, bond prices, interest rates and exchange rates. The process by which prices and interest rates evolve over time is known as a stochastic process, and this is a fundamental concept in finance theory.1 Essentially a stochastic process is a time series of random variables. Generally the random variables in a stochastic process are related in a non‐random manner, and so therefore we can capture in a probability density function. A good introduction is given in Neftci (1996), and following his approach, we very briefly summarize the main features here.
Consider the function y = f(x). Given the value of x, we can obtain the value of y. If we denote the set W as the state of the world, where w ∈ W, the function f(x, w) has the property that given a value w ∈ W, it becomes a function of x only. If we say that x represents the passage of time, two functions f(x, w1) and f(x, w2) will be different because the second element w in each case is different. With x representing time, these two functions describe two different processes that are dependent on different states of the world W. The element w represents an underlying random process, and so therefore the function f(x, w) is a random function. A random function is also called a stochastic process, one in which x represents time and x ≥ 0. The random characteristic of the process refers to the entire process, and not any particular value in that process at any particular point in time.
Examples of functions include the exponential function denoted by y = ex and the logarithmic function loge(y) = x.
The price processes of shares and bonds, as well as interest rate processes, are stochastic processes. That is, they exhibit a random change over time. For the purposes of modeling, the change in asset prices is divided into two components. These are the drift of the process, which is a deterministic element,2 also called the mean, and the random component known as the noise, also called the volatility of the process.
We introduce the drift component briefly as follows. For an asset such as an ordinary share, which is expected to rise over time (at least in line with assumed growth in inflation), the drift can be modeled as a geometric growth progression. If the price process has no “noise”, the change in price of the stock from over the time period dt can be given by:
where the term μ describes the growth rate. Expression (5.1) can be rewritten in the form:
which can also be written in integral form. For interest rates, the movement process can be described in similar fashion, although as we shall see interest rate modeling often takes into account the tendency for rates to return to a mean level or range of levels, a process known as mean reversion. Without providing the derivation here, the equivalent expression for interest rates takes the form
where α is the mean reversion rate that determines the pace at which the interest rate reverts to its mean level. If the initial interest rate is less than the drift rate, the rate r will increase, while if the level is above the drift rate it will tend to decrease.
For the purposes of employing option pricing models, the dynamic behavior of asset prices are usually described as a function of what is known as a Weiner process, which is also known as Brownian motion. The noise or volatility component is described by an adapted Brownian or Weiner process, and involves introducing a random increment to the standard random process. This is described next.
The stochastic process we have briefly discussed above is known as Brownian motion or a Weiner process. In fact, a Weiner process is only a process that has a mean of 0 and a variance of 1, but it is common to see these terms used synonymously. Weiner processes are a very important part of continuous‐time finance theory, and interested readers can obtain more detailed and technical data on the subject in Neftci (1996) and Duffie (1996)3 among others. It is a well‐researched subject.
One of the properties of a Weiner process is that the sample pathway is continuous, that is, there are no discontinuous changes. An example of a discontinuous process is the Poisson process. Both are illustrated in Figures 5.1 and 5.2 below.
In the examples illustrated, both processes have an expected change of 0 and a variance of 1 per unit of time. There are no discontinuities in the Weiner process, which is a plot of many tiny random changes. This is reflected in the “fuzzy” nature of the sample path. However the Poisson process has no fuzzy quality and appears to have a much smaller number of random changes. We can conclude that asset prices, and the dynamics of interest rates, are more akin to a Weiner process. This, therefore, is how asset prices are modeled. From observation we know that, in reality asset prices and interest rates do exhibit discontinuities or jumps, however, there are other advantages in assuming a Weiner process, and in practice because continuous‐time stochastic processes can be captured as a combination of Brownian motion and a Poisson process, analysts and researchers use the former as the basis of financial valuation models.
The first step in asset pricing theory builds on the assumption that prices follow a Brownian motion. The properties of Brownian motion W state that it is continuous, and the value of Wt(t > 0) is normally distributed under a probability measure P as a random variable with parameters N(0, t). An incremental change in the asset value over time dt, which is a very small or infinitesimal change in the time, given by Ws+t − Ws, is also normally distributed with the parameters N(0, t) under P. Perhaps the most significant feature is that the change in value is independent of whatever the history of the price process has been up to time s. If a process follows these conditions, it is Brownian motion. In fact, asset prices do not generally have a mean of 0, because over time we expect them to rise. Therefore modeling asset prices incorporates a drift measure that better reflects asset price movement, so that an asset movement described by:
would be a Brownian motion with a drift given by the constant μ. A second parameter is then added, a noise factor, which scales the Brownian motion by another constant measure, the standard deviation σ. The process is then described by:
which can be used to simulate the price path taken by an asset, as long as we specify the two parameters. An excellent and readable account of this is given in Baxter and Rennie (1996, Chapter 3), who also state that under (5.5) there is a possibility of achieving negative values, which is not realistic for asset prices. However, using the exponential of the process given by (5.5) is more accurate, and is given by (5.6):
Brownian motion or the Weiner process is employed by virtually all option pricing models, and we introduce it here with respect to a change in the variable W over an interval of time t. If W represents a variable following a Weiner process and ΔW is a change in value over a period of time t, the relationship between ΔW and Δt is given by (5.7):
where ε is a random sample from a normal distribution with a mean 0 and a standard deviation of 1. Over a short period of time the values of ΔW are independent and therefore also follow a normal distribution with a mean 0 and a standard deviation of . Over a longer time period T made up of N periods of length Δt the change in W over the period from time 0 to time T is give by (5.8):
The successive values assumed by W are serially independent, therefore from (5.8) we conclude that changes in the variable W from time 0 to time T follow a normal distribution with mean 0 and a standard deviation of . This describes the Weiner process, with a mean of zero or a zero drift rate and a variance of T. This is an important result because a zero drift rate implies that the change in the variable (in this case, the asset price) in the future is equal to the current change. This means that there is an equal chance of an asset return ending up 10% or down 10% over a long period of time.
