Learning from interest rate implied volatilities

A relative value analysis of swaptions

Master Thesis/ Kandidat afhandling

Copenhagen Business School
Cand.merc.mat

Author:
Johan Jørgensen

Thesis supervisor:
Søren Bundgaard Brøgger

Date: 12. September 2017.

Pages: 77

1
 Abstract

No previous research has firmly been conducted on relative value analysis on

longer expiry European swaptions. This paper conducts an empirical analysis
underpinning the concept of relative value analysis for ATM European swaptions on
EUR and USD market by studying time series dataset of implied volatilities. We
investigate the EUR and USD market for 2010-2017 by applying a principal
component analysis (PCA) framework.

The thesis investigates longer expiry swaption straddles. All with a low
exposure towards changes in the underlying swap rate. This gives the opportunity to
focus on hedging implied volatility level rather than delta hedging. In the analysis
we have found three significant dynamics represented on the implied volatility
surface. The first dynamic are interpreted as an overall implied volatility level factor.
The two other dynamics captures the implied volatility curves, each specifying a
dimension on the implied volatility surface. Also, we give evidence that these three
dynamics on the implied volatility surface explain over 95 % of the total variance for
both USD and EUR markets. This is a consistent result over the entire period from
2010-2017. Furthermore, an investigation of the linkage between principal
component scores and economic variables shows a high correlation between the
first principal component (PC) and the US stock market.

To put these result into perspective, an unorthodox multiple regression model

for hedging PC’s are introduced. The model is constructed by different very liquid
asset classes as a cheap and effective hedge. Results show that this hedging
strategy is only efficient during subperiods when using a rolling 30 week data
window. Furthermore, an example of a relative value trade is illustrated. However
relative value trading opportunities are difficult to spot as these seems to occur
rarely between 2010 and 2017.

Lastly, modelling with PCA on implied volatility surface is discussed. Here we

introduce the major pitfalls of a PCA setup.

2
 Contents

Abstract .................................................................................................................... 2
Introduction .............................................................................................................. 5
What is relative value? ........................................................................................... 9
Negative interest rates .........................................................................................12
Interest rate derivatives ...........................................................................................14
Interest rate swaps ............................................................................................... 15
European swaptions .............................................................................................18
The Normal model ............................................................................................21
Risk measures under the normal model ........................................................... 22
Quoting swaptions ............................................................................................23
Choosing normal or log-normal ........................................................................25
Methodology .......................................................................................................... 26
Mean Reversion ................................................................................................... 26
Calibrating MR with MLE ...................................................................................... 27
Principal component analysis ...............................................................................30
Using PCA in R .................................................................................................. 33
Distribution of the volatility surface .....................................................................34
Data .........................................................................................................................36
Description ...........................................................................................................36
Empirical Results .....................................................................................................38
Analysis ................................................................................................................38
Interpretation of the eigenvectors ....................................................................41
Scaled eigenvalues........................................................................................... 48
Clustering ......................................................................................................... 51
Factor Residuals ................................................................................................52

3
 Correlation with market indicators ...................................................................... 54
Applying the setup .................................................................................................. 62
Hedging with very liquid assets ........................................................................ 62
Strategies for relative value trades ...................................................................... 66
First PC ............................................................................................................. 67
Second PC........................................................................................................ 68
Third PC ........................................................................................................... 68
Optimal holding period .................................................................................... 69
Hedging of factors ............................................................................................ 70
An example of a trading opportunity ................................................................ 71
Pitfalls of a PCA trade........................................................................................... 74
Conclusion ............................................................................................................... 76
Bibliography ............................................................................................................ 78
Appendix.................................................................................................................. 79
Derivation of solution for maximum likelihood estimates .................................... 79
Trading signals for USD and EUR .........................................................................81
Selected R code ....................................................................................................83
Code: MLE of OU process for third principal component ..................................83
Eigenvectors on vega sector for EUR................................................................... 84

4
Introduction

This thesis covers both statistical and financial theory. As this is an empirical
analysis, this paper will introduce concepts of mathematical finance and not dwell
on advanced mathematical techniques within option pricing theory. We investigate
implied volatilities for European swaptions from a statistical perspective applying
PCA conceptually. This is possible by restricting the implied volatility surface to
longer expiry options.

A swaption is an option on a swap, which is characterized by the expiry of the

option and the underlying swap rate (also called tenor). Looking exclusively on at-
the-money-forward swaptions an implied volatility surface can be produced be
applying expiry and tenor of the options into a three dimensional space.

In the literature an extensive documentation on identifying Principal

Components in term structure dynamics has been investigated, such as capturing
swap curve dynamics. Only a few papers document similar studies on volatility
dynamics for the interest rate swaption market. However this paper focuses on
longer expiry swaptions with low gamma by introducing a vega segmentation of
these options. Our interest lies within capturing the dynamics of the implied
volatility surface of swaptions.

At-the-money-forward swaption straddles are very liquid derivatives, especially

on the EUR and USD markets. These are often used for hedging and investment
purposes for institutions and corporate customers as well as banks and hedge funds.
They operate on the same market but mostly in different arrays of the implied
volatility surface. Several different types of market participants in the fixed income
markets have different regulatory jurisdictions, and as regulation differs across
participants this can create relative value opportunities for perhaps some of these
participants.

5
 In September 1998, the widely known relative value hedge fund Long Term
Capital Management (LTCM) collapsed. LTCM defaulted on liquidity having highly
leveraged positions for example on swaptions.

Volatility can be viewed as an asset class and while volatility have a tendency to
being negative correlated to asset returns, investors seeks opportunities in the
volatility class to hedge their downside risk cheaply in periods of volatile markets.
Political unrest in the Western part of the world such as Brexit and presidential
election of Donald Trump in the United States of America has created even more
uncertainty surrounding different financial markets.

When discussing volatility one should often be more specific as this can refer to
different terminologies. Volatility can either be realized or implied. Implied volatility
is the anticipated future volatility in a market, meanwhile realized volatility can be
defined as the actual volatility observed in the market until expiry of the option. This
paper concentrates on options mainly determined by the implied volatility. Changes
in the implied volatility are mainly driven by supply and demand for volatility. An
example could be a large pension fund actively buying swaptions as a part of a
hedging strategy, which could increase volatility for the traded swaptions.

Commonly the market quotes swaptions in reference to its implied volatility

instead of its price. By quoting prices in implied volatility, effects regarding
parameters not related to volatility are removed, and henceforth allows for
assessing prices across different swaptions. Implied volatility can be derived from
the Black model which assumes lognormal distributed rates and therefore only
allows rates to be positive. Although it can be argued that normal distributed rates
should be applied instead, as rates are allowed to be negative. This is one of the
motivations for this paper.

In the paper by Frankena (2016) log-normal volatilities tend to present jumps

and variance considerably when strike level of interest rates approaches zero or
negative entries, while normal volatility is observed to be more stable in the same
interest rate environment.

Longstaff et al. (2001) provide empirical evidence of relative value analysis of

caps and swaptions through a string market model setup. They find that the relative
valuations of most swaptions in the period of 1992 to 1999 are fairly priced.
However after the announcement of crash of LTCM in 1998, they find strong

6
evidence of longer dated swaption being significantly undervalued relatively to
other swaptions in a 12 weeks period.

Principal component analysis can be used as a dimension reducing tool. A PCA

analysis can therefore create a great overview of changes in swaptions with
different tenors and expiries for a longer period. PCA has proven to be a useful tool
when applying it on interest rates, implied volatilities, swap spreads, equities and
several commodities. Evidence of mean reversion has been observed for various
financial variables over longer time periods, such as interest rates, butterfly spreads,
spreads between gold and silver, implied volatilities etc.

In our analysis, ATMF swaption data are used to compute three principal
components, where we find that during the selected period at least 95 % of total
variance is explained by these three factors for the entire period. We will interpret
the principal components by investigating each component’s eigenvectors.
Moreover we will show clusters in the data to explain a differential in classification
of instruments in a vega or gamma segment controlled by the instruments expiries.

Specifically, a PCA will be performed to follow up on these research questions:

1.1. Can we explain the most important dynamics of the implied volatility surface
for swaption by implementing a principal component analysis for all tenors and
expiries?

1.2. Is there any mispricing in the interest rate option market when looking at
relative value analysis and can a mispricing be a trading opportunity?

1.3. Are we able to construct a hedge against directional exposure on the volatility
surface with cheap and liquid assets, such as equity, bonds, FX rates and commodities?

The structure of the thesis is constructed as followed:

Section 1 begin with a description of the concept of relative value analysis and
discuss parallels to asset pricing theory. After this we follow up on the transition to
negative rates markets. Then we will go through the fundamentals for swaps and
swaptions. Section 2 will describe the Ornstein-Uhlebeck mean reversion process
and the derivation of the maximum likelihood estimation for the parameters in this
process. This is followed by an explanation of the methodology for a principal
component analysis, as these are the ground tools for this relative value analysis. In
7
the end of this section a segmentation of the volatility surface is styled for the
purpose of restricting the analysis to the swaptions that are most influenced by
vega. Section 3 describes the data input, while section 4 gives an overview of the
application of the principal component analysis and how to interpret the results.
Section 5 introduces a hedging strategy reducing the exposure against shift changes
in the implied volatility level of the implied volatility surface. As a closing section of
section 5, we explore a rare relative value opportunity but also encounter a trade
with a major pitfall within the relative value analysis framework. Lastly, section 6
concludes on the analysis and the three research questions.

8
What is relative value?1
Concept

The concept of relative value is based on quantitative analysis comparing two

assets on a financial market. A relative value strategy directly involves finding
relatively mispriced securities caused by dislocations and anomalies in a market,
which eventually will revert to its fair value. This is a strategy used by several market
participants on various markets such as for equity, fixed income, credit etc.

Proposition 1

If two securities have identical payoffs in every future state of the world, then they
should have identical prices today.2

If this statement3 is violated the existence of arbitrage opportunities is present.

Arbitrage opportunities are not consistent with the theory of equilibriums in
financial markets. Today proposition 1 is an established part of financial theory.
However in 1997 Myron Scholes and Robert Merton won a Nobel Prize in economics
by applying this result in their work on valuating options. Myron Scholes and Fisher
Black used this proposition to determine the value of options by creating a self-
financing portfolio which dynamically replicated the payoff of an option. Through
the valuation of the self-financing portfolio they could determine the value of a
specific option.

Proposition 2

If two securities present investors with the identical risk, they should offer identical
expected returns.

This is another important proposition used for defining the relative value
component. If two assets states identical risks, the expected return of these should
be identical. This statement is in fact more complex to prove than the first
statement. However it can be proved by using the Arbitrage Pricing Theory (APT),
where unobservable linear factors are drivers of return.

1
Fixed Income Relative Analysis by Dough Huggins Christian Schaller 2013
2
Stated in Fixed Income Relative Analysis by Dough Huggins Christian Schaller 2013
3
The law of one price.
9
 Our first proposition says if two assets have the same payoff in all future states
they should have the same price. If this isn’t the case an arbitrage opportunity
prevails. An arbitrage opportunity can be viewed as a window of opportunity for a
trading strategy where we have a guaranteed profit without any risk taking, also in
popular terms called a free-lunch. As we stated above, the existence of an arbitrage
opportunity is inconsistent with equilibrium pricing of financial assets. Nevertheless,
arbitrage opportunities can be used to identify relative value opportunities.

Clearly we have two propositions that assume no-arbitrage and we use these
models to identify relative value opportunities. So why do people search for these
opportunities when models assume no arbitrage? This can be explained through
two minor logical observations.

Firstly arbitrage opportunities rarely occur due to the fact that market
participants are looking at the same opportunities. However, if we did have
consistent non-arbitrage markets, no one would be searching for profitable
opportunities or inconsistencies.

Secondly in practice arbitrage opportunities always carry some risk. Explaining

this in a theoretical and simple manner, we observe the relation between bond
prices, bond futures prices and repo rates. If a bond future is too expensive, we can
exploit this by selling the contract, and then buy the bond by borrowing the funds in
the repo market and setting the bond as collateral for the loan. The bond is returned
at expiry by the counterparty in the repo market. This party can now deliver the
bond into future contracts. This is a riskless arbitrage opportunity in theory, but
these positions are associated with risk when done in practice. One of the most
significant risks may be that the counterparty from the repo market fail to deliver
the bond at expiry. This will make it difficult for the party to deliver the bond into
future contracts. If the delivery is uncompleted this can lead to penalties that can be
costly. Therefore this should be a risk factor that needs to be considered before
entering this strategy in practice.

Insight

Relative value analysis can be viewed as a way of gaining insight into

relationships between different financial instruments but also to develop an
understanding of which market forces that drives the prices of different instruments

10
and how different markets are interconnected. Relative value analysis has its origin
in arbitrage trading but has a much wider spectrum of applications. As it can identify
reason for a security priced in a certain manner, expose the source of certain
relationships and compare the relative value pricing of one financial instrument
against the price of other instruments.

Applications

The applications of relative value analysis scopes within numerous areas, but
is mostly used for trading and hedging purposes. Within the trading aspect, one can
identify rich and cheap value of securities. As securities can become even richer or
cheaper a trader must understand the reasoning for why these securities are rich or
cheap to form a reasonable expectation of future richening and cheapening of these
securities.

From a hedging or immunization perspective, relative value analysis

considers hedging or immunizing a position against several risk exposures. As an
example a flow trader could improve the expected risks of a portfolio by considering
hedging alternatives from a relative value analysis of German Bunds. If the trader
believes cash Bunds are going to cheapen compared to other alternatives, a hedge
could be created by selling cash Bunds. While if the trader expects Bund futures are
going to cheapen relative to cash Bunds the hedge can be constructed through
future contracts instead of using cash Bunds.

Some other applications worth mentioning is that relative value analysis can
express a macro view, signal the timing of unwinding a position or help investment
managers on security selection to increase alpha of a portfolio.

Risks

When applying relative value when engaging in a financial market, several

risks needs consideration, such as:

 Liquidity risk - trading fixed income instruments often have different

liquidity. Some trading strategies involve taking a long position in a rather
illiquid security while taking a short position in a liquid security and hereby

11
 earning a liquidity premium. If financial distress should occur, it can be hard
to find liquidity in the market.
 Model risk needs to be considered as many fixed income instruments are
priced by models that can be misspecified or misapplied, where signaling
incorrect valuation can occur.
 Event risk can be viewed as market stress situations such as financial distress
or in extreme cases flight-to-quality events. These events can change market
perspectives and perhaps break historical correlations between similar
securities.
 Interest rate risk depends of course on the chosen strategy, but even market
neutral strategies can suffer from higher financing costs if interest rate starts
to rise to a higher level.
 Credit risk is also a risk factor to be considered as the counterparty can
default and not be able to fulfil the contract.
 Legal/regulative risks can destroy relationships between two instruments
due to perhaps new regulatory charges.

Summary

Overall, relative value analysis should be considered as a craftsmanship, as it is

neither science nor art. Applying relative value analysis in a market, knowledge of
statistical foundations and a less scientific related understanding of the market is
needed in mastering this craft.

Negative interest rates

In classical thinking of interest rate, it is expected that a lender would receive
interest of a borrowed amount. The same expectation would apply for a person
depositing funds onto a bank account. However after the financial crisis in 2008,
situations where lenders pay interest to borrowers and banks charging people for
their deposits are the scenario of 2017 and not a theoretical debate between experts
anymore on a world with negative interest rates.

12
 Still the concept of negative interest rates is not unfamiliar. Tracing back to the
1970s, the Swiss National Bank carried out an experiment on negative interest rates
for the main purpose of controlling capital inflow as an action of preventing the
Swiss Franc from appreciating. Today this is common ground for several developed
countries. ECB was the first to implement negative rates in June 2014 due to weak
growth and inflation in Euro area. In December 2014 Switzerland moved into a
negative rate environment as a result of managing upward pressure on the franc,
fear of deflation and weak growth. Other countries in Europe have also turned to
negative rate such as Sweden and Denmark. The central bank of Sweden,
Riksbanken, adopted negative interest rates in 2009 by lowering the deposit rate to
-0.25 % with the ongoing financial crisis. Bank of Japan has also implemented a
negative rate policy.

Motivation for implementing negative interest rate policies differs for each
central bank, but one overall significant rationale has been to improve the economy
and to control inflation.

In the Euro area, Sweden, Denmark, Switzerland – negative interest rate

environments affect valuation models of interest rate derivatives.

The Black model has for years been the main framework for option pricing, but
the rise of negative interest rate environments has marked multiple shortcomings of
the model for handling interest rate options. One key feature of the Black model is
that it assumes lognormal distributed rates, which only allows rates to be positive.

Adapting to the new normal of negative interest rates has created alternative
models by either constructing a shifted lognormal distributed model or a normal
distributed model. Lognormal-shifted Black volatilities have also been applied by
some market participants but there are no model consensuses regarding the precise
value of the constant shift parameter. Throughout this paper, we will use the normal
distributed model due to its properties within the negative rate environment.