The next step in the analysis involves using stochastic calculus. Without going into this field here, we summarize from Baxter and Rennie (1996) and state that a stochastic process X incorporates a Newtonian term that is based on dt and a Brownian term based on the infinitesimal increment of W that is denoted by dWt. The Brownian term has a “noise” factor of αt. The infinitesimal change of X at Xt is given by the differential equation:
where σt is the volatility of the process X at time t and μt is the drift of X at time t. For interest rates that are modeled on the basis of mean reversion, the process is given by:
where the mean reverting element is as before. Without providing the supporting mathematics, which we have not covered here, the process described by (5.10) is called an Ornstein–Uhlenbeck process, and has been assumed by a number of interest rate models.
One other important point to introduce here is that a random process described by (5.10) operates in a continuous environment. In continuous‐time mathematics the integral is the tool that is used to denote the sum of an infinite number of objects, that is where the number of objects is uncountable. A formal definition of the integral is outside the scope of this book, but accessible accounts can be found in the texts referred to previously. A basic introduction is given at Appendix 5.4. However the continuous stochastic process X described by (5.9) can be written as an integral equation in the form:
where σ and μ are processes as before. The volatility and drift terms can be dependent on the time t but can also be dependent on X or W up to the point t. This is a complex technical subject and readers are encouraged to review the main elements in the referred texts.
The continuously compounded rate of return is an important component of option pricing theory. If r is the continuously compounded rate of return, we can use the lognormal property to determine the distribution that this follows. At a future date T, the asset price S may be written as (5.2):
and
Using the lognormal property we can describe the distribution of the risk‐free rate as:
We noted at the start of the chapter that the price of an option is a function of the price of the underlying stock and its behavior over the life of the option. Therefore, this option price is determined by the variables that describe the process followed by the asset price over a continuous period of time. The behavior that asset prices follow is a stochastic process, and so option pricing models—and term structure models—must capture the behavior of stochastic variables behind the movement of asset prices. To accurately describe financial market processes, a financial model depends on more than one variable. Generally, a model is constructed where a function is itself a function of more than one variable. Itô's lemma, the principal instrument in continuous time finance theory, is used to differentiate such functions. This was developed by a mathematician, K. Itô, in 1951. Here we simply state the theorem, as a proof and derivation are outside the scope of the book. Interested readers may wish to consult Briys et al. (1998), and Hull (1997) for a background on Itô's lemma. We also recommend Neftci (1996). Basic background on Itô's lemma is given in Appendices 5.2 and 5.3.
Brownian motion is very similar to a Weiner process, which is why it is common to see the terms used interchangeably. Note that the properties of a Weiner process requires that it be a martingale, while no such constraint is required for a Brownian process. A mathematical property known as the Lévy theorem allows us to consider any Weiner process Wt with respect to an information set Ft as a Brownian motion Zt with respect to the same information set.
We can view Brownian motion as a continuous time random walk, visualized as a walk along a line, beginning at X0 = 0 and moving at each incremental time interval dt either up or down by an amount . If we denote the position of the walk as Xn after the nth move, the position would be:
where the + and − signs occur with an equal probability of 0.5. This is a simple random walk. We can transform this into a continuous path by applying linear interpolation between each move point, so that:
It can be shown (but not here) that the path described at (5.24) has a number of properties, including that the incremental change in value each time it moves is independent of the behavior leading up to the move, and that the mean value is 0 and variance is finite. The mean and variance of the set of moves is independent of dt.
What is the importance of this? Essentially this—the probability distribution of the motion can be shown, as dt approaches 0, to be normal or Gaussian.
Ito's theorem provides an analytical formula that simplifies the treatment of stochastic differential equations, which is why it is so valuable. It is an important rule in the application of stochastic calculus to the pricing of financial instruments. Here we briefly describe the power of the theorem.
The standard stochastic differential equation for the process of an asset price St is given in the form:
where a(St, t) is the drift coefficient and b(St, t) is the volatility or diffusion coefficient. The Weiner process is denoted dWt and is the unpredictable events that occur at time intervals dt. This is sometimes denoted dZ or dz.
Consider a function f(St, t) dependent on two variables S and t, where S follows a random process and varies with t. If St is a continuous‐time process that follows a Weiner process Wt, then it directly influences the function f( ) through the variable t in f(St, t). Over time, we observe new information about Wt as well the movement in S over each time increment, given by dSt. The sum of both these effects represents the stochastic differential and is given by the stochastic equivalent of the chain rule known as Itô's lemma. So for example, if the price of a stock is 30 and an incremental time period later is 30.5, the differential is 0.5.
If we apply a Taylor expansion in two variables to the function f(St, t) we obtain:
Remember that ∂t is the partial derivative while dt is the derivative.
If we substitute the stochastic differential equation (5.25) for St, we obtain Itô's lemma of the form:
What we have done is taken the stochastic differential equation (“SDE”) for St and transformed it so that we can determine the SDE for ft. This is a valuable mechanism by which we can obtain an expression for pricing derivatives that are written on an underlying asset whose price can be determined using conventional analysis. In other words, using Ito's formula enables us to determine the SDE for the derivative, once we have set up the SDE for the underlying asset. This is the value of Itô's lemma.
The SDE for the underlying asset St is written in most textbooks in the following form:
which has denoted the drift term a(St, t) as μSt and the diffusion term (St, t) as σSt In the same way, Itô's lemma is usually seen in the form:
although the noise term is sometime denoted dZ. Further applications are illustrated in Example 5.2(i).
A variable (such as an asset price) may be assumed to have a lognormal distribution if the natural logarithm of the variable is normally distributed. Therefore, if an asset price S follows a stochastic process described by:
how would we determine the expression for ln S? This can be achieved using Itô's lemma.
If we say that F = ln S, then the first derivative:
and as there is no t we have .
The second derivative is .
We substitute these values into Itô's lemma given at (5.29) and this gives us:
So we have moved from dF to dS using Itô's lemma, and (5.31) is a good representation of the asset price over time.
The continuously compounded gross redemption yield at time t on a default‐free zero‐coupon bond that pays £1 at maturity date T is x. We assume that the movement in x is described by:
where a, α and s are positive constants. What is the expression for the process followed by the price P of the bond? Let us say that the price of the bond is given by:
We have dx, and we require dP. This is done by applying Itô's lemma. We require:
From Itô's lemma:
which gives:
which simplifies to:
Therefore, using Itô's lemma we have transformed the SDE for the bond yield into an expression for the bond price.
Whilst in no way wishing to trivialize the mathematical level, we will not consider the derivations here, but simply state that the observed values of the Brownian motion up to the point at time t determine the process immediately after, and that this process is Gaussian. Stochastic integrals are continuous path martingales. As described in Neftci (1996), the integral is used to calculate sums where we have an infinite or uncountable number of items, in contrast with the Σ sum operator, which is used for a finite number of objects. In defining integrals, we begin with an approximation, where there is a countable number of items, and then set a limit and move to an uncountable number. A basic definition is given in Appendix 5.4. Stochastic integration is an operation that is closely associated with Brownian paths—a path is partitioned into consecutive intervals or increments, and each increment is multiplied by a random variable. These values are then summed to create the stochastic integral. Therefore, the stochastic integral can be viewed as a random walk Brownian motion with increments that have varying values, a random walk with non‐homogeneous movement.
It is possible to generalize Itô's formula in order to produce a multidimensional formula, which can then be used to construct a model to price interest rate derivatives or other asset class options where there is more than one variable. To do this, we generalize the formula to apply to situations where the dynamic function f(.) is dependent on more than one Itô process, each expressed as standard Brownian motions.
Consider where are independent standard Brownian motions and WT is an n‐dimensional Brownian motion. We can express Itô's formula mathematically with respect to p Itô processes as:
Where the function f(.) contains second‐order partial derivatives with respect to x and first‐order partial derivatives with respect to t, which are a continuous function in (x, t), the generalized Itô formula is given by:
with:
Reproduced from Analysing and Interpreting the Yield Curve (2004)
This extract from Analysing and Interpreting the Yield Curve (2004)
Bond market participants take a keen interest in both cash and zero‐coupon (spot) yield curves. In markets where an active zero‐coupon bond market exists, much analysis is undertaken into the relative spreads between derived and actual zero‐coupon yields. In this chapter we review some of the yield curve analysis used in the market, with respect to bonds that are default‐free, such as US and UK government bonds. We then look at specific case study examples in the next (and final) chapter.
Market‐makers in government bond markets analyze various factors in the market in deciding how to run their book. Customer business apart, decisions to purchase or sell securities is a function of their views on:
All three areas are related but react differently to certain pieces of information. A report on the projected size of the government's budget deficit for example, will not have much effect on two‐year bond yields, whereas if the expectations come as a surprise to the market it may have an adverse effect on long‐bond yields. The starting point for analysis is the yield curve, both the traditional coupon curve plotted against duration and the zero‐coupon curve. Figure 12.1 illustrates the traditional yield curve for gilts in October 1999.
For a first‐level analysis, many market practitioners go no further than Figure 12.1. An investor who has no particular view on the future shape of the yield curve or the level of interest rates may well adopt a neutral outlook and hold bonds that have a duration that matches their investment horizon. If they believe interest rates are likely to remain stable for a time, they might hold bonds with a longer duration in a positively‐sloping yield curve environment, and pick up additional yield but with higher interest rate risk. Once the decision has been made on which part of the yield curve to invest in or switch in to, the investor must decide on the specific securities to hold, which then brings us on to relative value analysis. For this, the investor analyzes specific sectors of the curve, looking at individual stocks. This is sometimes called looking at the “local” part of the curve.