13
Interest rate derivatives

xIBOR rates

xIBOR stands for x Interbank Offered Rate, hence x refers to a specific currency.
These rates are submitted by a group of prime banks each bank/business day at
11:00 GMT and vary with an expiry ranging from one business day to 12 months. The
fixing/reference rate is computed as an average the submitted bank rates. The
EURIBOR fixings are set by the European Banking Federation, while the reference
rate is the LIBOR for USD. The rates are should reflect the price in which prime
banks can loan money to each other.

Fixing methodology differs respectively, but all xIBOR fixings are quoted using
the money market convention. By this it can be concluded that interest paid on a
loan at expiry is calculated as 𝛿 × 𝑁 × 𝐿 where 𝛿 is the coverage4, N is the notional
and L the xIBOR rate. First, if 𝐷(𝑡, 𝑇) is the price of a zero coupon bond at time 0,
bought at time t and maturing at time T. Secondly, if 𝐹(0, 𝑇, 𝑇 + 𝛿 ) denotes the
forward xIBOR rate at time t, where we can loan funding between time T and 𝑇 + 𝛿.
Now by using the argument of no arbitrage, we can derive the forward xIBOR rates
as

1
1 + 𝛿𝐹(0, 𝑇, 𝑇 + 𝛿 ) = <=>
𝐷(𝑇, 𝑇 + 𝛿)

1 𝐷(0, 𝑇) − 𝐷(𝑇, 𝑇 + 𝛿)
𝐹(0, 𝑇, 𝑇 + 𝛿 ) = ⋅ <=>
𝛿 𝐷(𝑇, 𝑇 + 𝛿)
1 𝐷(0, 𝑇)
𝐹(0, 𝑇, 𝑇 + 𝛿 ) = ( − 1)
𝛿 𝐷(𝑇, 𝑇 + 𝛿)

Overnight index swap

The overnight index swap (OIS) is an interest rate swap, where a fixed rate is
exchanged over an agreed notional given the geometric average of an overnight

4
Coverage is seen as the year fraction expressed in years.
14
rate (fx. Fed Funds effective rate) for a chosen payment period. OIS discounting is
used for USD and EUR swaptions.

Interest rate swaps

Interest rate swaps is today by far one of the most traded derivatives. Bank of
International Settlement5 (BIS) reported a total amount of outstanding interest rate
swaps around 421 trillion USD with a estimated gross value of all contracts around
17 trillion USD. Swap contracts are traded as over-the-counter (OTC), this means
that the deal is directly between counterparties and not executed from an
exchange.

Valuation of an interest rate swap

A plain vanilla interest rate swap (IRS) an agreement/contract between two

parties to exchange their series of payments for a pre-agreed period – one of the
parties pays a fixed interest rate while the other party pays a floating interest rate,
also called a fixed-for-floating swap. This common swap contract contains therefore
of a fixed and a floating leg respectively. The party paying the floating rate has
entered into a receiver swap, while the counterparty has entered into a payer swap.

In abstract can a swap somehow be compared to a linear combination of

forward rate agreements (FRA’s). Whereas FRA’s are settled and fixed in advance,
IRS is only fixed in advance but paid-in-arrears. Here may an IRS also have different
day count conventions and payment schedules.

The floating leg is linked to some xIBOR rate fixed-in-advance and paid-in-
arrears6. Plain vanilla IRS market conventions for both currencies can be seen below
in table X.X.

Floating leg Fixed leg

Currency Index name Sport Start Roll Term Freq. Day count Freq Day count
USD USD Libor 2B MF 3M Q Act/360 S 30/360
EUR Euribor 2B MF 6M S Act/360 A 30/360

5
https://www.bis.org/statistics/dt21a21b.pdf
6
Fixed income Derivatives. M. Linderstrom
15
 We can find the present value of the floating leg by visualize start and end dates
𝑓𝑙𝑜𝑎𝑡 𝑓𝑙𝑜𝑎𝑡
of an IRS, denoted from 𝑇𝑆 to 𝑇𝐸 , and calculate a set of coverages 𝛿𝑆+1 , … , 𝛿𝐸 .
The floating leg present value can be obtained as

𝑃𝑉𝑡𝐹𝑙𝑜𝑎𝑡 = ∑𝐸𝑖=𝑆+1 𝛿𝑖𝐹𝑙𝑜𝑎𝑡 𝐹(𝑡, 𝑇𝑖−1 , 𝑇𝑖 )𝑁𝑖 𝐷(𝑡, 𝑇𝑖 ) (1)

Here 𝑁𝑖 is the notional at period i, 𝐷(𝑡, 𝑇𝑖 ) is the discount factor and

𝐹(𝑡, 𝑇𝑖−1 , 𝑇𝑖 ) is the forward xIBOR rate between period 𝑇𝑖−1 and 𝑇𝑖 .

We can derive the fixed leg of the IRS in a similar fashion. From the table above,
we can see the payment frequency and day count for the fixed leg. These
parameters don’t necessarily match the same parameters for the floating leg, so we
typically create a new set of coverages and dates due to the difference in payment
frequency and day count conventions. The leg still have the same start and end
𝑓𝑖𝑥𝑒𝑑 𝑓𝑖𝑥𝑒𝑑
date, but the coverages now looks like 𝛿𝑆+1 , … , 𝛿𝐸 . By letting 𝐾 represent the
fixed rate paid in the swap, we can write the present value of the fixed leg as

𝑃𝑉𝑡𝐹𝑖𝑥𝑒𝑑 = ∑𝐸𝑖=𝑆+1 𝛿𝑖𝐹𝑖𝑥𝑒𝑑 𝐾𝐷(𝑡, 𝑇𝑖 ) (2)

From the equation above, it is easy to see that the fixed leg is the fixed rate
payment with a discounting factor.

The present values of both legs in the swap are now obtained, hence now we
are able to calculate the value of swap by combining the two equations, (1) and (2).
Then referring to a swap contract, we focus on the fixed leg, therefore discussing a
payer swap, will refer to a party paying the fixed rate and receives floating rate
respectively. The present value of the a payer swap starting a 𝑇𝑆 and maturing at 𝑇𝐸
is therefore
𝑃𝑎𝑦𝑒𝑟
𝑃𝑉𝑡 = ∑𝐸𝑖=𝑆+1 𝛿𝑖𝐹𝑙𝑜𝑎𝑡 𝐹(𝑡, 𝑇𝑖−1 , 𝑇𝑖 )𝑁𝑖 𝐷(𝑡, 𝑇𝑖 ) − ∑𝐸𝑖=𝑆+1 𝛿𝑖𝐹𝑖𝑥𝑒𝑑 𝐾𝐷(𝑡, 𝑇𝑖 ) (3)

If we instead wish to obtain the price of a receiver swap, the positions are
reverse, so the fixed leg can be seen as an asset and the floating leg as a liability.
Hence we have that 𝑃𝑉 𝑅𝑒𝑐𝑒𝑖𝑣𝑒𝑟 = −𝑃𝑉𝑃𝑎𝑦𝑒𝑟 .

Entering into a swap, it is typical to trade swap with a net present value of zero.
This means that setting 𝑃𝑉 𝐹𝑖𝑥𝑒𝑑 = 𝑃𝑉𝐹𝑙𝑜𝑎𝑡𝑖𝑛𝑔 , we can isolate the fixed rate 𝐾 by
setting the present value of the IRS equal to zero in equation (3) above. The isolated
fixed rate can be denoted as the par swap rate and is defined as

16
 ∑𝐸𝑖=𝑆+1 𝛿𝑖𝐹𝑙𝑜𝑎𝑡 𝐹(𝑡, 𝑇𝑖−1 , 𝑇𝑖 )𝑁𝑖 𝐷(𝑡, 𝑇𝑖 )
𝑅(𝑡, 𝑇𝑆 , 𝑇𝐸 ) =
∑𝐸𝑖=𝑆+1 𝛿𝑖𝐹𝑖𝑥𝑒𝑑 𝐾𝐷(𝑡, 𝑇𝑖 )

Let us assume we have entered into a payer swap at T_S, paying the fixed rate
K and receiving a floating xIBOR rate until expiry T_E of the swap. If entering into a
matching receiver swap at some time t. Here we would receive the fixed par swap
rate 𝑅(𝑡, 𝑇𝑆 , 𝑇𝐸 ) and paying some floating xIBOR rate, but the floating rates will
cancel out and we will have a net cashflow of the fixed legs instead, where
𝑅(𝑡, 𝑇𝑆 , 𝑇𝐸 ) − 𝐾. Setting 𝐴(. ) as the time t sum of discounted fixed coverages,
𝐴(𝑡, 𝑇𝑆 , 𝑇𝐸 ) = ∑𝐸𝑖=𝑆+1 𝛿𝑖𝐹𝑖𝑥𝑒𝑑 𝐾𝐷(𝑡, 𝑇𝑖 ), the present value at time t for the initial
payer swap can be stated as
𝑝𝑎𝑦𝑒𝑟
𝑃𝑉𝑡 = 𝐴(𝑡, 𝑇𝑆 , 𝑇𝐸 )(𝑅(𝑡, 𝑇𝑆 , 𝑇𝐸 ) − 𝐾) (4)

Where 𝐴(𝑡, 𝑇𝑆 , 𝑇𝐸 ) is the so called annuity factor or level of the swap. By

differentiating this equation w.r.t par swap rate, the result is precisely𝐴(. ). 𝐴(. ) can
therefore be seen as the value of receiving one basis point over the period 𝑇𝑆 −
𝑇𝐸 and hereby thought as the payers sensitivity towards the par swap rate.

The swap rate is constructed as a government bond yield plus a spread

representing the increased credit risk relative to sovereign risk decided by
participants in the market. We can see the swap rate as a rate including both a risk-
free rate and a swap spread. Here we assume that the risk-free rate calculated on
observed government bonds as they are assumed to be safe investments due to the
fact that a government will never fill bankruptcy (broader speaking about highly
developed countries such as USA). The swap spread reflects therefore a premium
that can be explained by numerous drivers such as liquidity risk, BIS weighting,
credit ratings, credit situation and demand/supply.

17
European swaptions
A plain vanilla interest rate option like a European swaption, labelled swaptions
further on, is an option with the right (not obligated) to enter into an IRS starting on
a future date 𝑇𝑆 at predetermined fixed rate 𝐾, where the IRS is maturing at 𝑇𝐸 . As
an example if a party buys a payer swaption with 𝑇𝑆 = 2 and 𝑇𝐸 = 7, this means that
this party has the right to pay the fixed rate 𝐾, while receiving a floating xIBOR rate
at 2Y with expiry at 7Y, therefore the name 2Y7Y payer swaption. We can say this is
the same as buying at put option on a bond. Therefore the party buying a payer
swaption has the right to pay a fixed rate and receive floating rate, while it is vice
versa for a receiver swaption. Swaptions are also traded as OTC instruments.

Before entering at swaption contract both parties must agree on the type of
settlement. We have two types of settlement

 Cash settlement – here on exercise date the swap price on the underlying
swap is determined by an average of prices from five predetermined banks.
Both the highest and the lowest quotes are excluded in the calculation of the
average price. This determines the amount of cash exchanges between the
two parties.
 Physical settlement or swap settlement – here the buyer of the option will
enter into the underlying swap with actual cash flow exchanges until expiry
of the swap.

Swaption Pricing

As mentioned in the section above, a swaption can be settled differently. The

choice of settling has impact on the cash flows but also in the valuation of swaption
contract. We will price a swaption under the assumption that the holder has entered
a payer swaption.

Physical settlement

For a physical settled swaption the option holder can enter into a swap contract
predetermined by the conditions of the contracted option. A swap contract will only
be entered if the swaption is in-the-money to the swaption holder. This reminds us
of equation (4) and the ability to decide not to enter into the swap. The holder
receives at 𝑇𝑆

18
 𝑃ℎ𝑦𝑠𝑖𝑐𝑎𝑙
𝑃𝑎𝑦𝑒𝑟 𝑠𝑤𝑎𝑝𝑡𝑖𝑜𝑛 𝑃𝑉𝑇𝑆 = 𝐴(𝑇𝑆 , 𝑇𝑆 , 𝑇𝐸 )[(𝑅(𝑇𝑆 , 𝑇𝑆 , 𝑇𝐸 ) − 𝐾)]+
𝐸

𝑤ℎ𝑒𝑟𝑒 𝐴(𝑇𝑆 , 𝑇𝑆 , 𝑇𝐸 ) = ∑ 𝛿𝑖𝐹𝑖𝑥𝑒𝑑 𝐷(𝑇𝑆 , 𝑇𝑖 )

𝑖=𝑆+1

Cash settlement

At expiry of the option the holder of a cash settled swaption will receive the
present value of the underlying swap. Here we are discounting using the par swap
rate. The holder receives at 𝑇𝑆

𝑃𝑎𝑦𝑒𝑟 𝑠𝑤𝑎𝑝𝑡𝑖𝑜𝑛 𝑃𝑉𝑇𝐶𝑎𝑠ℎ

𝑆
= 𝐴̃(𝑇𝑆 , 𝑇𝑆 , 𝑇𝐸 )[(𝑅(𝑇𝑆 , 𝑇𝑆 , 𝑇𝐸 ) − 𝐾)]+
𝐸
𝛿𝑖𝐹𝑖𝑥𝑒𝑑
𝑤ℎ𝑒𝑟𝑒 𝐴̃(𝑇𝑆 , 𝑇𝑆 , 𝑇𝐸 ) = ∑ 𝑇𝑖 −𝑇𝑆
𝑖=𝑆+1 (1 + 𝛿𝑖𝐹𝑖𝑥𝑒𝑑 𝑅(𝑇𝑆 , 𝑇𝑆 , 𝑇𝐸 ))

Overall the difference between these two settlement agreements lies in

method of discounting. Following on, we will state 𝐴(. ) as a reference for both
cases.

We are now able to apply a swaption pricing model. Several pricing models
exist, varying in the assumptions regarding the behavior of the underlying swap
rates. The most known is the Black model, which assumes lognormal distributed
behavior. The derivation of the result behind the Black model will not be enlighten,
but solitary stated by result

𝑃𝑎𝑦𝑒𝑟 𝑠𝑤𝑎𝑝𝑡𝑖𝑜𝑛 𝑃𝑉𝑡 = 𝐴(𝑡, 𝑇𝑆 , 𝑇𝐸 )[(𝑅(𝑡, 𝑇𝑆 , 𝑇𝐸 )𝑁(𝑑1 ) − 𝐾𝑁(𝑑2 ))]

𝑅(𝑡, 𝑇𝑆 , 𝑇𝐸 ) 1
log ( 𝐾 ) − 2 𝜎 2 (𝑇𝑠 − 𝑡)
𝑑1 =
𝜎√𝑇𝑠 − 𝑡

𝑑2 = 𝑑1 − 𝜎√𝑇𝑠 − 𝑡

We can find the value of the swaption by determining the forward swap rate
𝑅(𝑡, 𝑇𝑆 , 𝑇𝐸 ), the annuity factor (cash or physical) 𝐴(𝑡, 𝑇𝑆 , 𝑇𝐸 ), the time to expiry and
lastly the volatility 𝜎.

19
 Pricing the receiver swaption can be done easily through the put-call-parity for
plain vanilla European options. If the payer and receiver option has identical strike
rate 𝐾 and identical timeframe (entered into a swaption at time t, start date at 𝑇𝑆
and end date at 𝑇𝐸 ), the parity states

𝐹𝑜𝑟𝑤𝑎𝑟𝑑 𝑆𝑡𝑎𝑟𝑡𝑖𝑛𝑔 𝑃𝑎𝑦𝑒𝑟 𝑆𝑤𝑎𝑝(𝐾)

= 𝑃𝑎𝑦𝑒𝑟 𝑆𝑤𝑎𝑝𝑡𝑖𝑜𝑛(𝐾) − 𝑅𝑒𝑐𝑒𝑖𝑣𝑒𝑟 𝑆𝑤𝑎𝑝𝑡𝑖𝑜𝑛(𝐾)

Rearranging this parity, we are able to derive the price of the receiver swaption
as followed

𝑅𝑒𝑐𝑒𝑖𝑣𝑒𝑟 𝑠𝑤𝑎𝑝𝑡𝑖𝑜𝑛 𝑃𝑉𝑡 = 𝐴(𝑡, 𝑇𝑆 , 𝑇𝐸 )[(𝐾𝑁(−𝑑2 ) − 𝑅(𝑡, 𝑇𝑆 , 𝑇𝐸 )𝑁(−𝑑1 ))]

Volatility in the pricing model can be formulated as the anticipated future

volatility in an option. From Black Scholes option framework, implied volatility for
swaptions are the core element in calculating premiums as implied volatility is the
only unobservable parameter in this framework. Hereby the price is only influenced
by the implied volatility due to the fact that the underlying, strike and expiry are all
observable.

Swaptions are in the United States all almost cash settled while in Europe
approximately 50 % are in fact cash settled. However all European swaptions
included in this paper are cash settled.

Due to the nature behind lognormality, the Black model measures implied
volatility in relative approach, i.e. relative changes of the forward swap rate, while
the normal model measures the implied volatility as the absolute changes of the
forward swap rate. We will now investigate the normal model.

20
The Normal model
The normal model is a rather simple model, but still a benchmark model when
working with negative interest rates. It was introduced in 1900 by a French
mathematician, named Bachelier. The evolution of the forward swap rate 𝐹𝑡 follows
the stochastic differential equation

𝑑𝐹𝑡 = 𝜎𝑁 𝑑𝑊𝑡

With 𝜎𝑁 as the normal volatility and 𝑊𝑡 as a Wiener process. Introducing Ito

calculus7, we are able to find a solution to this equation.

𝐹𝑡 = 𝐹0 + 𝜎𝑁 𝑊𝑡

Here is the forward swap rate assumed to be normally distributed or the

process behaves as a standard Brownian motion.

The price of a payer swaption in the normal model can be found as

𝑑2
𝑒− 2
𝑃𝑉𝑇𝑆 = 𝐴(𝑇𝑆 , 𝑇𝑆 , 𝑇𝐸 )𝜎𝑁 √𝑇𝑆 + 𝑑𝑁(𝑑)
√2𝜋
( )
Where 𝑁(… ) is the probability distribution of a standard normal variate and

𝑓−𝐾
𝑑=
𝜎𝑁 √𝑇𝑆

By introducing the put-call parity we can easily calculate the receiver swaption
as

𝑑2
−
𝑒 2
𝑃𝑉𝑇𝑆 = 𝐴(𝑇𝑆 , 𝑇𝑆 , 𝑇𝐸 )𝜎𝑁 √𝑇𝑆 − 𝑑𝑁(−𝑑)
√2𝜋
( )
The normal volatility is not comparable with black volatility as one is quoting an
absolute volatility level while the other is quoting a relative volatility level. However
an approximation can be derived for converting normal volatilities into black
volatilities and the other way around.

7
Stochastic Calculus
21
 As mentioned earlier, the normal model assumes swap rates to be normal
distributed instead of log normal distributed in the Black model. Black model
measures implied volatility in relatively order, as the relative changes of the forward
swap rate. Conversely to the normal model, which measures implied volatility as
absolute changes of the forward swap rate. It is important to note that implied
volatility itself isn’t the volatility of movements of the swap rate, but instead the
markets opinion regarding this. Still, the linkage between these two sizes is
expected to be strong.

Risk measures under the normal model

Delta

Delta is defined as the change in the value of the option when the price of the
underlying asset increases.

Delta for a payer swaption is given as

𝜕𝑉
∆𝑃𝑆 = = 𝐴(𝑇𝑆 , 𝑇𝑆 , 𝑇𝐸 )𝑁(𝑑)
𝜕𝐹
𝑓−𝐾
𝑑=
𝜎𝑁 √𝑇𝑆

Gamma

Gamma is defined as the sensitivity of the delta to the underlying asset.

𝜕 2 𝑉 𝐴(𝑇𝑆 , 𝑇𝑆 , 𝑇𝐸 )
𝛤𝑃𝑆 = 𝛤𝑅𝑆 = 2 = 𝑁(𝑑)
𝜕 𝐹 𝜎𝑁 √𝑇𝑆

Vega

Vega is defined as the change in the option price due to change in the volatility of
the underlying asset.

Vega is the same for put and call options, which can be given from the put-call
parity.

𝜕𝑉
𝜈𝑃𝑆 = 𝜈𝑅𝑆 = = 𝐴(𝑇𝑆 , 𝑇𝑆 , 𝑇𝐸 )√𝑇𝑆 𝑁(𝑑)
𝜕𝜎
22
Theta

Theta is defined as the change in the option prices with the respect to time to
maturity. A purchased option will decrease as time goes by, while holding all other
parameters constant. At expiry of the option, the time value of the option is zero
and the value of the option is only given by its intrinsic value.

Quoting swaptions
As mentioned earlier, swaptions are OTC products and are quoted for various
currencies. They are quoted as straddle options premia, Black or Normal implied
volatilities. On the figure below, we see ATM straddles quoted in normal volatility.
This is what most market participants today uses when quoting prices on swaptions.

Figure 1: Normal Implied Volatilities for EUR Swaption, Date 01-25-2017, Source: ICAP

23
 A straddle strategy is a combination of buying or selling both a call option (here
receiver swaption) and a put option (here payer swaption) on a swap with the same
strike rate and expiry. A receiver swaption is receiving a fixed swap rate, so the
holder of the swaption has the right to receive a fixed rate in a swap contract in a
determined future. A payer swaption is a put option as the holder has the right to
pay a fixed rate in a swap contract in a determined future.

A long position in a swaption straddle can then be expressed as

𝑆𝑡𝑟𝑎𝑑𝑑𝑙𝑒 𝑃𝑉𝑇𝑠 = 𝑃𝑎𝑦𝑒𝑟 𝑠𝑤𝑎𝑝𝑡𝑖𝑜𝑛 𝑃𝑉𝑇𝑠 + 𝑅𝑒𝑐𝑒𝑖𝑣𝑒𝑟 𝑠𝑤𝑎𝑝𝑡𝑖𝑜𝑛 𝑃𝑉𝑇𝑠

We can easily derive an ATMF straddle price quoted in normal volatility, since
the forward rate and strike price are the same. The NPV for a straddle is then

2𝑇𝑆
𝑆𝑡𝑟𝑎𝑑𝑑𝑙𝑒 𝑃𝑉𝑇𝑠 = 𝑁𝐴(𝑇𝑆 , 𝑇𝑆 , 𝑇𝐸 )𝜎𝑁 √
𝜋

Where 𝑁 is the notional,𝐴(𝑇𝑆 , 𝑇𝑆 , 𝑇𝐸 ) is the swap annuity factor or basis point

value and 𝜎𝑁 is the normal implied volatility.

Straddles are often traded as to speculate how volatility in the future changes.
We could look at a long short combination of two ATM straddles with different
maturities. The option writer of a straddle has unlimited downside risk while having
an upside limited to a premium received from the two options. Straddles are a
common option strategy for investment purpose, here with the expectation of
market stability/neutralization and stable or declining volatility. Hereby such a
position would only be profitable before expiry decline in implied volatility or time
value decay. A straddle is often applied in speculating on future changes of the
volatility.

However straddle prices don’t depend on forward swap rate in a directly

manner, but purely on volatility. Therefore they can be viewed as a good indicator of
implied volatility level.

There is no delta risk when entering at straddle, but large market volatility can
cause an exposure to delta risk. ATMF straddles are also more liquid than
standalone payer and receiver swaptions, due to limited dollar value of 01 basis
point (DV01 risk). Often are ATMF straddles with different expiries the cheapest way
to hedge gamma, vega and theta risk.

24
Choosing normal or log-normal
We have both introduced the the normal model and the lognormal model in the
swaption framework for pricing swaptions. As European ATM swaptions are
classified as a plain vanilla option, simple approaches are used on trading desks such
as these two models.

As lognormality can’t deal with negative interest rates and also has difficulties
with low interest rates, we favor to use normal volatility.

As this paper is more interested in the changes of the implied volatility, the
perspective on selecting a pricing framework isn’t important for the analysis. The
difference between the models are however important to understand in explaining
the basis of the unit of measurement.

However an interesting issue follows when converting Black volatilities into

Normal volatilities. The relationship for ATMF swaptions can be expressed as

𝜎𝑁 ≈ 𝑓𝜎𝐵𝑙𝑎𝑐𝑘
1
If 𝜎𝑁 is constant, 𝜎𝐵𝑙𝑎𝑐𝑘 is proportional to , where f is forward rate
𝑓

This can be a convenient way to approximate in either normal or black volatility.

If normal volatility is constant, lognormal will be decreasing in rates. This is a so-
called skew effect.

25
Methodology

Mean Reversion
In relative value analysis one of the most fundamental tools is the mean
reversion process. This can be explained by expecting over at time period that one
or several variables to follow a long-run average.

Ornstein Uhlenbeck stochastic differential equation can be found as

𝑑𝑆𝑡 = 𝜆(𝜇 − 𝑆𝑡 )𝑑𝑡 + 𝜎𝑑𝑊𝑡

Here is 𝑑𝑆𝑡 the change in the value of the random variable 𝑆 over time t, 𝜆 is
known as the speed of mean reverting process, the long term equilibrium/average
of the variable 𝑆 is given by the parameter 𝜇, 𝜎 is the instantaneous volatility of the
random variable 𝑆 and the 𝑑𝑊𝑡 is the change in the standard Wiener process 𝑊𝑡

Generally a stochastic differential equation is formulated as

𝑑𝑥𝑡 = 𝑓(𝑥𝑡 )𝑑𝑡 + 𝑔(𝑥𝑡 )𝑑𝑊𝑡

The first part of the formula expresses the drift process 𝑓(𝑥𝑡 ) and defines the
mean of the process while the term 𝑔(𝑥𝑡 ) expresses a diffusion process where the
volatility of the process is stated.

The mean reverting process can also calculate a stopping time also known as
first time passage. In this paper maximum likelihood is used to calculate the
parameters of the Ornstein-Uhlenbeck process. In finding these parameters, the
stochastic differential equation exact solution of the OU process above is discretized
and approximated as

(1 − 𝐞−𝟐𝛌∆𝐭 )
St+1 = St 𝐞−𝛌∆𝐭 + μ(1 − 𝐞−𝛌∆𝐭 ) + σ√ ∆𝐖𝐭
2𝛌

26
 Here is ∆𝐭 an infinitesimal change, while ∆𝐖𝐭 are independent identically
distributed Wiener process. This formula is useful in simulation of generating paths,
we want to analyze later on.

Calibrating MR with MLE

From Calibrating the Ornstein-Uhlenbeck (Vasicek) model, Van Den Berg 2011
paper it states, 𝑆𝑡+1 conditional probability density function is given as

1 1 2
𝑃(𝑁0,1 = 𝑥) = 𝑒 −2𝑥
√2𝜋

This is a combination between the normal distribution probability density

function and the solution to the stochastic differential equation above.

Now looking at the conditional probability density function of an observation

𝑆𝑡+1 given a previous observation 𝑆𝑡 , here with 𝛿 as the time interval between the
two observations. The equation of the conditional probability density function is
therefore given as
2
1 (𝑆𝑡 − 𝑆𝑡−1 𝑒 −𝜆𝛿 − 𝜇(1 − 𝑒 −𝜆𝛿 ))
𝑓(𝑆𝑡+1 |𝑆𝑡 ; 𝜇, 𝜆, σ
̂) = exp[−
̂
√2𝜋σ ̂2
2σ

Now we can derive the log likelihood function of a set of observations

(𝑆0 , 𝑆1 , … , 𝑆𝑛 ) from the conditional probability density function. Showed here as
𝑛

̂) = ∑ ln 𝑓(𝑆𝑡 𝑆𝑡−1 ; 𝜇, 𝜆, 𝜎̂)

ℒ(𝜇, 𝜆, σ
𝑡=1
𝑛
𝑛 1 2
̂) −
= − ln(2𝜋) − 𝑛𝑙𝑛(σ ∑[𝑆𝑡 −𝑆𝑡−1 𝑒 −𝜆𝛿 − 𝜇(1 − 𝑒 −𝜆𝛿 )]
2 2σ
̂
𝑡=1

We then find the first order conditions for the maximum likelihood estimation
by setting these equal to zero

̂)
𝜕ℒ(𝜇, 𝜆, σ
=0
𝜕𝜇

27
 𝑛
1 2
= 2 ∑[𝑆𝑡 −𝑆𝑡−1 𝑒 −𝜆𝛿 − 𝜇(1 − 𝑒 −𝜆𝛿 )]
̂
σ
𝑡=1
∑𝑛𝑡=1[𝑆𝑡 −𝑆𝑡−1 𝑒 −𝜆𝛿 ]
𝜇=
𝑛(1 − 𝑒 −𝜆𝛿 )

̂)
𝜕ℒ(𝜇, 𝜆, σ
=0
𝜕𝜆
𝑛
𝛿𝑒 −𝜆𝛿
=− ∑[(𝑆𝑡 −𝜇)(𝑆𝑡−1 − 𝜇) − 𝑒 −𝜆𝛿 (𝑆𝑡−1 − 𝜇)2 ]
̂2
σ
𝑡=1
1 ∑𝑛𝑡=1(𝑆𝑡 −𝜇)(𝑆𝑡−1 − 𝜇)
𝜆 = − ln
𝛿 ∑𝑛𝑡=1(𝑆𝑡−1 − 𝜇)2

̂)
𝜕ℒ(𝜇, 𝜆, σ
=0
𝜕σ
̂
𝑛
𝑛 1 2
= − 3 ∑[𝑆𝑡 − 𝜇 − 𝑒 −𝜆𝛿 (𝑆𝑡−1 − 𝜇)]
̂ σ
σ ̂
𝑡=1
𝑛
1 2
σ2 =
̂ ∑[𝑆𝑡 − 𝜇 − 𝑒 −𝜆𝛿 (𝑆𝑡−1 − 𝜇)]
𝑛
𝑡=1

The three solutions depends on each other but as seen on the equations above
𝜆 and 𝜇 are independent of σ ̂, wherefore if we either have the value of 𝜆 or 𝜇, the
other parameter can be derived. Lastly, we are able to derive σ ̂ when both 𝜆 and 𝜇
are known. For solving the equations, finding either 𝜆 or 𝜇 will be sufficient.

A derivation of the following results can be found in the appendix. Solving these
equations gives following maximum likelihood estimators

𝑆𝑦 𝑆𝑥𝑥 − 𝑆𝑥 𝑆𝑦
𝜇=
𝑛(𝑆𝑥𝑥 − 𝑆𝑥𝑦 ) − (𝑆𝑥2 − 𝑆𝑥 𝑆𝑦 )

And for speed mean reversion rate

1 𝑆𝑥𝑦 − 𝜇𝑆𝑥 − 𝜇𝑆𝑦 + 𝑛𝜇 2

𝜆 = − ln
𝛿 𝑆𝑥𝑥 − 2𝜇𝑆𝑥 + 𝑛𝜇 2
28
 Lastly we get the variance as

2𝜆
σ2 = σ
̂2
1 − 𝛼2
𝛼 = 𝑒 −𝜆𝛿

Where

1
σ2 =
̂ [𝑆 − 2𝛼𝑆𝑥𝑦 + 𝛼 2 𝑆𝑥𝑥 − 2𝜇(1 − 𝛼)(𝑆𝑦 − 𝛼𝑆𝑥 ) + 𝑛𝜇 2 (1 − 𝛼)2 ]
𝑛 𝑦𝑦
These results are used to compute the expected future path and the standard
deviations.

29
Principal component analysis8
When looking at financial data such as swaptions quotes, we sometimes deal
with large data sets that are driven by only a couple of factors. These factors can be
found through reducing the dimensionality of variables and still preserving the
variables explaining most of the variance in the data set. Principal component
analysis (PCA) is a statistical tool used for the purpose of identifying patterns and
expressing the data in formal approach to expose their differences and similarities.
This can help us in analyzing and identifying relative value opportunities, where the
relative value of one or several instruments are independent of market direction or
hedging solutions etc.

The main assumption for PCA is that factors/principal components as 𝑌𝑖 are

uncorrelated. This is done by finding a rotation of the variables. This is a rather weak
assumption and therefore gives the market the ability to add additional information
regard about shape and strength of each factor.

We see the principal components as linear combinations of random variables in

a p-dimensional space 𝑋1 , 𝑋2 , … , 𝑋𝑝 . By rotating the original system of the random
variables, a new coordinate system is created as a representation of the linear
combinations. These new coordinate axes characterize the directions with
maximum variability and gives a stricter but also simpler portrayal of the covariance
structure. As a matter of fact the principle components focus exclusively on a
covariance matrix Σ (or correlation matrix) and don’t require any assumptions
regarding multivariate normality.

If we have a random vector 𝑿′ = [𝑋1 , 𝑋2 , … , 𝑋𝑝 ] with the covariance matrix 𝚺 and

eigenvalues 𝜆1 ≥ … ≥ 𝜆𝑝 ≥ 0.

Let us consider the linear combinations

𝑌1 = 𝒆1𝑇 𝑿 = 𝑒1,1 𝑋1 + ⋯ + 𝑒1,𝑝 𝑋𝑝

𝑌2 = 𝒆𝑇2 𝑿 = 𝑒𝑖,1 𝑋1 + ⋯ + 𝑒2,𝑝 𝑋𝑝
⋮=⋮
𝑌𝑝 = 𝒆𝑇𝑝 𝑿 = 𝑒𝑝,1 𝑋1 + ⋯ + 𝑒𝑝,𝑝 𝑋𝑝

8
Richard A. Johnson & Dean W. Wichern, “Applied Multivariate Statistical “
30
𝑌1 , 𝑌2 , . . , 𝑌𝑝 form the principal components of 𝑿. The principal component 𝑌1
explains most of the variance in the used dataset, therenext 𝑌2 and so on. Hereby
we are interested in choosing the fewest number of factors while keeping most
variance, as these factors gives us information about movements in the data set.