An assessment of a local part of the yield curve includes looking at other features of individual stocks in addition to their duration. This recognizes that the yield of a specific bond is not only a function of its duration, and that two bonds with near‐identical duration can have different yields. The other determinants of yield are liquidity of the bond and its coupon. To illustrate the effect of coupon on yield, let us consider Table 12.1. This shows that, where the duration of a bond is held roughly constant, a change in coupon of a bond can have a significant effect on the bond's yield.
Table 12.1 Duration and yield comparisons for bonds in a hypothetical inverted curve environment, October 1999.
Coupon | Maturity | Duration | Yield |
8% | 20‐Feb‐02 | 1.927 | 5.75% |
12% | 5‐Feb‐02 | 1.911 | 5.80% |
10% | 20‐Jun‐10 | 7.134 | 4.95% |
6% | 1‐Jul‐10 | 7.867 | 4.77% |
In the case of the long bond, an investor could, under this scenario, both shorten duration and pick up yield, which is not the first thing that an investor might expect. However, an anomaly of the markets is that, liquidity issues aside, the market does not generally like high coupon bonds, so they usually trade cheap to the curve.
The other factors affecting yield are supply and demand, and liquidity. A shortage of supply of stock at a particular point in the curve has the affect of depressing yields at that point. A reducing public sector deficit is the main reason why such a supply shortage might exist. In addition, as interest rates decline, for example, ahead of or during a recession, the stock of high coupon bonds increases, as the newer bonds are issued at lower levels, and these “outdated” issues can end up trading at a higher yield. Demand factors are driven primarily by the investor's views of the country's economic prospects, but also by government legislation, for example the Minimum Funding Requirement in the UK compelled pension funds to hold a set minimum amount of their funds in long‐dated gilts, which had the effect of permanently keeping demand high.1
Liquidity often results in one bond having a higher yield than other, despite both having similar durations. Institutional investors prefer to hold the benchmark bond, which is the current two‐year, five‐year, ten‐year or thirty‐year bond and this depresses the yield on the benchmark bond. A bond that is liquid also has a higher demand, thus a lower yield, because it is easier to convert into cash if required. This can be demonstrated by valuing the cash flows on a six‐month bond with the rates obtainable in the Treasury bill market. We could value the six‐month cash flows at the six‐month bill rate. The lowest obtainable yield in virtually every market2 is the T‐bill yield, therefore valuing a six‐month bond at the T‐bill rate will produce a discrepancy between the observed price of the bond and its theoretical price implied by the T‐bill rate, because the observed price will be lower. The reason for this is simple, because the T‐bill is more readily realizable into cash at any time, it trades at a lower yield than the bond, even though the cash flows fall on the same day.
We have therefore determined that a bond's coupon and liquidity level, as well as its duration, will affect the yield at which it trades. These factors can be used in conjunction with other areas of analysis, which we look at next, when deciding which bonds carry relative value over others.
As many readers would have concluded, the yield versus duration curve illustrated in Figure 12.1 is an ineffective technique with which to analyze the market.
This is because it does not highlight any characteristics of the yield curve other than its general shape and this does not assist in the making of trading decisions. To facilitate a more complete picture, we may wish to employ the technique described here. Figure 12.2 shows the bond par yield curve3 and T‐bill yield curve for gilts in October 1999. Figure 12.3 shows the difference between the yield on a bond with a coupon that is 100 basis points below the par yield level, and the yield on a par bond. The other curve in Figure 12.3 shows the level for a bond with a coupon that is 100 basis points above the par yield. These two curves show the “low coupon” and “high coupon” yield spreads. Using the two figures together, an investor can see the impact of coupons, the shape of the curve and the effect of yield on different maturity points of the curve.
Constructing a zero‐coupon yield curve provides the framework within which a market participant can analyze individual securities. In a government bond market, there is no credit risk consideration (unless it is an emerging market government market), and therefore no credit spreads to consider. There are a number of factors that can be assessed in an attempt to identify relative value.
The objective of much of the analysis that occurs in bond markets is to identify value, and identify which individual securities should be purchased and which should be sold. At the overview level, this identification is a function of whether one thinks interest rates are going to rise or fall. At the local level though, the analysis is more concerned with a specific sector of the yield curve, whether this will flatten or steepen, whether bonds of similar duration are trading at enough of a spread to warrant switching from one into another. The difference in these approaches is one of identifying which stocks have absolute value, and which have relative value. A trade decision based on the expected direction of interest rates is based on assessing absolute value, whether interest rates themselves are too low or too high. Yield curve analysis is more a matter of assessing relative value. On (very!) rare occasions, this process is fairly straightforward, for example if the three‐year bond is trading at 5.75% when two‐year yields are 5.70% and four‐year yields are at 6.15%, the three‐year appears to be overpriced. However, this is not really a real‐life situation. Instead, a trader might assess the relative value of the three‐year bond compared to much shorter‐ or longer‐dated instruments. That said, there is considerable difference between comparing a short‐dated bond to other short‐term securities and comparing say, the two‐year bond to the thirty‐year bond. Although it looks like it on paper, the space along the x‐axis should not be taken to imply that the smooth link between one‐year and five‐year bonds is repeated from the five‐year out to the thirty‐year bonds. It is also common for the very short‐dated sector of the yield curve to behave independently of the long end.