From linear algebra we know that linear combinations we can find

𝑉𝑎𝑟[𝑌𝑖 ] = 𝒆𝑇𝑖 𝚺𝒆𝑖 = 𝜆𝑖 𝑖 = 1, … , 𝑝

𝐶𝑜𝑣[𝑌𝑖 , 𝑌𝑘 ] = 𝒆𝑇𝑖 𝚺𝒆𝑘 = 0 𝑓𝑜𝑟 𝑖 ≠ 𝑘, 𝑖, 𝑘 = 1, … , 𝑝

And Σ has the eigenvalue-eigenvector pair (𝜆1 , 𝑒1 ) … , (𝜆𝑝 , 𝑒𝑝 )

Further we have that

𝐸(𝑌𝑖 ) = 0
𝑝 𝑝
∑ Var[𝑌𝑖 ] = ∑ Var[𝑋𝑖 ]
𝑖=1 𝑖=1

We have the following linear combination of X

𝒀𝒊 = 𝒂𝑇𝑖 𝑿 where 𝑉𝑎𝑟(𝒀𝒊 ) = 𝒂𝑻𝒊 𝜮𝒂𝒊

The first principal component is a linear combination with most variance, where we
maximize 𝑉𝑎𝑟(𝒀𝒊 )

𝑎𝑇 𝛴 𝑎
𝑀𝑎𝑥𝑎≠0 [ 𝑇 = 𝜆1 ]
𝑎 𝑎

As 𝑒1𝑇 𝑒1 = 1, we get that

𝑒1𝑇 𝛴 𝑒1
𝜆1 = 𝑇 = 𝑒1𝑇 𝛴 𝑒1 = 𝑉𝑎𝑟[𝑌1 ]
𝑒1 𝑒1

The next principal components are linear combinations which maximize their
variance and are uncorrelated to previous k components

𝑎𝑇 𝛴 𝑎
𝑀𝑎𝑥𝑎⊥𝑒1,…,𝑒𝑘 [ 𝑇 = 𝜆𝑘+1 ] , 𝑘 = 1,2, … , 𝑝 − 1
𝑎 𝑎

𝑡
When 𝑒𝑘+1 𝑎𝑛𝑑 𝑒𝑘+1 𝑒𝑘+1 = 1

31
 Then
𝑇
𝑒𝑘+1 𝛴𝑒𝑘+1 𝑇
𝜆𝑘+1 = 𝑇 = 𝑒𝑘+1 𝛴𝑒𝑘+1 = 𝑉𝑎𝑟[𝑌𝑘+1 ]
𝑒𝑘+1 𝑒𝑘+1

Where we have 𝑉𝑎𝑟[𝑌𝑖 ] = 𝜆𝑖

The sum of all eigenvalues are equal to the sum of the variance for all variables,
here the i-th principal component explains a proportion of the total variance

𝜆𝑘
𝑘 = 1,2, … , 𝑝
𝜆1 + ⋯ + 𝜆𝑝

What does the scores represents? As the relationship between each eigenvector
of the covariance matrix in a PCA is orthogonal, the eigenvectors are used to project
the data from its original axes into the ones characterized by the computed principal
components. The factor scores is a rebasing of the original coordinate system of the
entire dataset into a space given by new axes defined by the eigenvectors with the
greatest variance.

In deriving a PCA, correlation or covariance matrix can be used for the

derivation. Using the covariance matrix, the computed principal component scores
will be presented in the original units, while choosing the correlation matrix instead
will produce unitless results. This is due to the fact that the correlation matrix
standardizes all dimensions of the dataset and this paper chooses to operate with
the covariance matrix.

The data used for a PCA is not required to follow a Gaussian distribution, but
assumes linearity as PCA computes a projection of the data on dimension reduced
linear subspace. Any non-linear relationships between variables are therefore not
regarded in this process.

For hedging purposes these uncorrelated factors can give strong hedging
opportunities. The factors can be viewed as risk factors. Here we use the weights
and loadings from our PCA in calculating hedge ratios that immunize a portfolio of
securities against changes in factors. Remember this is a linear combination.

32
 Using PCA in R
All calibrations are done in the statistical program R. Here we have used the
function Prcomp(), in the calibration of PCA tool, from a library called PCA in the
statistical program R. Prcomp executes a PCA by a singular value decomposition
method of the data matrix, which is calculated from the covariance matrix. Other
methods are a possibility but won’t be regarded in this paper. The singular value
decomposition is given as

𝑋 = 𝑈𝐷𝑉 𝑇

𝑈 is a 𝑚 𝑥 𝑚 where the columns are orthonormal vectors

𝑉 is 𝑛 𝑥 𝑛 where the columns are orthonormal vectors

𝐷 is 𝑚 𝑥 𝑛 diagonal with diagonal elements called singular values of 𝑋.

We are especially looking at the arguments $rotation and $x in R function

prcomp. The argument $rotation gives a matrix of the eigenvectors or factor
loadings which shows the weighting of variables/instruments meanwhile argument
$x gives the PC or scores/factors as the new variables.

Algorithm for PCA based on SVD

1) Collect the p observed data samples into a matrix 𝑿 = [𝑋1 , 𝑋2 , … , 𝑋𝑝 ]

2) Compute the singular value decomposition of the matrix by
𝑋 = 𝑈𝐷𝑉 𝑇
3) Find the principal directions in the columns 𝑈
4) Find the principal components that are stored in the columns of matrix
𝑍 = 𝐷𝑉 𝑇

33
Distribution of the volatility surface
We can divide ATM options into two different types of options by their expiry.
Looking at ATM options with relatively short time to expiry, prices are mainly
depended on moves in the underlying asset. These options have high gamma and
therefore will delta changes significantly when the underlying asset moves.
Whereas ATM options with longer maturities are mainly depended on moves in the
implied volatility level. To summarize this, options with a shorter expiry are greatly
affected by changes in the underlying and therefore can be classified to be exposed
to changes in the realized volatility. Here we can state the realized volatility as the
movement in the underlying asset. We define this as the gamma sector of the
volatility surface.

However, options with a longer expiry horizon are less impacted by changes in
the underlying as but instead have a significant exposure to changes in the implied
volatility level. We define these options as vega sector options on the implied
volatility surface.

We can therefore differentiate between these two markets by the time to

expiry. There is no conclusive line between the classifications of these two sectors
on the volatility surface. Options with around two years to expiry can’t be distinctly
classified and falls into a gray area between the gamma and vega sector. These
‘gray area’ options are considered in each particular analysis of the volatility surface
whenever they belong in the gamma or vega sector. This will of course be followed
up on in the analysis. The expiry of the underlying is irrelevant in this discussion. The
classifications of the two segments are illustrated in figure 2 below.

Options classified as in the gamma sector can be traded actively within a Black-
Scholes or Bachelier valuation framework using delta hedging. By focusing on the
vega sector instead, this gives rise for another mindset that is unlike the well-known
delta hedging strategy. Introducing a statistical analysis on the implied volatility
surface can be a powerful tool in the decision making of relative value opportunities,
hedging etc. A common tool such as a PCA can identify the drivers of the implied
volatility that affects the whole surface for vega sector options and use this
information in the decision making of trading on the implied volatility or in hedging
a portfolio of swaptions. As implied volatility is the overarching parameter, it makes
little sense to search for a superior valuation model that can undermine the Black

34
Scholes framework but instead use this statistical tool to identify value in the vega
sector. Realized volatility in the underlying are of a little concern for pricing
straddles in the vega sector, as these are more or less strictly exposed to changes in
the implied volatility level. Given this characteristic, we can discard linkages to
external variables such as movement in the underlying swap rate, and instead
concentrate on internal statistical relationships between instruments on the implied
volatility surface by considering these as time series used for statistical analysis.

Expiry 1Yr 2Yr 3Yr 4Yr 5Yr 6Yr

1Mo 35.5 53.8 63.5 69 73.7 73.3
3Mo
6Mo
46.6
51.4
61
65
GAMMA
71.3
73.5
76.2
77.2
80.2
80.8
79.9
81.2
1Yr 61.5 70.1 76.2 79.3 82.3 82.6
2Yr 76.5 80.7 GRAY
81.4 AREA
82.8 84.2 84.2
3Yr 85.6 85.4 85.4 85.5 85.6 85.4
4Yr
5Yr
89
90.1
88.1
89.6
VEGA
87.7
88.9
87.2
88.2
86.8
87.5
86.3
86.8
7Yr 88.5 88.5 87.6 86.7 85.7 84.8
10Yr 83.9 82.6 81.7 80.8 79.9 79.1
15Yr 72.9 71.3 70.8 70.3 69.8 69.2

Figure 2 - Classification of ATM Swaption volatility matrix for USD, date 5-18-2016

This paper focuses on volatility surface of at-the-money-forward (ATMF)

swaptions and does not take different strikes levels into consideration. Adding the
feature would create a cube instead of a surface, where we would need to account
skew as an extra dimension. It is however a possibility to add this dimension in the
vega sector without implementing an option price model as for the gamma sector.
However given limited liquidity for out-the-money-forward swaptions, data don’t
necessarily reflect actual tradable prices but more often constructed prices from
brokers. Therefore will a PCA with this extra dimension not represent the actual
movements in the implied volatility market but an artificial reality decided by
models quoting the OTMF swaptions volatilities.

35
 Data

Description
Our concentration lies on EUR and USD European Swaptions. Here is implied
normal volatility data on both markets extracted. Data includes swaptions with
expiry ranging from 3 months to 30 years, while tenors are ranging from 1 year to 30
years. These options can be exercised at expiry as an X year swap. If we have 5y10y
swaption, this is an option to pay or receive fixed for five years in 10 years’ time.

The data for the swaption volatility matrix is extracted from a Bloomberg
terminal with a data license to ICAP. The ICAP broker is considered by several
market makers to be a valid distributer of rather accurate ATMF swaption quotes,
reflecting actual tradable implied volatilities. The volatility quotes extracted are all
mid quotes. The following Bloomberg function VCUB was used for finding ATMF
implied volatilities for the EUR and USD markets. Both markets trades European
swaptions actively and are considered to be two of the largest and most liquid
markets trading European swaptions. As mentioned earlier, OTM swaptions can be
very illiquid, which is one of the main reasons that this thesis only focuses on ATM
swaptions.

The underlying of a swaption is the forward swap rate, which are the
cornerstone when pricing these instruments, but this rate are not directly traded on
the OTC market. The rates are actually computed by a yield curve model and the
computed forward swap rates will not always represent the actual market forward
swap rate. But such brokers such as ICAP supply both implied forward swap rates
with quoted volatilities.

When modelling the estimated parameters in a PCA with high accuracy, it is of

great importance that the chosen market data has an adequate lifespan. There was
no data available for EUR swaption from ICAP or other brokers before the 1.
November 2010 and this is therefore the starting point for EUR swaptions. Starting
date for USD swaptions is the 1st of January 2010 with ending date at the 25th of
January 2017. Data is collected daily and the chosen timestamp for the extracting is

36
at end-of-day. We start off with investigating implied volatility level data in basis
points units.

Stock indexes, currency crosses, fixed income data and commodity data are
collected on a daily basis from 2010 to 2017. Some equity data are not available for
all days considered for swaptions. Therefore will days without data, be given the
same value as the latest former trading day. This should not conflict with the
collected swaption data as many of these holidays; we see no change in normal
volatility quotes. This is not a problem for the currency crosses as these are traded 7
days a week.

37
Empirical Results

Analysis
PCA is run for a 2.5 year timespan starting by taking data from the 1st of January
2010 and running each single day until the 25th of January 2017 to avoid biases when
evaluating trading signals or for the interpretation of the results constructed from a
PCA. Longer or shorter period could be chosen. Mainly focusing on capturing
structural changes in level data, a longer period is preferable. Fewer instruments
can be regarded depending on the area of interest. We want to construct a
consistent framework, which can identify and extract value on the USD and EUR
ATMF swaption markets. This framework is projected to help develop informal
decisions regarding relative value analysis for purposes such as trading strategies,
asset allocation or hedging risk factors.

As the swaption matrix are a two dimensional surface defined by expiry and
tenor, data must be transformed into a one dimensional vector, as the loadings or
the covariance matrix in a PC analysis otherwise would compute a multi-
dimensional covariance matrix, which would be impossible to work with. The
solution is to compile all instruments into one dimensional vector by combining
expiry and tenor for all possible outcomes and stack them together.

𝜎𝑁1𝑀1𝑌
𝜎𝑁3𝑀1𝑌
1𝑥𝑁 𝑣𝑒𝑐𝑡𝑜𝑟 = ⋮
𝜎𝑁1𝑌2𝑌
⋮
(𝜎𝑁30𝑌30𝑌 )

After this is obtained, a two dimensional space can be introduced as all

instruments is set in a one dimensional vector and including time as the new
dimension. This enables us to visualize each principal component eigenvectors of all
instruments as a 1xN vector, where n is the number of instruments.

There is an important choice to consider before running a PCA. Thus is deciding

if data for each variable should be standardized. This is mentioned as PCA is not
38
scale invariant, because the calculated eigenvectors are not scale invariant. Notably,
this can give instruments with higher implied normal volatility larger weighting in
the analysis. However this is not conclusive since the difference between the lowest
and highest volatilities across instrument isn’t more divided. On figure 4 and 5, the
scaling of all swaption with at least 2Y to expiry for USD and EUR are shown as
boxplots. We see implied volatilities varies most for short expiry options, but all
instruments moves within the same range. As the normal implied volatility across
the whole surface regarding different maturities and underlying swap rates gives
values in a comfortable range, standardization are not acknowledged. Moreover,
constructing a PCA with a covariance matrix is preferable when dealing with
variables stated in the same unit of measurement and same scaling as the
computed factor scores can be interpreted in the same unit as the input data.

Figure 3 – USD boxplots of the scaling of normal implied volatility values for vega sector

39
 Figure 4 - EUR boxplots of the scaling of normal implied volatility values for vega sector

The USD normal volatility surface for ATM swaption are enlighten in a three
dimensional surface on figure 6. The y-axis presents the underlying swap rate if the
option and the x-axis describe the expiry of the option. Clearly, implied normal
volatility decreases as expiry increases. Volatility is low in the very short end for a 1
year tenor and only one month to expiry. But all other short expiry swaptions offers
higher implied normal volatilities as a volatility risk premium. The implied normal
volatilities are shown in basis points (bps).

40
 Figure 5 Normal volatility surface USD date 9 November 2016

The surface features a downward slope, decreasing as expiry rises. Focusing

only on changes in tenor, we see a spike at 10Y tenor. If we exclude expiries less
than 2Y only, we only see smaller variations on the implied volatility surface.

Interpretation of the eigenvectors

Firstly, by running a daily 630 days PCA for the whole implied volatility surface
for USD, no instruments are excluded from the analysis. Here the purpose is to
illustrate the dynamics of the first eigenvector and give an interpretation of this.
This will be followed up by an interpretation of the eigenvectors for the three most
influential principal components for the restricted vega sector. The interpretation is
done by evaluating the weights, also called sensitivities, for each instrument. The
weights illustrated are for that specific day, as weights for all instruments changes
over time while staying in the same pattern.

41
First eigenvector

From figure 7 below, the weights of the first eigenvector of a PCA on the whole
surface USD are shown. The instruments are grouped and colored by their tenors.
For every tenor, instruments are arranged in ascending expiry order starting with
1M.

We observe that sensitivities of the eigenvector decreases with expiry and

negative entries are introduced for 10Y to 30Y expiry options. As the entries for
input variables has opposite signs this can be seen as a strong indication for
introducing a segmentation of the gamma and vega sector, as we explained in the
section ‘Distribution of the volatility surface’.

Restricting the input data to options with at least 2 years expiry resolves the
problem as seen in figure 8 below, where we have illustrated the weights of first
eigenvector, which is generally calculating negative weights. The interpretation of
the negative weights for the first principal component tells us that all instruments
are affected in the same direction. However some instruments are more affected
than others, e.g. the first principal can be understood as a level factor. This level
factor describes the dynamics in the overall implied volatility level.

42
Figure 6 - First eigenvector of the whole USD volatility surface, Date: 2015-07-03

43
 Analogously, we see that almost every 3Y expiry options are undoubtedly a part
of the vega sector as the sensitivity values all are negative. However we observe that
2Y1Y, 3Y1Y, 2Y2Y and 2Y3Y swaptions all have positive entries. This can be explained
by these shorter expiry options still must have some gamma influence during this
period. But the positive sensitivities of the 3Y1Y and 2Y3Y are quite close to zero,
giving us no reason not to include these instruments as a part of the vega sector,
which a cluster analysis, described later on, will validate.

While the rest of the 2Y expiry options all have a negative sensitivity, the figure
below gives a pattern of the attributes of the first eigenvector. The longer the expiry
of the option (more or less), it gets less important what swap rate is the underlying,
when being certain of an option could be classified as part of the vega segment.
Especially for swap rates 1Y to 7Y, sensitivity becomes more negative due to longer
option expiry. We can say that the longer the tenor is, the shorter the expiry needs to
be in the extent of an option being a part of the vega sector.

Figure 7 First eigenvector of vega sector, USD surface, Date: 2015-07-03

The sensitivities are roughly uniform distributed, which in essence gives

associations towards empirical evidence of swap curve shifts from PCA. This support

44
our interpretation on how different ATMF swaptions are affected by shifts on the
implied volatility surface. PCA is often used to show how a shift in a curve empirically
differs from theory. In theory a parallel shift occurs when expectation changes at once
on a curve for all tenors by the same amount, but in empirically the shift differs for
short-term and long-term tenors. In this example, the result is the similar but instead
with a combination of short expiry-tenor and long expiry-tenor. This is the difference
between the most volatile and least volatile implied volatility swaptions. The
sensitivities for 4Y3Y and 7Y3Y are -0.0619 and -0.0778. This means that if 4Y3Y falls
with one basis point, the 7Y3Y swaption is expected to fall with 1.26 basis points
calculated from -0.0778/-0.0619=1.26.

We actually find that 3Y1Y, 2Y2Y and 2Y3Y swaptions all have negative
sensitivities as all the other instruments except for shorter periods. We are illustrating
the positive values in the figure above in as a statement of determine which
instruments in the ‘grey’ area there should be included into the vega sector. As only
one of the 2Y expiry options has a positive sensitivity in smaller subperiods, we will
remark the 2Y expiry as a decisive part of the vega sector for the rest the analysis.

We are now able to identify and analyze how the second and third eigenvectors
impacts the implied volatility surface.

Second eigenvector

The interpretation of the second principal component revealed from a PCA,

describes the steepness of the volatility curves. Here with different option expiries
but with the same underlying swap rate. This eigenvector captures the dynamics of
the term structure of each tenor for instrument in the vega sector. An increase in the
second factor will lead to an increase in options with a longer swap rate such as 7, 10
or 15 years relatively to shorter swap rates such as 1, 2 or 3 years. These behaviors are
also shown on figure below, where we observe the second factors sensitivities for all
input variables.

45
 Figure 8 Second eigenvector of vega sector, USD surface, Date: 2015-07-03

Here we see that the expiry dimension determines the level of the sensitivity for
all instruments. As an illustration look at all instruments with a 15 year expiry
regardless of the underlying swap rate. All same expiry instruments surfaces around
the same sensitivity more or less. Still the shorter expiry options, the more decrease
the sensitivity for all long end tenors. As the tenor increases, options tend to move in
a positive direction. Here could a curve trade involve a short expiry swaption with a
large negative weighting and a long expiry swaption with the same tenor, but with a
positive weighting as we expect these two positons to go in opposite directions. The
largest movements are given by the largest absolute weights. This means that the
most interesting curve trades is given for shorter tenors as these volatility curves tend
to change the most. As an example, if the sensitivities for 4Y3Y and 15Y3Y are -0.1367
and 0.0401, this means that if 4Y3Y falls with one basis point, the 15Y3Y swaption
implied volatility is expected to increase with 0.29 basis points by dividing the two
numbers.

Third eigenvector

The third factor describes the steepness of the implied volatility curves, likewise
to the second principal component but with a focal point on different tenors and

46
same option expiries instead. The instruments are now grouped and colored by their
expiries, while sorted by tenor in ascending order, as illustrated on figure 10.

The coefficients for third eigenvectors illustrate that same expiry options with a
2Y tenor starting point has a negative entry but increases to a positive value when
finishing at the 30Y tenor.

Figure 9 Third eigenvector of vega sector, USD surface, Date: 2015-07-03

The value of the weights is highest for shorter expiry options while long expiry
options seems to less affected by changes in the steepness than their counterparties.
A steepener trade must involve a short expiry swaption with a large negative
weighting and a long expiry swaption with the same tenor, but with a positive
weighting as we expect these two positons to go in opposite directions. It is a less
structural pattern that can be observed compared to the previous eigenvectors, and
2Y expiry options only all have positive entries for all tenors. The 2Y expiries options
doesn’t capture the same dynamic as the rest of the expiry options and cannot be
used in a relative value trade on the third principal component here. But as
mentioned earlier, sensitivities changes over time and we find days where 2Y expiry
options can be used in a relative value trade on the third principal component.

47
 The 2Y1Y and 2Y2Y have extra high positive entries; this could be related to the
gamma influence we mentioned earlier. If a relative value trade was interesting for
the third principal component on the 3rd of July 2015, a position would eventually be
taken on either 3Y, 4Y, 5Y or 7Y expiry options as these seems most suitable for a
trade on the steepness of the volatility curves with same expiry.

As an example, if the sensitivities for 4Y3Y and 4Y15Y are 0.1219 and -0.0795, this
means that if 4Y3Y falls with one basis point, the 4Y15Y swaption is expected to
increase with 0.6525 basis points by dividing the two numbers.

Summary

The eigenvectors of three principal components of the vega sector for EUR
market are exemplified in the appendix. They depict the same dynamics as
pinpointed for the USD market.

Given this knowledge, we can conclude that the first principal component uses
both dimensions, expiry and underlying, meanwhile the second and third principal
components are dimension specific factors, explaining each a facet of curve
movements. By diversifying the different dynamics, these finding can now be used to
identifying trading opportunities but also for risk managers interested in hedging
against specific risk exposures.

Scaled eigenvalues
The market mechanism of the implied normal volatilities during the sample
period of 2010 t0 2017 has shown significant commonality. Throughout the sample
period, the first principal component explains over 81 % on average for EUR and on
average 73 % for USD swaption matrix after introducing the two year expiry
restriction. Interestingly, the first three principal components together explain
around 98 % of the vega sector for the entire period investigated for both USD and
EUR. Nevertheless, the three principal components for both USD and EUR appear to
consistently explain practically all of the variance found in the implied volatility
surface data.

48
 Figure 10 - Scaled eigenvalue on a PCA for vega sector USD

Figure 11 - Scaled eigenvalue on a PCA for vega sector EUR

The figures above reveals the structure of the implied volatility surface, sorted by
scaled eigenvalue summing to nearly one each day. The scaled eigenvalue of the first
principal component fluctuates much more for USD than EUR, but we see a decrease
on both market around primo 2014 and again primo 2016 followed by an increase in
Medio 2016. The distribution of scaled eigenvalue between the three first

49
components often varies over time due to changes in the eigenvectors. These
changes often occur due to macroeconomic events such as changing market regimes.

It is by far the first two principal components that explain almost all of the
movements in the two datasets. Given the interpretation of the eigenvectors earlier,
it must imply that most of the exposure for vega sector swaptions on the volatility
surface is given by changes in the overall implied volatility level. Also, there is an
inverse relationship between the first and second scaled eigenvalues for both EUR
and USD, where a decrease in the first scaled eigenvalue will result in a similar
increase in the scaled second eigenvalue. The third eigenvector are quite constant
during the whole period for both markets. The second principal component explains
around 7.9 % of the variance while the third component explains almost 3.5 % of the
variance. It is difficult to interpret the fourth principal component and further as these
often are difficult to identify as there percentage of the overall variance are so low.
We are limiting ourselves to only three principal components as these accounts for
most of the variance explained.

As the numbers, shown on the two figures above, are computed from a daily 630
days PC analysis, deviations in the explaining ratio are an excellent and robust test for
warnings about model issues or even sudden changes in the behavior of the volatility
surface.

From observing the sensitivities of the first eigenvector of the USD volatility
surface from 1. June 2012 (first day of the 630 day period) and six month forward the
interpretation of the first two eigenvectors are not as given in the section before.
Here is the interpretation of the first eigenvector expressed as the steepness of
volatility curves, with same tenor but different expiries, while the second eigenvector
expresses the overall implied volatility level. This demonstrates a switching tendency
between the interpretation of the first and second eigenvector. The exact same
condition applies for the two eigenvectors during 2014. This is a clear signal of the
eigenvectors being unstable in subperiods. Interestingly, this instability only occurs
for the USD market, while all three eigenvectors are stable for EUR market.

The instability of eigenvectors is somehow understandable when dealing with

instruments driven by different dynamics. If we divide between a mix of short expiry
and long expiry options where longer expiry options volatilities not always are not
affected in the same direction as the short end options that have larger dispersion. If
the long end expiry and tenor instruments were excluded, eigenvectors would
become more stable due to the similarity in the dynamics. This means an
50
investigation of only long expiry options could also be applied, depending on the
desired specifications. For example are pension funds more concerned with buying
long term protection on their portfolio and participate in the long tenor and long
expiry swaption market.

Clustering
As we discussed, the first and second have by far the largest scaled eigenvalues
compared to the third eigenvalue for the vega sector. An interesting discovery can be
found by comparing options sensitivities of the first two eigenvectors for the whole
volatility surface. We found that the first two eigenvectors determines the majority of
the shape for all instruments on the implied volatility surface. The eigenvectors given
by a PCA can be used as a visualization tool for detecting clusters in the dataset. By
investigating the positioning of the first two eigenvectors of the whole implied
volatility surface and not only the vega sector, we discover clusters for both groups -
short and long expiry options, as seen on figure 13. Notice how instruments with
either long or short expiry tend to group together on the plot. These clusters support
the decision of dividing options into a gamma and a vega segment, as we discussed in
the section ‘Distribution of the volatility surface’. PCA can commonly be used for
identifying clusters within datasets. This is a powerful tool if we only need to consider
a few principal components, as in this situation. A comparison with the sensitivity of
the third eigenvector can also be applied for better understanding of the behavior of
the different instruments, but seems less interesting as this component only
describes a little part of the total variance in the data.

51
 Figure 12 Cluster analysis of the whole USD volatility surface, Date 18 Oct. 2016

As seen on the figure above, a clear picture can be observed, where we

introduces three classifications. The sensitives of the two eigenvectors are displayed
on the axes. The green marks indicate swaptions with an expiry under two years, the
blue ones are options with precise two year expiry and the red marks are options
maturing in three years or more. On the right side of the figure are all short options
placed, while a compact clusters for the long expiry options are shown on the left
side. Two years options are placed in the middle of the two clusters but are more
identifiable with the long expiry options cluster. The clusters appear clearly in each
daily PCA of the whole surface for both USD and EUR.

Factor Residuals
The interpretation of the three first factors has now been considered and we will
now focus on looking on the residuals by redefining factors as residuals.

52
 From PCA, we start with n input variables/dimensions, and modelling this n
factors will appear (k=n). No residual is outlined in the n factor PCA, however the
unused factors as we only choose three factors, can be expressed as residual. Here

𝜀1𝑡 𝑛 𝑒𝑖1
𝑡
( ⋮ ) = ∑ 𝛼𝑖 ∙ ( ⋮ )
𝜀𝑛𝑡 𝑖=𝑘+1 𝑒𝑖𝑛

Our first look at a PCA looks like

𝑦1𝑡 𝑛 𝑒𝑖1
𝑡
( ⋮ ) = ∑ 𝛼𝑖 ∙ ( ⋮ )
𝑦𝑛𝑡 𝑖=1 𝑒𝑖𝑛

But after adding the residuals it converts into

𝑦1𝑡 𝑘 𝑒𝑖1 𝜀1𝑡

𝑡
( ⋮ ) = ∑ 𝛼𝑖 ∙ ( ⋮ ) + ( ⋮ )
𝑦𝑛𝑡 𝑖=1 𝑒𝑖𝑛 𝜀𝑛𝑡

When investigating the factor residual, we are interested in one of the most
fundamental part of relative value analysis, the indication of relatively rich and cheap
instruments. If an instrument has high negative residual value, measured in basis
points, this indicates that this specific instrument is relatively cheaper than the rest of
the market. Meanwhile if an instrument has a high positive residual, this instrument is
rich and therefore relatively more expensive than the rest of the market. Applying the
factor residuals in this paper, we are able in identifying one or several instruments
appropriate candidates for a chosen relative value trading strategy.

After moving rearrange the equation above, a one factor residual is calculated as

𝑟𝑒𝑠𝑖𝑑𝑢𝑎𝑙 = 𝐿𝑎𝑠𝑡 𝑜𝑏𝑠𝑒𝑟𝑣𝑒𝑑 𝑣𝑎𝑙𝑢𝑒 − 𝑀𝑒𝑎𝑛 − 𝐿𝑜𝑎𝑑𝑖𝑛𝑔 1 𝑓𝑜𝑟 𝑒𝑎𝑐ℎ 𝑣𝑎𝑟𝑖𝑎𝑏𝑙𝑒

∗ 𝑆𝑐𝑜𝑟𝑒 1

As the equation shows, the residuals can basically be regarded as the original
projected data points minus the fitted data points. First we subtract the mean from
the original data. By multiplying each coefficient in the factor loading with the
equivalent factor score, we are able construct new coordinates as a fitted projection
given the original coordinate system. This leaves us with the residual from a one
factor model.

53
 A two factor residual is calculated from the same principal as the first factor
residual, here adding our projection from the second factor as

𝑟𝑒𝑠𝑖𝑑𝑢𝑎𝑙 = 𝐿𝑎𝑠𝑡 𝑜𝑏𝑠𝑒𝑟𝑣𝑒𝑑 𝑣𝑎𝑙𝑢𝑒 − 𝐶𝑜𝑙𝑀𝑒𝑎𝑛 − 𝐿𝑜𝑎𝑑𝑖𝑛𝑔 1 𝑓𝑜𝑟 𝑒𝑎𝑐ℎ 𝑣𝑎𝑟𝑖𝑎𝑏𝑙𝑒

∗ 𝑆𝑐𝑜𝑟𝑒 1 − 𝐿𝑜𝑎𝑑𝑖𝑛𝑔 2 𝑓𝑜𝑟 𝑒𝑎𝑐ℎ 𝑣𝑎𝑟𝑖𝑎𝑏𝑙𝑒 ∗ 𝑆𝑐𝑜𝑟𝑒 2

A three factor residual is calculated as

𝑟𝑒𝑠𝑖𝑑𝑢𝑎𝑙 = 𝐿𝑎𝑠𝑡 𝑜𝑏𝑠𝑒𝑟𝑣𝑒𝑑 𝑣𝑎𝑙𝑢𝑒 − 𝐶𝑜𝑙𝑀𝑒𝑎𝑛 − 𝐿𝑜𝑎𝑑𝑖𝑛𝑔 1 𝑓𝑜𝑟 𝑒𝑎𝑐ℎ 𝑣𝑎𝑟𝑖𝑎𝑏𝑙𝑒

∗ 𝑆𝑐𝑜𝑟𝑒 1 − 𝐿𝑜𝑎𝑑𝑖𝑛𝑔 2 𝑓𝑜𝑟 𝑒𝑎𝑐ℎ 𝑣𝑎𝑟𝑖𝑎𝑏𝑙𝑒 ∗ 𝑆𝑐𝑜𝑟𝑒 2
− 𝐿𝑜𝑎𝑑𝑖𝑛𝑔 3 𝑓𝑜𝑟 𝑒𝑎𝑐ℎ 𝑣𝑎𝑟𝑖𝑎𝑏𝑙𝑒 ∗ 𝑆𝑐𝑜𝑟𝑒 3

Factor residual can also be used to identify cheap trades expressed as a macro
view. Residuals can pinpoint a cheap entry position for an investor in a specific
instrument. As an example, if implied volatility falls, which instruments has fallen
more relatively to others and therefore best to take a short position in.

Correlation with market indicators

We can further explore the interpretation of the three principal components by
creating a time series for these. This is done by using the factor scores of each
component. A statistical tool such as PCA assimilates market dynamics by designing
a backward view, but incorporating fundamental variables will enlighten us to
visualize forward looking expectations. Let us here define fundamental variables as
economic variables. As mentioned earlier, PCA composes mean-reverting
uncorrelated factors which often can be interpreted though economic viewpoints
such as business cycles etc. This is a strong property of PCA which enables us to
investigate and understand the ever-changing dynamics of the financial markets by
gaining a symbiotic relationship between economic and statistical analysis. The
viewpoints are constructed by analyzing the correlation between the fundamental
variables and the uncorrelated factor scores. We are therefore able to pinpoint
potential future risks.

The two figures below illustrate the evolution of factor scores over the 630 day
period for USD and EUR. The factors scores can be seen as volatility in bps as
mentioned in the method section. The first principal components behavior is cyclical
and moves within the same level in longer periods eventually reverting back to its

54
mean. In the section in scaled eigenvalues, in Medio 2015 the second PC for USD
percentage of total variance explained was around 40 % and almost explaining the
same amount of total variance as the first principal component. We can also observe
correlation between the first and second PC for EUR during subperiods. The third PC’s
fluctuates around its mean and moves smoothly throughout this period. All principal
components are mean reverting but at a different pace.

From the figure below, we see the three factors scores over time. The interesting
part is to determine if any of these factors should revert to its long term mean in the
nearest future. The first factor often produces a low speed of mean reverting, which
makes sense since this can be considered as a macroeconomic determined factor.

55
Figure 13 –Movement in factor scores for USD and EUR

56
 We are now interested in exploring the principal components further, by
comparing the scores with fundamental economic variables. As an example we will
look at the correlation between principal components and fundamental variables.
Firstly we consider a 630 days correlation on the 3rd of July 2015. Secondly we regard
the absolute average of the rolling 630 days correlation for the entire period as well to
identify consistent market interactions with the principal components. Absolute
average is chosen as some indicator varies in being negatively or positively correlated
to the components. Calculating with a regular average can compose result with small
correlation in these situations

One of the chosen fundamental variables is the VIX index, which is regarded as
an investors ‘Fear gauge’. It is an index that measures the implied volatility for the
current stock market index of S&P 500. Given this index, we can investigate if there is
any relationship between equity market volatility risk and the principal components
of the implied volatility surface. The following stock markets are included as
fundamental variables; DAX, NASDAQ, S&P 500 (SPX) and Euro Stoxx 50 (SX5E).