One method used to identify relative value is to quantify the coupon effect on the yields of bonds. The relationship between yield and coupon is given by (12.1):
where;
rm | is the yield on the bond being analyzed; |
rmP | is the yield on a par bond of specified duration; |
CPD | is the coupon on an arbitrary bond of similar duration to the part bond. |
and c and d are coefficients. The coefficient c reflects the effect of a high coupon on the yield of a bond. If we consider a case where the coupon rate exceeds the yield on the similar‐duration par bond (CPD > rmP), (12.1) reduces to (12.2):
Equation (12.2) specifies the spread between the yield on a high coupon bond and the yield on a par bond as a linear function of the spread between the first bond's coupon and the yield and coupon of the par bond. In reality this relationship may not be purely linear, for instance the yield spread may widen at a decreasing rate for higher coupon differences. Therefore (12.2) is an approximation of the effect of a high coupon on yield where the approximation is more appropriate for bonds trading close to par. The same analysis can be applied to bonds with coupons lower than the same‐duration par bond.
The value of a bond may be measured against comparable securities or against the par or zero‐coupon yield curve. In certain instances the first measure may be more appropriate when for instance, a low coupon bond is priced expensive to the curve itself but fair compared to other low coupon bonds. In that case, the overpricing indicated by the par yield curve may not represent unusual value, rather a valuation phenomenon that is shared by all low coupon bonds. Having examined the local structure of a yield curve, the analysis can be extended to the comparative valuation of a group of similar bonds. This is an important part of the analysis, because it is particularly informative to know the cheapness or dearness of a single stock compared to the whole yield curve, which might be somewhat abstract. Instead we may seek to identify two or more bonds, one of which is cheap and the other dear, so that we might carry out an outright switch between the two, or put on a spread trade between them. Using the technique we can identify excess positive or negative yield spread for all the bonds in the term structure. This has been carried out for our five gilts, together with other less liquid issues as at October 1999 and the results are summarized in Table 12.2.
Table 12.2 Yields and excess yield spreads for selected gilts, 22 October 1999.
Coupon | Maturity | Duration | Yield % | Excess yield spread (bp) |
8% | 07/12/2000 | 1.072 | 5.972 | −1.55 |
10% | 26/02/2001 | 1.2601 | 6.051 | 4.5 |
7% | 07/06/2002 | 2.388 | 6.367 | −1.8 |
5% | 07/06/2004 | 4.104 | 6.327 | −3.8 |
6.75% | 26/11/2004 | 4.233 | 6.351 | 2.7 |
5.75% | 07/12/2009 | 7.437 | 5.77 | −4.7 |
6.25% | 25/11/2010 | 7.957 | 5.72 | 1.08 |
6% | 07/12/2028 | 15.031 | 4.77 | −8.7 |
From Table 12.2 as we might expect the benchmark securities are all expensive to the par curve, and the less liquid bonds are cheap. Note that the 6.25% 2010 appears cheap to the curve, but the 5.75% 2009 offers a yield pick‐up for what is a shorter‐duration stock—this is a curious anomaly and one that disappeared a few days later.4
Table 12.2 Bond basis point value, 22 October 1999.
Coupon | Maturity | Duration | Yield % | Price | BPV |
8% | 07/12/2000 | 1.072 | 5.972 | 102.17 | 0.01095 |
10% | 26/02/2001 | 1.2601 | 6.051 | 105.01 | 0.01880 |
7% | 07/06/2002 | 2.388 | 6.367 | 101.5 | 0.02410 |
5% | 07/06/2004 | 4.104 | 6.327 | 94.74 | 0.03835 |
6.75% | 26/11/2004 | 4.233 | 6.351 | 101.71 | 0.03980 |
5.75% | 07/12/2009 | 7.437 | 5.77 | 99.84 | 0.07584 |
6.25% | 25/11/2010 | 7.957 | 5.72 | 104.3 | 0.07526 |
6% | 07/12/2028 | 15.031 | 4.77 | 119.25 | 0.17834 |
In the earlier section on futures trading, we introduced the concept of spread trading, which is not market‐directional trading, but rather the expression of a view point on the shape of a yield curve, or more specifically the spread between two particular points on the yield curve. Generally, there is no analytical relationship between changes in a specific yield spread and changes in the general level of interest rates. That is to say, the yield curve may flatten when rates are both falling or rising, and equally may steepen under either scenario as well. The key element of any spread trade is that it is structured so that a profit (or any loss) is made only as a result of a change in the spread, and not due to any change in overall yield levels. That is, spread trading eliminates market‐directional or first‐order market risk.
Table 12.3 shows data for our selection of gilts but with additional information on the basis point value (BPV) for each point. This is also known as the “dollar value of a basis point” or DV01.