The generic MOVE index is also introduces. This is constructed by Merrill Lynch
and measures a weighted average9 of ATM normalized implied volatility of the yields
to expiry of long term (maturities of two, five, ten and 30 years) Treasury options over
a one month period. Other notable fundamental variables are Crude Oil, German and
American government bonds, 5Y swap rates on EUR and USD. We would also like to
compare the dynamics to two other huge swaptions markets. Here Japan and United
Kingdom, so 5Y5Y swaptions volatilities are also included for JPY and GBP.

If the third principal component is uncorrelated with chosen macroeconomic

variables this could indicate that this is a “pure” relative value factor, meaning it is
dictated by a statistical component rather than macroeconomic changes. Equally
factors are often more uncorrelated to these variables due to the fact that higher
number factors are more driven by statistical analysis than by macroeconomic
events. So if we could give a reasonable interpretation of the fourth principal
component, this could be used as a relative value factor.

The table below shows that the first USD principal component is highly
correlated with several economic variables on the 3rd of July 2015. Seemingly, there is
a negative linkage with all stock indexes used, which we mentioned above. This can

9
The average weighting is set as 20 % two-year, 20 % five-year, 40 % ten-year and 20 % 30-year bond.
This weighting is chosen by estimates of option trading volumes for each of the different maturities.
57
be compiled to a high negative correlation with USD stock markets and EUR stock
markets, where correlation is most significant with the USD stock market. The
situation is actually the same for the first principal component for EUR. In the case of
absolute average rolling correlation, USD stock markets tends to be better at
explaining the most dominate dynamic on the volatility surface for both EUR and
USD swaptions compared to EUR stock markets. This states that these two
components have a high linkage toward essential risk variables.

Importantly, the first principal component for both EUR and USD are highly
positive correlated to the overall level of underlying swap rates in their denominated
currency. We see that the first principal component for USD is correlated the overall
level of the implied volatility for JPY swaptions.

We see that the second USD principal component is highly correlated to the US
Generic Government 5Y and 10Y bonds on this day, but less correlated over the entire
period. Moreover there is a high correlation between the second principal component
on EUR and the foreign exchange rate between EUR/USD. The second factor for USD
can be considered as a less definite macroeconomic factor as correlation is high at the
current date, but is overall at a lower level on average for USD, while the second
principal component for EUR is clearly more affected by market indicators
consistently.

Interestingly, crude oil future prices and PC3 are highly negatively correlated on
the observed date, but in general the differential between tenors are unaffected by
fundamental variables. The MOVE index signals interest rate movements as
uncertainty change in the bond markets, and are overall fairly correlation to the three
principal components for USD. The index is however not an independent
fundamental variable when comparing to the dynamics of the implied volatility
surface. The MOVE index is widely used as an indicator to proxy tensions in fixed
income markets (inflation, deflation and concerns regarding government debt) and
could be loosely considered as a liquidity indicator, which affects all dynamics found
for USD.

Looking at the 630 day rolling correlation for the entire period, taking the mean
of the absolute correlation for each market indicator towards the third principal
component, we observe an overall lower level of correlation for all outcomes. Some
of the crosses have a significant correlation to EUR or USD for the whole period, but
all at level lower than 50 %. We therefore roughly postulate that the third principal
components as a purely statistical component. This gives some possibilities in
58
constructing relative value trades, as this component have relatively little exposure
towards market indicators and therefore market directions. This is nice feature in a
trading perspective, as this shows that taking a position on this component,
performance will not be affected by macroeconomic trends or changes. Statistically,
macroeconomic exposure will rises by trading on the second principal component and
even more if a position is made on the first principal component.

We see the first principal components for USD and EUR are highly correlated
with a value of 0.737 measured on 03-07-2015. Overall for the entire period, for these
components are not as correlated compared to this date. The third principal
component for USD and EUR are not significantly correlated to any other principal
components, which isn’t surprising.

By using the times series of the factor scores on the daily PCA as input for an
Ornstein-Uhlebeck process we are able to assess the speed of mean reversion for
each factor. Here, a much higher level of speed of mean reversion is observed for the
third factor than for the second and first factor. The specifications for speed of mean
reversion, variance, days to converge 50 % and 90 % to the mean can be found in the
appendix. This will be introduced further in the strategy section on trading signals.

59
Table 1 - 630 days rolling correlation for USD and EUR factor on 03-07-2015 and absolute mean for entire period
Correlation on 03-07-2015
USD VIX SPX OIL 5Y US Gov 10Y US Gov DAX SX5E MOVE 10Y GER Gov 5y5y GBP 5y5y JPY EURUSD USD 5Y swap rate EUR 5Y swap rate NASDAQ
Factor 1 0.034 -0.863 -0.432 0.626 -0.001 -0.690 -0.728 0.199 0.599 0.245 0.652 0.315 0.605 0.718 -0.830
Factor 2 -0.126 -0.176 -0.178 -0.716 -0.799 -0.156 -0.155 -0.482 -0.317 -0.758 -0.141 -0.186 -0.721 -0.365 -0.213
Factor 3 0.223 0.340 0.680 -0.004 -0.452 0.435 0.369 0.732 -0.614 0.169 -0.429 -0.841 0.070 -0.468 0.378

Absolute average rolling correlation

USD VIX SPX OIL 5Y US Gov 10Y US Gov DAX SX5E MOVE 10Y GER Gov 5y5y GBP 5y5y JPY EURUSD USD 5Y swap rate EUR 5Y swap rate NASDAQ
Factor 1 0.203 0.596 0.408 0.731 0.743 0.398 0.526 0.423 0.658 0.543 0.621 0.531 0.762 0.646 0.608
Factor 2 0.393 0.324 0.337 0.432 0.434 0.148 0.392 0.472 0.403 0.426 0.273 0.279 0.380 0.415 0.355
Factor 3 0.182 0.309 0.283 0.081 0.167 0.426 0.338 0.435 0.217 0.169 0.171 0.432 0.109 0.202 0.305

Correlation on 03-07-2015
EUR VIX SPX OIL 5Y US Gov 10Y US Gov DAX SX5E MOVE 10Y GER Gov 5y5y GBP 5y5y JPY EURUSD USD 5Y swap rate EUR 5Y swap rate NASDAQ
Factor 1 -0.105 -0.715 0.669 0.155 0.564 -0.536 -0.546 0.083 0.881 0.534 0.809 0.593 -0.174 0.948 -0.657
Factor 2 0.062 0.117 -0.562 -0.292 -0.472 0.388 0.309 0.474 -0.399 0.203 -0.113 -0.711 -0.259 -0.254 0.190
Factor 3 0.074 -0.544 0.035 -0.703 -0.511 -0.510 -0.547 -0.074 0.086 -0.377 0.281 0.041 -0.694 0.130 -0.574
Absolute average rolling correlation
EUR VIX SPX OIL 5Y US Gov 10Y US Gov DAX SX5E MOVE 10Y GER Gov 5y5y GBP 5y5y JPY EURUSD USD 5Y swap rate EUR 5Y swap rate NASDAQ
Factor 1 0.372 0.490 0.267 0.181 0.318 0.507 0.414 0.428 0.609 0.556 0.491 0.425 0.253 0.729 0.539
Factor 2 0.575 0.379 0.371 0.622 0.754 0.534 0.501 0.355 0.579 0.417 0.386 0.739 0.622 0.415 0.537
Factor 3 0.257 0.257 0.298 0.261 0.186 0.304 0.245 0.282 0.133 0.273 0.262 0.162 0.233 0.159 0.324

Correlation on 03-07-2015
USD\EUR Factor 1 Factor 2 Factor 3
Factor 1 0.737 0.159 0.456
Factor 2 0.396 0.219 0.554
Factor 3 -0.326 0.696 0.038
Average correlation for entire period
USD\EUR Factor 1 Factor 2 Factor 3
Factor 1 0.450 0.575 0.257
Factor 2 0.490 0.380 0.256
Factor 3 0.267 0.371 0.298

60
 The correlation between the external variables and the principal components
evolves as the relationship changes. Monitoring rolling correlation is therefore
needed to be considered before trading on any of the principal components.

From the start of the investigated period, we are able to register an overall very
high negative correlation between the first factor and SPX index but this correlation
decreases over time to a moderate level. Meanwhile the commodity Oil moves with a
small margin during this period at a moderate level of correlation. The graph below
helps in identifying the first factor as highly correlated to risk variables such as
commodity and stock prices denominated in USD. By following up on this, factor 1 -
stated as the overall volatility level behaviors in both dimensions are regarded in the
same manner as some very risky assets. Rolling correlation between the first principal
component and SPX fluctuates highly and when correlation is low in 2014, the second
principal component is highly correlated with the SPX index. The transition happens
at the exact same time we encounter instability between the first and second
principal component.

Figure 14 - Rolling correlation between USD PC’s against SPX and OIL

At most, we illustrate that external data regarding macroeconomic trends and

events are helpful in explaining movements in the principal components and also
gaining valuable forward looking information as part of the analysis. This information
can be used to forecast future behaviors of the first and second principal components
and visualize future expected risks.

61
Applying the setup

Hedging with very liquid assets

With the discovery of high correlation between factors and several fundamental
variables interesting possibilities arises. Typically, an investor would hedge a portfolio
of swaptions with swaptions. This can be a quite expensive hedge. Instead if we could
use fundamental variables to reduce exposure against the dynamics of the first
principal component, as this would definitely be a cheaper alternative. The most ideal
hedging assets have the following properties; they are cheap to take a long or short
position, easy access and liquid market for the security and highly correlated to the
risk exposure (here the three factors).

Here could one of the USD stock market indexes, SPX or NASDAQ, be used as
hedging asset. We witnessed negative correlation between the two indexes and the
first principal component and would therefore buy one of the indexes as a hedge.

We will now introduce a multiple regression on principal components for hedging

different exposures by using liquid assets such as equities, currency crosses and fixed
income bonds. Testing different combinations, only a few variables are considered
and our goal is to find a solution with a high adjusted R squared value. A bias on
finding variables denominated in same currency is created for not introducing
currency risk. As a setup, we are interested in a dynamic hedge, where we are
rebalancing the hedge portfolio over time. As principal components are uncorrelated,
due to the orthogonal eigenvectors, any position can be regressed directly to each of
the found risk factors on the implied volatility surface. The factor scores for the first
principal component are carried out as the dependent variable in the regression
analysis.

So far we have only worked with level data, but differenced data is preferred for
risk management purposes, as focus here lies in capturing short term dynamics and
rebalancing the hedge over a period. Level data are however better at capturing long
term structural behavior of relationship between the implied volatilities. Statistically,
there is a loss in information regarding level impact if differenced data is used. But
non stationarity are a concern when choosing level data, especially for the first
principal component, where we will apply the factor score to a multiple regression
analysis. Here is differenced data naturally more stationary and not a concern.
Therefore is differenced data applied instead, as this gives the craved prerequisites
for the hedge. Applying a hedge on differenced data on a short dataset is useful for

62
short term hedging purposes. However for long term hedging circumstances, level
data on implied volatilities could be an option.

We are only interested in hedging the first principal component as this explains
on average 75 % of the total variance in the data for USD. Given the scaled
eigenvalues during the investigated period from 2010 to 2017 we can see that the first
factor is a dominate parameter in explaining the majority of the movements in the
implied volatility surface. A hedge against the first principal component will give a
non-directional position against shifts in the overall implied volatility surface.

Weekly changes are composed from Wednesday to Wednesday as prior to avoid

effects from Monday opening and Friday close. These changes can be expressed as
linear combinations of the form

𝑥𝑥𝑌𝑥𝑥𝑌(𝜎𝑁 (𝑇) − 𝜎𝑁 (𝑇 − 1)) = 𝒆𝑇𝑝 𝑿 = 𝑒𝑝,1 𝑋1 + ⋯ + 𝑒𝑝,𝑝 𝑋𝑝

where 𝑥𝑥𝑌𝑥𝑥𝑌(𝜎𝑁 (𝑇) − 𝜎𝑁 (𝑇 − 1)) is the absolute difference in implied volatility for
a specific instrument 𝑥𝑥𝑌𝑥𝑥𝑌 from time 𝑇 − 1 to 𝑇. Here are 𝑿 the uncorrelated
principal components with the factor loadings or weights 𝒆𝑇𝑝 .

When testing the duration of the length of weekly changes, by increasing the
interval, our attempt to regress the first principal component against fundamental
variables results in a lower degree of rolling adjusted R-squared value. As the weekly
changes are differenced data, PCA performs better for short datasets when regressed
against several fundamental variables. This corresponds to similar observations from
“Market Musings”10 report on Relative value across the US swap surface. Intuitively
this makes sense as differenced data capture short term changes in implied volatility
dynamics and short datasets provide such information. From a hedging perspective,
traders often want to eliminate some exposure on their trading book against short
term changes in the implied volatility dynamics, so we are using weekly change data
and only concentrating on a short historical view.

The computed factor scores, used for the rolling multiple regression analysis, are
settled from a PCA on weekly changes for a 120 days period. The rolling correlation
between the factors scores of the first principal component and the fundamental

10
TD Securities (2015), “Market Musings — Relative Value Across the U.S. Swap Surface: A PCA
Approach”

63
variables discussed earlier, are not as significant as with level data. However several
of the fundamental variables have been transformed into differenced data as well. As
stated below, NASDAQ, 5 year USD swap rate and VIX index.

𝑦 = 𝛽1 ∆(𝑈𝑆𝑆𝑊5) + 𝛽2 ∆(𝑉𝐼𝑋) + 𝛽3 ∆(𝑁𝐴𝑆𝐷𝐴𝑄)

where 𝑦 represents PC 1 scores, 𝛽1 is the amount of 5y USD payer interest swap , 𝛽2

is the amount of VIX futures and 𝛽3 is the amount of NASDAQ index stocks used to
produce a linear hedge on changes in the overall implied volatility level, which is the
PC 1 scores. NASDAQ and SPX are almost similar, but NASDAQ was a slightly better
fit here. Important to know is that the different securities can be traded in both a long
and a short position.

After converting from level data to differenced data and reducing the period to
30 weeks instead, the first eigenvector for USD is now stable throughout the entire
period. This tell us that, over a short term, risk managers main focus should be in
reducing the exposure toward the overall implied volatility level.

Figure 15 - Rolling beta coefficients on 5y swap rate, VIX and NASDAQ to first principal
component scores. Last figure shows rolling adjusted r squared.

The beta coefficients are not stable during the whole period and expresses
against scores in bps, therefore all coefficients should be divided by 100. Some

64
subperiods however are the hedge nonexistence as the r squared value close to zero.
By converting from level data to differenced data and only investigating a 30 week
period, eigenvectors are stable during the entire period, where the three eigenvectors
clearly depicts the interpretation given earlier.

Finding several numbers of liquid hedging assets in the market with adequately
significant correlations to the first principal component - implied volatility level risk
can be reduced additionally with a multiple regression. From a statistical point of
view, selecting more hedging assets will boost the coefficient of determination (r-
squared) closer to one.

The simplicity of the method using r-squared should be considered in a multiple

regression when choosing more than one hedging asset. Choosing more than one will
introduce a risk of multicollinearity. This means that hedging assets or independent
variables are in risk of not being entirely independent. As it is commonly known
markets can be more or less correlated. The risks of introducing multicollinearity can
therefore be quite substantial when applying several hedging assets in one
regression. To reduce this problem we can focus on maximizing the adjusted R
squared when applying different hedging assets. The adjusted R squared differs from
the simple R squared, because it adjusts for degrees of freedom. By focusing on
maximizing the adjusted R squared hedging assets will only be allowed in a multiple
regression if it makes a noteworthy contribution to the value of the adjusted R
squared.

The coefficient of determination expresses the percentage of risk possible to

eliminate by trading with the hedging assets. So this indicates how effective the
linear hedge is by using VIX futures, NASDAQ index and 5 y USD swap rates.

The adjusted R squared value range around 0.2 and 0.51 75% of the time,
however the distance between minimum and maximum shows great inconsistent for
a linear hedge against the changes in the overall implied volatility if a weekly hedged
is chosen.

Min 1st Qu. Median Mean 3rd Qu. Max.

0.00099 0.2015 0.3501 0.3476 0.5113 0.7033

This shows us that some of the directional exposure can be dynamically hedged
cheaply with stocks, VIX futures and swaps. However the fluctuations in the r squared
value varies much and this indicates the multiple regression don’t generate the

65
presumed hedge for the entire period. Other hedging intervals can be used to
perhaps find a more significant result.

We need however to acknowledge a 30 weeks factor scores presented as the

independent variable, possibly aren’t extremely robust. PCA is sensitive to outliers
since the principal components are constructed from a covariance matrix, which is
highly sensitive to outliers. Sudden jumps in the implied volatility surface could leave
us with a rigid result.

Turning the regression around between explanatory and response variables, we

have an asset portfolio consisting of equity and bonds instead, where we could hedge
downside risk by taking a position in a swaption. Here as a hedge spanning over
multiple asset classes. This can be an efficient way to hedge risk exposure such as
volatility risk during volatile periods.

A portfolio of swaption exposure towards changes in the implied volatility level

will be reduced to some extent, by dynamically hedging against the first principal
component scores with the dependent variables used in the regression. As these
variables are asset that are considered very liquid and can be seen as cheaper hedge
alternative relatively to hedging with swaptions. The hedge is however non-existent
in subperiods.

Strategies for relative value trades

Overall, there is a great existence of different volatility trades. The most common
strategy is by selling options and dynamically delta hedging the exposure towards the
underlying asset. This is a play for an investor to profit from market known feature,
where implied volatility tends to exceed realized volatility.

In relation to the vega sector for swaptions, constructing a relative value trade is
recognized as vega trading. The pricing of ATM options mainly depended on changes
in the implied volatility as expiry increases. As vega increases with expiry, decreases
gamma. If gamma is low for an option, changes in the underlying have no significant
effect on delta. This implies that delta hedging has no impact on the P&L for vega
sector swaptions.

66
 Due to the low exposure to changes in the underlying for straddles in the vega
sector, this allows for conceptual base the volatility surface with a statistic tool as
PCA and therefore hedge against the implied volatility level, from the interpretation
of the first principal component, instead of delta hedging. Given the knowledge of
the dynamics of the implied volatility surface after an interpretation of the
eigenvectors for the three principal components, three different kinds of trades can
be constructed.

Let us summarize the different exposures from different types of option trades.
Firstly, trading with one instrument gives an exposure to the absolute level of the
implied volatility. Secondly, trading with at least two instruments gives a hedge
against the exposure to the absolute level of the implied volatility and an exposure on
an implied volatility curve.

Daily run of mean reverting process for the three principal components to find
opportunities by evaluating the speed of mean reversion. After choosing the desired
position on one of the principal components, we display the relevant residuals. The
residuals expresses if an option is relatively cheap or rich and therefore helps
identifying assets from a relative value viewpoint. A larger difference between the
rich and cheap instrument will signify a profitable opportunity. If the trade consists of
two instruments, a relevant hedging ratio is calculated. Lastly, the position will be
unwound after a shorter holding period.

Salomon Smith Barney (2000) shows that trading signals differs for level data
and change data. The timing of trading signals is different for the two approaches.
This is however applied by tracking the P&L for butterfly trades on swap rates and not
for swaptions. Given these observations, they choose level data for timing of trades
as these produces generally higher Sharpe ratios. We will however only focus on level
data for trading purposes.

First PC
Given a position in a single instrument is susceptible to a shift in volatility surface
level. If high speed of mean reversion was observed for the first principal component
and this is far away from the mean, we can trade on a change in the direction of the
implied volatility surface. Such a situation is however rather unlikely given the
outlined properties, mainly driven by macroeconomic fundamentals. But perhaps this
relationship can be exploited. If a high correlation between the factor score and 5Y

67
swap rate is observed and we anticipate the swap rate to move in favor of the mean
reversion of the principal component very soon this could be an opportunity of a
profitable trade. As directional positions normally have a low speed of the mean
reversion, investors need to be patient and expect longer holding periods. Or else
attractive relative value trade on the first principal component seems diluted.

The first principal component is the main interest factor from a risk management
perspective; meanwhile the second and third principal component is the main
interest for relative value trading opportunities as they serve a higher speed of mean
reversion.

Second PC
As the second principal component represents the dynamics of the implied
volatility curves for dimensional equivalent tenors, a positioning for a change in the
slope could be considered. Here either by anticipating a curve steepener or flattener.
If a relative value opportunity suddenly appears, an investor could take a position by
selling a rich and buying a cheap swaption with same tenor and different expiry. The
rich and cheap assets can be identified for the one factor residuals. Before an investor
trade, the second eigenvector sensitivities of the two instruments have to be moving
in different directions when trading on changes in the slope. The investor isn’t
exposed the direction of the implied volatility level, but are exposed to the dynamics
of the third principal component. The exposure should be hedged if the scaled
eigenvalue of the third principal component is high.

Third PC
Last strategy takes a view on the implied volatility curves, with the same expiry
options but with different tenors. Trading on the tenor dimension is also
characterized as a position on the slope. So here is a curve steepener or flattener also
considered. Here rich and cheap instruments are selected from a two factor residual.
For example, buying a 10Y2Y straddle and selling a 10Y10Y straddle and perhaps
unwind after one month, the position would be a 10Y2Y minus 10Y10Y implied
volatility position, with almost none exposure towards curve steepness and towards
realized volatility. The weighting of the eigenvector for the third principal

68
components should always be checked before entering a position, in the sense as
trading on the second principal component.

Implied volatility curves set on the same tenor tend to be stable in the longer
tenors. This is captured by the sensitivities for the second eigenvector. Here were
weights close to zero for the longest tenors. This gives a picture of trading on long
tenor curves doesn’t seem to be an attractive choice.

Optimal holding period

The upper limit of the holding period is set to one month. A short holding period
is important as we can postulate that time decay of theta and changes in the
underlying swap rate will not influence the P&L for a chosen position. By only trading
with straddles a non-directional position is taken on movements in the underlying
tenor. We are able to hedge against the first principal component while not being
concerned with delta hedging as this is not required. This allows a relative value
trader, maneuvering in the vega sector, to focus on hedging exposures on principal
components instead.

For longer holding periods, large changes in the underlying can introduce an
unhedged jump risk in the P&L. The holding period should therefore be short so
straddle returns isn’t affected by both volatility and jump risk.

As mentioned before, theta is low for swaptions in the vega sector, but if holding
periods becomes very long, theta will begin to matter. This is primarily the case for
positions on the second principal component where theta is very different for the
options with same tenor but different expiries.

We are only able to calculate returns on relative value trades as P&L for changes
in implied volatility for any strategies. The implied volatility for ATMF swaptions on a
traded date will be compared to implied volatility for new ATMF swaptions at the end
of the holding period. We are aware that the underlying swap rates changes but is
assumed to be ignored due to the very short holding period and no precise way to
determine the value of a “old” non ATMF swaption straddle.

69
Hedging of factors
As the principal components can be viewed as risk factors, hedging ratios can be
calculated to reduce exposure towards these risk factors. This can be done by using
linear algebra by applying both eigenvalues and eigenvectors.

We want an optimal hedge ratio between two instruments. Whereas 𝑥 and 𝑦 are
the two instruments traded, being long in 𝑥 and short in 𝑦. Here can 𝑥 and 𝑦 be
expressed as linear combinations of factor loadings and scores in a PCA.

𝑥𝑡 = 𝑐𝑡1 𝑒𝑥1 + 𝑐𝑡2 𝑒𝑥2

𝑦𝑡 = 𝑐𝑡1 𝑒𝑦1 + 𝑐𝑡2 𝑒𝑦2

Here are 𝑥𝑡 and 𝑦𝑡 implied volatility levels of our two chosen instruments. These
are vectors as they are time series. 𝑐𝑡1 and 𝑐𝑡2 are the factor scores for the first and
second principal component. Both scores are given as time series vectors. 𝑒(𝑥,𝑦),(1,2) is
the factor loadings for the two instruments for each direction and is given as
numbers.

A solution for the minimum variance function11 w.r.t 𝛾 can be found as

min 𝑉𝑎𝑟(𝑥 − 𝛾𝑦)

𝛾∈ℝ

Where 𝛾 is defined as the hedge ratio. So hedging against the first principal
component, the sensitivities for the first eigenvector of two instruments are used. We
find the hedge ratio as
𝑒𝑥1
𝛾=
𝑒𝑦1

Where 𝛾 is the amount of instrument x while the amount of y is one, 𝑒1 is the first
factor sensitivity for the specific instrument.

While hedging against both the first and second principal components, we find
the hedge ratio as

𝑒𝑥1 𝑒𝑦1 𝜆12 + 𝑒𝑥2 𝑒𝑦2 𝜆22

𝛾= 2 2 2 2
𝑒𝑥1 𝜆1 + 𝑒𝑦1 𝜆2

Where 𝜆12 is the variance for the first principal component and 𝜆22 is the variance
for the second principal component. 𝛾 is the amount of instrument x while the

11
PCA unleashed

70
amount of y is one, 𝑒# is either the first or second eigenvectors sensitivity for the
specific instrument.

An example of a trading opportunity

For our construction of trade ideas, we have used the time series of all three
principal components scores for EUR and USD on each day from 2012 to 2017 created
from our daily PCA, to identify days with highest speed of mean reversion. A
maximum likelihood estimation method is used for producing the parameters of an
Ornstein-Uhlebeck mean reverting process. This gives a possibility signaling trading
opportunities by calculating speed of mean reversion of the computed principal
components scores. This gives us a statistical reason in deciding which principal
component we should investigate further on a specific date. This is illustrated by the
trading signal tables given in the appendix. The 06-10-2015 is revealed as a day with a
relatively high speed of mean reversion on the second principal components scores
for USD and the current factor score point is also far away from its mean. This day is
chosen as this is the first day in a six day period with high speed of mean reversion.
Moreover the expected count of days for the process to converge to its mean is 19.4
days12 with a 50 % probability. This seems like a reasonable opportunity given the 21
days holding period setup explained earlier.

Taking a position on the second principal component, a long and short position in
two instrument with the same tenor but different expiries will be considered. As
described in the interpretation of the second eigenvector, slope trades are only
interesting for short tenors because instrument sensitivities for longer tenors all
moves in the same direction. Swaptions with a 2Y tenor is listed in the table below as
this implied volatility curve reveals the largest differences for the one factor residuals.
The table also includes the sensitivities for each instrument towards the first and
second eigenvector. Lastly, implied hedge ratios against the first principal component
are calculated from the minimum variance framework for each instrument.

12
See trade signals in the appendix.

71
 Table 2 – Properties of the PCA for trade date 06-10-2015 on USD volatility surface.

2Y2Y 3Y2Y 4Y2Y 5Y2Y 7Y2Y 10Y2Y 15Y2Y 20Y2Y 25Y2Y 30Y2Y
One Factor Residual for USD
Bp -9.037 -6.487 -3.814 -3.050 -2.197 -1.469 -0.083 1.491 0.545 0.090

Sensitivity of the First Eigenvector

Weight -0.013 -0.084 -0.109 -0.110 -0.091 -0.055 -0.061 -0.056 -0.069 -0.073
Sensitivity of the Second Eigenvector
Weight -0.282 -0.217 -0.141 -0.093 -0.042 0.020 0.061 0.078 0.071 0.066

Implied Hedge Ratio against First PC

2Y2Y 3Y2Y 4Y2Y 5Y2Y 7Y2Y 10Y2Y 15Y2Y 20Y2Y 25Y2Y 30Y2Y
2Y2Y 1.000 0.152 0.117 0.117 0.140 0.232 0.208 0.227 0.186 0.175
3Y2Y 6.587 1.000 0.770 0.768 0.925 1.526 1.372 1.493 1.223 1.154
4Y2Y 8.558 1.299 1.000 0.998 1.202 1.983 1.783 1.940 1.590 1.499
5Y2Y 8.579 1.302 1.002 1.000 1.205 1.988 1.787 1.945 1.593 1.502
7Y2Y 7.118 1.081 0.832 0.830 1.000 1.649 1.483 1.613 1.322 1.247
10Y2Y 4.316 0.655 0.504 0.503 0.606 1.000 0.899 0.978 0.802 0.756
15Y2Y 4.801 0.729 0.561 0.560 0.674 1.112 1.000 1.088 0.892 0.841
20Y2Y 4.412 0.670 0.516 0.514 0.620 1.022 0.919 1.000 0.819 0.773
25Y2Y 5.384 0.817 0.629 0.628 0.756 1.247 1.121 1.220 1.000 0.943
30Y2Y 5.710 0.867 0.667 0.666 0.802 1.323 1.189 1.294 1.061 1.000

The largest residual difference is between the two instruments 2Y2Y straddle and
20Y2Y straddle. 2oY2Y has the highest positive residual indicating a cheap trade.
There are several possible rich trades with high negative residual such as 2Y2Y, 3Y2Y,
4Y2Y etc. From the sensitivities of the second eigenvector we expect the 2Y-7Y expiry
options to decrease and 10Y-30Y expiry options to increase when trading on the slope
of the 2Y implied volatility curve. Buying a 20Y2Y straddle and selling a rich trade
straddle, with a hedge against the first principal component, will give the exposure of
implied 20Y2Y volatility minus an implied xY2Y volatility position with very little
exposure to realized volatility.

Importantly, the second principal is only significantly correlated with the 5Y USD
government bond on the trading day, which nevertheless isn’t an indicator of risk.
This means that the factor can currently be characterized as a statistical component
and will not include any macroeconomic exposure in this situation, however over
time, the current state would most likely change. If we reflect on our earlier
observations on rolling correlation with fundamental variables for the two first
principal components, a change seems rather plausible.

From the figure below we see the implied volatility curve for 2Y tenor, for three
different days; on trading day, 21 days before and day of unwinding (21 days after
trade day). It appears that short expiry options implied volatility decreases over time.

72
 Figure 16 – Changes in the term structure of the 2Y tenor in a two month spectrum

20Y2Y has gained 0.3 bps 21 days after the trade, while all volatilities for the
shorted positions except for 2Y2Y have decreased.

For the position of selling 3Y2Y and buying 20Y2Y, the hedge ratio can be found
in the table above. Columns are considered as one unity, while lines are considered as
beta unity. By buying 20Y2Y and hedging against the overall implied volatility level,
1.493 units 3Y2Y are needed. If we unwind the position after a holding period of 21
days (one month), we would profit 0.74 bps.

P&L can be measured as the change in the implied volatility position as stated
just above. Given the different combinations, P&L can be found as

Buying 20Y2Y
Selling 2Y2Y 3Y2Y 4Y2Y 5Y2Y 7Y2Y
P&L in bp -0.347 0.740 2.848 0.583 0.417

We see all positions profits except for a short 2Y2Y straddle. Notice that the
weight of the first principal component for 2Y2Y is -0.013, which is close to zero,
could possibly be influenced by gamma. Given this information, one could consider to
exclude 2Y2Y as this position bears a lot uncounted risk.

73
Pitfalls of a PCA trade
In this section, we will illustrate an example of a relative value trading
opportunity, which eventually can’t be considered to be an opportunity due to an
unhedgeable first principal component.

From the figure below, we see that the first factor scores for USD in a PC analysis
have decreased in 2016 from late summer until late autumn. This decrease can be
explained by an overall fall in the implied volatility level for this period. By comparing
implied volatility values for all instruments from the entire period to the levels in
autumn 2016, implied volatility for all instruments are at its lowest level since mid-
2013. However this changes in mid-November 2016, where we observe a turnaround
in the trend of the first factor. Here increases implied volatility for all swaptions
substantially. This change is happening exactly the days after the presidential
election in USA.

Figure 17 – Time series of factor scores for USD

74
 We find speed of mean reversion relative high for the second principal
component on 15th of December 2016. However significant correlations between all
three principal components are observed since late November 2016, which is also
observable on figure 19. This is a problem for relative value trades on either the
second or third principal component. Hedges are very likely to break down during this
subperiod. A trade on the second PC with a hedge against first PC would not work
and therefore be exposed as well to the first PC during this period. So performance of
this trade would be driven by direction. Generally correlations between principal
components are zero for the whole analyzed timespan, but sometime correlation can
occur during subperiods. This is a major pitfall within the PCA framework on relative
value trading. Therefore should relative value traders consistently monitoring
principal component correlation before entering a trade. This will reduce the risk of
setting up a trade with a directional exposure or else can such trades eventually lead
to significant loses over time.

Correlation between principal components is not an issue from a risk

management perspective. Firstly, a manager wants to reduce the risk in first principal
component as this exposure is the most dominate risk. If this is done with other asset
classes, as we modelled earlier, a positive correlation between the first principal
component and another principal component will automatically also be a hedge
against the exposure of other risk factors in these subperiods. This will overall reduce
the exposure against multiple dynamics of the implied volatility surface
unintentionally.

The main pitfalls of trading options with the assistance of a statistical tool such as
PCA, is if we compare this to a one dimensional yield curve PCA setup, the structure
of the principal components has a more clear structure for the yield curve. Both the
second and third principal component of a PCA on the EUR and USD volatility surface
signifies curve steepness, but within different dimensions. Going back to the one
dimensional yield curve, the second principal component represents all information
regarding curve steepness, while the third principal component is less influential and
denotes all information on curvature in the exact same dimension. This can result in
that higher order eigenvectors on the volatility surface tends to be less stable
compared to the one dimensional yield curve. The instability of the first and second
eigenvectors for USD is somewhat a perfect example of this.