If a trader believed that the yield curve was going to flatten, but had no particular strong feeling about whether this flattening would occur in an environment of falling or rising interest rates, and thought that the flattening would be most pronounced in the two‐year versus ten‐year spread, they could put on a spread consisting of a short position in the two‐year and a long position in the ten‐year. This spread must be duration‐weighted to eliminate first‐order risk. At this stage we must point out, and it is important to be aware of, the fact that basis point values, which are used to weight the trade, are based on modified duration measures. From an elementary understanding of bond maths we know that this measure is an approximation, and will be inaccurate for large changes in yield. Therefore the trader must monitor the spread to ensure that the weights are not going out of line, especially in a volatile market environment.
To weight the spread, the trader should use the ratios of the BPVs of each bond to decide on how much to trade. Assume that the trader wants to purchase £10 million of the ten‐year. In this case, he must sell ((0.07584/0.02410) × 10,000,000) or £31,468,880 of the two‐year bond. It is also possible to weight a trade using the bonds' duration values, but this is rare. It is common practice to use the BPV.
The payoff from the trade depends on what happens to the two‐year versus ten‐year spread. If the yields on both bonds move by the same amount, there will be no profit generated, although there will be a funding consideration. If the spread does indeed narrow, the trade will generate profit. Note that disciplined trading calls for both an expected target spread as well as a fixed time horizon. So for example, if the current spread is 59.7 basis points, the trader may decide to take the profit if the spread narrows to 50 basis points, with a three‐week horizon. If, at the end of three weeks, the spread has not reached the target, the trader should unwind the position anyway, because that was their original target. On the other hand, what if the spread has narrowed to 48 basis points after one week and looks like narrowing further—what should the trader do? Again, disciplined trading suggests the profit should be taken. If contrary to expectations, the spread starts to widen, if it reaches 64.5 basis points the trade should be unwound, this “stop‐loss” being at the half‐way point of the original profit target.
The financing of the trade in the repo markets is an important aspect of the trade, and will set the trade's break‐even level. If the bond being shorted (in our example, the two‐year bond) is special, this will have an adverse impact on the financing of the trade. The repo considerations are reviewed in Choudhry (2002).
A bond spread has two fundamental characteristics; in theory there should be no profit or loss effect due to a general change in interest rates, and any profit or loss should only occur as a result of a change in the specific spread being traded. Most bond spread trades are yield curve trades where a view is taken on whether a particular spread will widen or narrow. Therefore it is important to be able to identify which sectors of the curve to sell. Assuming that a trader is able to transact business along any part of the yield curve, there are a number of factors to consider. In the first instance, the historic spread between the two sectors of the curve. To illustrate in simplistic fashion, if the 2–10 year spread has been between 40 and 50 basis points over the last six months, but very recently has narrowed to less than 35 basis points, this may indicate imminent spread widening. Other factors to consider are demand and liquidity for individual stocks relative to others, and any market intelligence that the trader gleans. If there has been considerable customer interest in certain stocks relative to others, because investors themselves are switching out of certain stocks and into others, this may indicate a possible yield curve play. It is a matter of individual judgement.
An historical analysis requires that the trader identifies some part of the yield curve within which he expects to observe a flattening or steepening. It is, of course, entirely possible that one segment of the curve will flatten while another segment is steepening, in fact this scenario is quite common. This reflects the fact that different segments respond to news and other occurrences in different ways.
A more exotic type of yield curve spread is a curvature trade. Let us consider for example, a trader who believes that three‐year bonds will outperform on a relative basis, both two‐year and five‐year bonds. That is, he believes that the two‐year/three‐year spread will narrow relative to the three‐year/five‐year spread, in other words that the curvature of the yield curve will decrease. This is also known as a butterfly/barbell trade. In our example, the trader will buy the three‐year bond, against short sales of both the two‐year and the five‐year bonds. All positions are duration‐weighted.
Reproduced from Analysing and Interpreting the Yield Curve (2004)
The construction of an internal yield curve is one of the major steps in the implementation of the internal funds pricing or “funds transfer pricing” (FTP) system in a bank. It should be done after the Trading book and the Banking book are split, but before internal funding methodologies for all balance sheet items and the internal bank result calculation methodology are approved and implemented in the balance sheet management information.
As is pointed out in regulatory recommendations, “the transfer prices should reflect current market conditions as well as the actual institution‐specific circumstances”2. Putting it simply, that means the curve should contain a market component responsible for interest rate risk and a bank‐specific spread over the market which will reflect the term liquidity cost for the bank.
There are several approaches for construction of an FTP curve. The choice of the approach depends on the scale of your bank, as well on the market where the bank operates. These main approaches are:
We consider each of these approaches in turn. We also show in the Appendix the ALCO submission prepared by one of the authors on implementing an internal curve methodology, when he was working in the investment banking division of a global multinational bank.
As implied by the name and definition of the approach, both components of an FTP curve should be derived from market quotes and indicators. The following is a list of requirements defining the characteristics of these market indicators:
Usually, as market indicators for FTP curve construction we understand Libor (Euribor, etc.), interest rate swap (IRS), cross‐currency swap (CCS), and credit default swap (CDS). Although as CDS is not available for each bank and sometimes is not representative when it is available, bond yields as market quotes can also be used. An FTP curve based on bond yields can be constructed using several methods which are deeply analysed in the special section. (See Application of Ordinary Least Squares method and Nelson‐Siegel family approaches.)
The market approach can be applied by large banks which have a lot of operations in financial markets. The deep and developed market of financial instruments (including derivatives) is required.
The advantages of this approach are:
Although this method also has limitations:
The topic of constructing a yield curve was widely investigated in scientific literature and by practitioners of central banks. The reason for such a deep investigation for these researches was that each country which has government bonds needs to construct a realistic, trustworthy, and flexible zero‐coupon yield curve to reflect the level of the country's debt cost. Such reasons and the grade of responsibility for fair curve construction do not stop debates around the best methods and approaches. However not all approaches are feasible for use in ALM applications.
What are the main criteria for choosing the optimal approach? This approach should be simple enough to be executed without complicated technical packages, but at the same time reflect the market of a particular country, even when bond quotes are not available for all tenors.
According to a Bank for International Settlements survey about zero‐coupon yield curve estimation procedures at central banks (2005), the most commonly used approaches are spline‐based (for example, McCulloch, 1971, 1975) and parametric methods (Nelson, Siegel, 1987 and Svensson, 1994). Although the spline methods gain more positive assessments due to their provision of smoothness and accuracy, they are more complex. At the same time, Nelson and Siegel state in their work that their objective is simplicity rather than accuracy. The following part of this section is devoted to comparison and contrasting of parametric approaches and their practical implementation.
All the parametric methods are based on minimisation of the price / yield errors. Among parametric methods the most basic approach is the Ordinary Least Squares (OLS) method. This is the simplest parametric method which applies calculation of squares of deviations of actual quotes from the calculated approximated values and minimisation of their sum. Formula used to construct the curve is , where x – duration, f(x) – yield, parameters a, b, and c should satisfy the following: sum of the squares of deviations from the mean should be minimal.
The implementation in practice consists of three steps:
The advantages of applying this method for ALM purposes are its simplicity and transparency. However, there are disadvantages which can significantly impact the result. The first and the main one is that the function which is used for approximation is parabolic – and, thus, is increasing in its first half, but after the extremum it starts to decline (see Figure 3.1). It's evident that this is not the best function to describe the market structure of interest rates – at least due to the reason that according to the expectations hypothesis theory, in the long run the rates tend to increase.
This method can, though, be used by an ALM unit of a bank in the following cases: a) the market does not have long tenors (so the declining part of the curve won't be used); b) the market yield curve is increasing (that means that it is not inverted and does not have troughs); c) the ALM unit doesn't have resources to try to implement any more complicated method (usually in small banks).
Figure 3.1 shows an example of the difference in output of the N‐S and OLS methodologies when using the same inputs.
In order to make the curve more flexible Nelson‐Siegel family curves are used.
The first approach was suggested by Nelson and Siegel (basic N‐S). They use four parameters in the equation to describe the yield curve:
where
β0 | is a long‐term interest rate; |
β1 | represents the spread between short‐term and long‐term rates (this parameter defines the slope of the curve: if parameter is positive, then the slope is negative). – is the starting point of the curve at the short end; |
β2 | is the difference between the middle‐term and the long‐term rates, defining the hump of the curve (if then the hump is observed in period ô, if then the curve will have an “U”‐shape); |
τ | is a constant parameter, representing the tenor at which the maximum of the hump is achieved. |
The main difference between the basic N‐S approach and the OLS method is the addition of dependence between short‐term and long‐term rates, which tend to make the curve more flexible. This model assumes that long‐term rates directly impact the short‐term rates and, thus, some segments of the curve can't change while other segments are stable. Moreover, this basic approach provides a curve with only one hump or trough. And that is not always true in practice.
Svensson suggested an extension to basic N‐S approach and implemented an additional component to provide a better description of the first part of the curve:
Thus the curve can have two extremums, β3 determines the size and the form of the second hump, τ2 specifies the tenor for the second hump.
The advantages of the Svensson method in comparison to OLS and basic N‐S approaches are even better flexibility and better accuracy. However even two humps do not perfectly reflect the market curve. That is why in practice further adjustments were made.
One of them is the adjusted Nelson‐Siegel (adjusted N‐S) approach with a set of seven parameters. The vector of curve's parameters is recalculated after each new bond / new quote is added. The first four parameters (βs) are responsible for the level of yields on short‐term, middle‐term, and long‐term segments of the yield curve (the shifts up and down), and the remaining three parameters (τs) are responsible for convexity / concavity of the appropriate segments.
Such adjustment provides even more advantages: such flexibility that the yield curve by adjusted N‐S approach can have all types of forms: monotonous increasing, or decreasing, convex, U‐form, or S‐form; memory, because the calculation is based on the previous parameters, so additional data doesn't change the form suddenly. This may be especially useful for markets with low liquidity of some bond issues – when on one day bonds are traded, there's a quote and the bond is included in the calculation, but on another day they are not traded, there's no quote and, thus, there's no input for calculation.
The disadvantage of all types of Nelson‐Siegel approaches is their complexity in comparison with the OLS method. As far as one needs to assign initial values to parameters and then apply the “goal seek” function – there's risk that at this moment a mistake is made and further calculations are incorrect.