75
Conclusion

Throughout this paper, we have investigated the relative valuation in the interest
rate options markets. We have focused on implied volatility dynamics for long
expiries, with at least two year. The result from our empirical analysis of the EUR and
USD market from 2010 to 2017 identifies three different dynamics offsetting the
movements on the implied volatility surface. We have found strong evidence that the
implied volatility surface for EUR is mostly influenced by a directional market
mechanism. This is however not the case for the USD market, where the most
influencing mechanism varies between a directional dynamic and a steepness of the
volatility curves for same tenor’s swaptions. Overall we have evidence of three
significant dynamics on the implied volatility surface explaining over 95 % of the total
variance for both USD and EUR markets over the entire period. This is a consistent
result over the entire period from 2010-2017. In the analysis we have found three
significant dynamics represented on the implied volatility surface. The first dynamic
are interpreted as an overall implied volatility level factor. The two other dynamics
captures the implied volatility curves, each specifying a dimension on the implied
volatility surface.

By linking the principal components with economic and fundamental variables,

we have gained valuable insight into identifying important macroeconomic drivers for
implied volatility, which can help predict future movements of each dynamic. This is a
strong property to PCA, which is based on historical data and can therefore normally
be viewed as a backward-looking method. We found that the movements of the first
principal component for both USD and EUR are highly correlated with the level of the
US stock markets.

The paper has provided evidence of specifying a hedge against the first principal
component for USD using a multiple linear regression. Here we repeat a PCA on
differenced data and can conclude that the hedging strategy with NASDAQ, VIX and
the 5y USD swap rate has been successful in some subperiods. While through other
periods the hedging portfolio has shown little or no effect on reducing the exposure
toward the dynamics of the overall level on the implied volatility surface. Three
hedging assets were most significant in reducing the exposure. However, there is a
trade-off between increasing or decreasing the number of independent variables in
the regression. Adding hedging assets can increase the adjusted R-squared

76
significantly. This means that the percentage of hedging performance will rise,
however fewer variables will reduce transaction costs on the weekly rebalancing.

If traded swaptions suddenly stops to covary, perhaps due to market distress or

structural changes, a hedging or trading strategy within the relative value framework
will be pointless due to future expectation of co-integrated movements, as the
principal components are designed from a covariance or correlation matrix. As we
demonstrated in the methodology section concerning PCA, this tool assumes that
the directions with the most variance are the most interesting.

Furthermore, we introduced a trading strategy to identify ‘true’ relative value

opportunities in the EUR and USD markets. We found out that investors tend to trade
on the second or third principal component due to the higher level of mean reversion.
However only very few opportunities were found and we are therefore not able to
validate the strategy as a viable trading opportunity. However it seems that these
methods have been used by practitioners for ages. If an investor wants higher return
on the implied volatility position, they should exclude the long expiry and the long
tenor options as variation in implied volatility is relatively low. But the exclusion
comes with a risk, as theta and gamma are more likely to influence your position.

Finally we can acknowledge PCA as a practical tool to identify risk factors,

macroeconomic co-tendencies and trading opportunities. It must however be treated
with caution as there are several pitfalls when applying PCA to a relative value
analysis. A PCA on an implied volatility surface extracts higher order eigenvectors,
which tends to be less stable compared to the eigenvectors in a one dimensional yield
curve setup. The discovered instability of the first and second eigenvectors for USD
depicts this issue when applying PCA on an implied volatility surface.

77
Bibliography

PCA Unleashed – Credit Suisse 24 October 2012

Duarte, J., F. Longstaff, and F. Yu (2007): “Risk and return in fixed income arbitrage:
Nickelsin front of a steamroller,”

Duyvesteyn, J, (2015), “Riding the swaption curve”

Franco, Jose Carlos Garcia, (2003), “Maximum likelihood estimation of mean

reverting processes”

Frankena, L.H, (2016), “Pricing and hedging options in a negative interest rate
environment"

Huggins, D and Schaller, C, (2013), “Fixed Income Relative Value Analysis”

Hull, J, (2005), “Options, Futures and other Derivatives”

Jolliffe, I.T. (2002), “Principal Component Analysis”

Linderstrøm, M, (2013), “Fixed Income Derivatives Lecture Notes”

Longstaff, Schwartz and Santa Clara, (2001), “Relative Value of Caps and Swaptions:
Theory and Empirical Evidence”

Nordea – Financial Instruments book (2004)

Richard A. Johnson & Dean W. Wichern, (2013), “Applied Multivariate Statistical “

Salomon Smith Barney, (2000), “Principles of Principal Components: A fresh look at

Risk, Hedging, and Relative Value”.

TD Securities (2015), “Market Musings — Relative Value Across the U.S. Swap
Surface: A PCA Approach”

Trolle, Anders B. and Schwartz, Eduardo S., (2010), “An Empirical Analysis of the
Swaption Cube”

Tuckman, Bruce, (2011), “Fixed Income Securities”

78
Appendix

Derivation of solution for maximum likelihood

estimates
We are able to find 𝜇 by substituting the derived 𝜆 equation into the derived 𝜇
equation. By changing the notation for 𝜆 and 𝜇, this will be helpful for the derivation.
We have that
𝑛

𝑆𝑥 = ∑ 𝑆𝑡−1
𝑡=1
𝑛

𝑆𝑦 = ∑ 𝑆𝑡
𝑡=1
𝑛
2
𝑆𝑥𝑥 = ∑ 𝑆𝑡−1
𝑡=1
𝑛

𝑆𝑥𝑦 = ∑ 𝑆𝑡−1 𝑆𝑡
𝑡=1
𝑛

𝑆𝑦𝑦 = ∑ 𝑆𝑡2
𝑡=1
This leads us to

𝑆𝑦 − 𝑒 −𝜆𝛿 𝑆𝑥
𝜇=
𝑛(1 − 𝑒 −𝜆𝛿 )

1 𝑆𝑥𝑦 − 𝜇𝑆𝑥 − 𝜇𝑆𝑦 + 𝑛𝜇 2

𝜆 = − ln
𝛿 𝑆𝑥𝑥 − 2𝜇𝑆𝑥 + 𝑛𝜇 2

Here we substitute 𝜆 into 𝜇

𝑆𝑥𝑦 − 𝜇𝑆𝑥 − 𝜇𝑆𝑦 + 𝑛𝜇 2

𝑆𝑦 − ( ) 𝑆𝑥
𝑆𝑥𝑥 − 2𝜇𝑆𝑥 + 𝑛𝜇 2
𝑛𝜇 =
𝑆𝑥𝑦 − 𝜇𝑆𝑥 − 𝜇𝑆𝑦 + 𝑛𝜇 2
1−( )
𝑆𝑥𝑥 − 2𝜇𝑆𝑥 + 𝑛𝜇 2

Then removing the denominators

79
 𝑆𝑦 (𝑆𝑥𝑥 − 2𝜇𝑆𝑥 + 𝑛𝜇 2 ) − (𝑆𝑥𝑦 − 𝜇𝑆𝑥 − 𝜇𝑆𝑦 + 𝑛𝜇 2 )𝑆𝑥
𝑛𝜇 =
(𝑆𝑥𝑥 − 2𝜇𝑆𝑥 + 𝑛𝜇 2 ) − (𝑆𝑥𝑦 − 𝜇𝑆𝑥 − 𝜇𝑆𝑦 + 𝑛𝜇 2 )

Now we collect the terms

(𝑆𝑦 𝑆𝑥𝑥 − 𝑆𝑥 𝑆𝑥𝑦 ) − 𝜇(𝑆𝑥2 − 𝑆𝑥 𝑆𝑦 ) + 𝜇 2 𝑛(𝑆𝑦 − 𝑆𝑥 )

𝑛𝜇 =
(𝑆𝑥𝑥 − 𝑆𝑥𝑦 ) − 𝜇(𝑆𝑦 − 𝑆𝑥 )

By isolating 𝜇 on the left side of the equation

𝑛𝜇(𝑆𝑥𝑥 − 𝑆𝑥𝑦 ) − 𝜇(𝑆𝑥2 − 𝑆𝑥 𝑆𝑦 ) = 𝑆𝑦 𝑆𝑥𝑥 − 𝑆𝑥 𝑆𝑥𝑦

And finally we find a solution

𝑆𝑦 𝑆𝑥𝑥 − 𝑆𝑥 𝑆𝑦
𝜇=
𝑛(𝑆𝑥𝑥 − 𝑆𝑥𝑦 ) − (𝑆𝑥2 − 𝑆𝑥 𝑆𝑦 )

And for speed mean reversion rate

1 𝑆𝑥𝑦 − 𝜇𝑆𝑥 − 𝜇𝑆𝑦 + 𝑛𝜇 2

𝜆 = − ln
𝛿 𝑆𝑥𝑥 − 2𝜇𝑆𝑥 + 𝑛𝜇 2

Lastly we get the variance as

2𝜆
σ2 = σ
̂2
1 − 𝛼2
𝛼 = 𝑒 −𝜆𝛿

Where

1
σ2 =
̂ [𝑆𝑦𝑦 − 2𝛼𝑆𝑥𝑦 + 𝛼 2 𝑆𝑥𝑥 − 2𝜇(1 − 𝛼)(𝑆𝑦 − 𝛼𝑆𝑥 ) + 𝑛𝜇 2 (1 − 𝛼)2 ]
𝑛

80
Trading signals for USD and EUR
Date Last Point Mean Lambda Sigma2 Exphalflife Exp90pct
10/14/2015 24.362 -3.793 0.037 11.585 18.707 62.144
10/7/2015 23.236 -4.602 0.037 12.393 18.748 62.281
10/19/2015 23.964 -3.420 0.037 10.969 18.887 62.741
10/16/2015 24.277 -3.530 0.036 11.174 19.275 64.030
10/6/2015 24.365 -4.760 0.036 12.551 19.400 64.444
10/13/2015 22.523 -4.000 0.035 11.757 19.532 64.886
10/15/2015 24.080 -3.661 0.035 11.364 19.632 65.216
10/9/2015 22.060 -4.325 0.035 12.049 19.644 65.257
10/8/2015 22.455 -4.477 0.035 12.194 19.694 65.420
10/12/2015 22.244 -4.163 0.035 11.899 19.837 65.896
12/15/2016 9.882 1.769 0.035 8.595 20.091 66.742
12/21/2016 4.649 1.326 0.034 8.565 20.139 66.899
12/20/2016 6.835 1.483 0.034 8.565 20.152 66.944
10/20/2015 23.251 -3.312 0.034 10.693 20.171 67.007
12/23/2016 2.615 1.145 0.034 8.554 20.188 67.062
12/13/2016 14.166 2.094 0.034 8.592 20.295 67.420

Figure 18 - Mean reversion on daily PCA for USD PC 2 scores. Sorted by - speed of mean
reversion (Lambda).

Date Last Point Mean Lambda Sigma2 Exphalflife Exp90pct

11/16/2015 -0.942 1.163 0.051 5.957 13.609 45.207
11/13/2015 -0.543 1.242 0.048 5.750 14.312 47.544
11/12/2015 -0.238 1.340 0.047 5.533 14.858 49.357
4/22/2016 6.259 -0.862 0.047 2.794 14.871 49.399
4/21/2016 -6.188 0.892 0.045 2.794 15.340 50.959
11/17/2015 -1.351 0.877 0.044 5.882 15.845 52.636
4/20/2016 -5.668 0.927 0.043 2.787 16.035 53.266
4/19/2016 -6.658 0.942 0.043 2.782 16.218 53.874
4/18/2016 -7.125 0.960 0.042 2.776 16.692 55.448
4/15/2016 -7.242 1.001 0.041 2.772 16.866 56.027
11/18/2015 -2.074 0.585 0.040 5.850 17.175 57.056
11/11/2015 -0.430 1.185 0.040 5.516 17.433 57.912
11/10/2015 -0.244 1.086 0.036 5.440 19.025 63.199
2/8/2016 -0.728 -1.331 0.036 3.695 19.117 63.507
2/3/2016 4.055 1.580 0.036 3.732 19.299 64.109
11/19/2015 -2.876 0.005 0.036 5.567 19.323 64.190

Figure 19 - Mean reversion on daily PCA for USD PC 3 scores. Sorted by - speed of mean
reversion (Lambda).

81
 Date Last Point Mean Lambda Sigma2 Exphalflife Exp90pct
1.10.2017 2.546 -4.585 0.017 1.658 41.975 139.437
12.28.2016 0.515 -5.488 0.016 1.708 42.482 141.121
1.4.2017 1.302 -5.027 0.016 1.680 42.842 142.317
1.9.2017 1.759 -4.768 0.016 1.663 43.223 143.585
12.26.2016 -0.425 -5.797 0.016 1.719 43.502 144.510
12.27.2016 -0.334 -5.704 0.016 1.713 43.584 144.782
1.6.2017 1.660 -4.866 0.016 1.667 43.662 145.041
1.3.2017 1.198 -5.149 0.016 1.685 43.811 145.537
12.23.2016 -0.532 -5.918 0.016 1.726 43.992 146.140
12.29.2016 0.767 -5.445 0.016 1.700 44.185 146.778
1.5.2017 1.729 -4.960 0.016 1.672 44.470 147.725
1.2.2017 1.164 -5.268 0.015 1.691 44.927 149.244
12.30.2016 0.992 -5.367 0.015 1.695 44.932 149.261
1.11.2017 2.766 -4.578 0.015 1.640 45.289 150.445
12.22.2016 -0.864 -6.110 0.015 1.735 45.578 151.407
12.21.2016 -1.030 -6.260 0.015 1.742 46.564 154.682

Figure 20 - Mean reversion on daily PCA for EUR PC 2 scores. Sorted by - speed of mean reversion
(Lambda).

Date Last Point Mean Lambda Sigma2 Exphalflife Exp90pct

3.10.2015 12.560 -0.389 0.043 5.988 16.216 53.868
3.9.2015 11.788 -0.466 0.043 5.901 16.259 54.010
3.5.2015 7.003 -0.711 0.042 5.699 16.405 54.497
3.6.2015 9.848 -0.548 0.042 5.802 16.465 54.695
3.4.2015 6.870 -0.768 0.042 5.610 16.592 55.118
3.11.2015 13.672 -0.227 0.042 6.057 16.629 55.239
6.28.2013 -7.135 0.165 0.041 5.191 16.865 56.025
6.27.2013 -8.304 0.144 0.041 5.185 16.937 56.265
6.26.2013 -7.319 0.096 0.041 5.188 17.106 56.824
3.26.2013 0.200 0.912 0.040 4.866 17.119 56.868
2.27.2015 4.651 -0.993 0.040 5.360 17.147 56.962
3.16.2015 17.178 0.176 0.040 6.455 17.148 56.963
3.22.2013 -1.747 -1.073 0.040 4.804 17.152 56.976
3.3.2015 5.562 -0.811 0.040 5.517 17.153 56.982
6.25.2013 8.200 -0.078 0.040 5.183 17.172 57.044
7.1.2013 -8.278 0.010 0.040 5.180 17.258 57.331

Figure 21- Mean reversion on daily PCA for EUR PC 3 scores. Sorted by - speed of mean reversion
(Lambda).

82
Selected R code

Code: MLE of OU process for third principal component

for (j in 630:dim(usdnew)[1]){
PC<-prcomp(usdnew[(j-629):j,2:dim(usdnew)[2]],scale.=FALSE)

p<-vector(mode="numeric",length=nrow(PC$x))
d<-vector(mode="numeric",length=(nrow(PC$x)-1))
for (i in 1:nrow(PC$x)){
p[i]<-
PC$x[i,3]*PC$x[i,3]
for (l in 2:nrow(PC$x)){
d[l]<-PC$x[l-1,3]*PC$x[l,3]
}
}

n<-nrow(PC$x)
Sx<-sum(PC$x[,3])-PC$x[nrow(PC$x),3]#last point
Sy<-sum(PC$x[,3])-PC$x[1,3]#first point
Sxx<-sum(p[1:nrow(PC$x)])-p[i]
Syy<-sum(p[1:nrow(PC$x)])-p[1]
Sxy<-sum(d[])

lastpoint[j]<-PC$x[nrow(PC$x),3]
middel[j]<-(Sy*Sxx-Sx*Sxy)/(n*(Sxx-Sxy)-(Sx*Sx-Sx*Sy))
deviation[j]<-(lastpoint[j]-middel[j])/middel[j]*10000
lambda[j]<-(-log((Sxy-middel[j]*Sx-middel[j]*Sy+n*middel[j]*middel[j])/(Sxx-
2*middel[j]*Sx+n*middel[j]*middel[j])))
alpha[j]<-exp(lambda[j]*-1)
sigmaquer[j]<-(Syy-2*alpha[j]*Sxy+alpha[j]*alpha[j]*Sxx-2*middel[j]*(1-
alpha[j])*(Sy-alpha[j]*Sx)+n*middel[j]*middel[j]*(1-alpha[j])*(1-alpha[j]))/n
sigma2[j]<-sigmaquer[j]*2*lambda[j]/(1-alpha[j]*alpha[j])
sigma[j]<-sqrt(sigma2[j])
exphalflife[j]<-log(2)/lambda[j]
exp90pct[j]<-log(10)/lambda[j]
}

83
Eigenvectors on vega sector for EUR

Figure 22 – First eigenvector of vega sector, EUR surface, Date: 2015-07-03

Figure 23 – Second eigenvector of vega sector, EUR surface, Date: 2015-07-03

84
Figure 24 – Third eigenvector of vega sector, EUR surface, Date: 2015-07-03

85