Nevertheless, one of the Nelson‐Siegel family approaches would be recommended to be used for ALM purposes in the following cases: a) for larger banks with bigger ALM units, equipped with automatic systems; b) on the markets where the curve is supposed to demonstrate several humps on different tenors; c) on bond markets with low liquidity.
To summarise, not all of the existing methods to construct a yield curve can be successfully applied by ALM units. Parametric approaches are most simple in their implementation. The choice between the most “plain vanilla” OLS method and more complicated Nelson‐Siegel family approaches should be done according to the size of the bank and market conditions and peculiarities.
An important reference yield curve today is the overnight index swap (OIS) curve. This is similar to the conventional swap curve except it refers to the overnight rate on the floating leg of the swap, compared to the 3‐month or 6‐month rate on the floating leg of a conventional swap.
Prior to the 2008 crash derivative pricing and valuations were based on the simple principle of the time value of money. A breakeven price was calculated such that when all the future implied cash flows were discounted back to today the Net Present Value of all the cash flows was zero. The breakeven price was used for valuations or spread by a bid–offer price to create a trading price. Bid–offers would generally incorporate a counterparty‐dependent mark‐up to cover various costs and a return on capital in a fairly ill‐defined way.
A breakeven price required the construction of a projection curve for the index being referenced, which could be a tenor of Libor, such as 3‐month, for 3‐month Libor swaps or an overnight rate in the case of overnight index swaps. The projection curve would use market observable inputs and various interpolation methodologies to derive market implied future rates used to project the future floating‐rate fixings. In other words:
Once the market implied future fixings are known, a discount curve was then used to discount cash flow mismatches in such a way that the breakeven fixed rate gives a result where all the cash flow mismatches have a present value of zero. Historically market practice was to discount cash flows using a Libor discount curve, the assumption being that all cash flow mismatches could be borrowed or reinvested at Libor. It was implicitly assumed that Libor was the risk‐free rate at which these cash flow mismatches could be funded.
Figure 3.2 shows the GBP OIS (known as SONIA) and GBP swap curves as at 13 October 2016. The OIS curve lies below the conventional swap curve because of the term liquidity premium (TLP) difference between the two tenors (the TLP is higher the longer the tenor).
In the case of a collateralised trade transacted under the terms of a CSA, the posting of collateral ensures that the NPV of a trade is always zero net of collateral held or posted. All future funding mismatches are therefore explicitly funded directly by the exchange of collateral and as a result do not require any external funding. The collateral remuneration rate is defined by the CSA and is normally the relevant OIS rate. As it is the relevant OIS rate that becomes the applicable rate for funding cash flow mismatches, market practice has evolved to discount future cash flows using OIS rates when pricing or valuing a collateralised trade, so‐called OIS discounting. It is impossible to determine when this market practice changed but it is generally accepted that when the London Clearing House (which clears derivative transactions) changed to OIS discounting in 2010 it was doing so to reflect best market practice.
Where multiple forms of collateral were permissible under a CSA, market practice evolved to take into account the embedded option for the collateral poster. Pricing assumed that a counterparty would always act rationally and post the cheapest to deliver collateral with the discount curve reflecting the relevant collateral OIS rate, so‐called CTD or CSA discounting. A further enhancement is often used where the collateral is non‐cash such as a government bond, in these instances the discount curve reflects the price at which the bonds can be traded in the repo market. For complex CSAs where collateral was multicurrency, the discount curve is often a multicurrency hybrid curve which reflects that the CTD may change at a future date.
For uncollateralised trades, not only was the use of a Libor rate to discount cash flow mismatches clearly inappropriate given the increase in funding spreads during the 2008 crisis but also, in the absence of collateral, the NPV of an uncollateralised trade represented a potential funding requirement. Pricing for uncollateralised trades now generally contains an upfront adjustment to the price to take into account the expected funding costs of non‐collateralisation referred to as the Funding Valuation Adjustment or FVA. FVA along with a Credit Valuation Adjustment CVA now represent a more rigorously defined part of the bid–offer price adjustment.
Collateral posted at the LCH has no optionality allowed – it must be in cash and in the currency of the transaction. Hence the swap screen price now reflects the price for a cleared swap at LCH. Because the collateral is unambiguous (cash in the currency of the trade) remunerated at the relevant OIS, so the discount / funding rate is always known and is not counterparty dependent. Anything other than a LCH cleared trade means the swap price is different to take into account the impact of the type of collateral. This is an important point: observable swap rates are for LCH cleared trades only.
The yield curve is the best snapshot of the state of the financial markets. It is not the sole driver of customer prices in banking, but it is the most influential. Hence it is important that all practitioners understand the behaviour of the curve and how to analyse and interpret it. Being aware of the relationship between spot rates, forward rates, and yield to maturity is also important. Ultimately, there should be no short‐cuts when it comes to understanding the yield curve.
In the example in the text the working is:
Par yeields | 0.05 | Df | 0.952381 | ||
0.0525 | 0.9026128 | ||||
0.0575 | 0.8447639 | Working | 0.1066621 | ||
0.8933379 | |||||
1.0575 |
With special thanks to Praveen Murthy, iflexsolutions, for assistance with the calculations.
3.147.140.206