Fixed Income Analytics

Renato Antonio Palma Valencia

April 2026

1
Renato Palma Valencia Fixed Income Analytics

Contents
1 Module 1: Bond Pricing 3
1.1 Introduction and Class Hierarchy . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 3
1.2 Definition and Fair Price . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 4
1.3 Generalization and Examples . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 5
1.4 Yield to Maturity (YTM) . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 6
1.5 Bond Parity . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 7
1.6 Duration & DV01 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 8
1.7 Convexity and Taylor Approximation . . . . . . . . . . . . . . . . . . . . . . . . . . . 10
1.8 Conclusions and Key Takeaways . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 10

2 Module 2: Yield Curve 12

2.1 Introduction and Class Hierarchy . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 12
2.2 Term Structure of Interest Rates and Discount Factor . . . . . . . . . . . . . . . . . 13
2.3 Forward Rates and Instantaneous Forward Rates . . . . . . . . . . . . . . . . . . . . 14
2.4 Bootstrapping . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 15
2.5 Curve Interpolation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 15
2.6 Conclusions and Key Takeaways . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 18

3 Module 3: Sensitivity Analysis 20

3.1 Introduction and Class Hierarchy . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 20
3.2 Parallel Shift DV01 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 21
3.3 Key Rate Durations . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 22
3.4 Portfolio Sensitivity . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 24
3.5 Conclusions and Key Takeaways . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 25

4 Module 4: PCA on Term Structure 27

4.1 Introduction and Class Hierarchy . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 27
4.2 Motivation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 28
4.3 PCA Framework . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 29
4.4 The Three Factors (Level, Slope and Curvature) . . . . . . . . . . . . . . . . . . . . 30
4.5 PCA Reconstruction and Risk Management . . . . . . . . . . . . . . . . . . . . . . . 32
4.6 Conclusions and Key Takeaways . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 34

5 Module 5: IRS Pricing 36

5.1 Introduction and Class Hierarchy . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 36
5.2 Interest Rate Swap Definition and Structure . . . . . . . . . . . . . . . . . . . . . . . 37
5.3 The Floating Leg equals Par . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 38
5.4 Swap as a Portfolio of FRAs . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 38
5.5 Swap Rate K* . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 39
5.6 Swap Value at t . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 40
5.7 Swap DV01 and Applications . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 40
5.8 Conclusions and Key Takeaways . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 42

6 Project Summary and Further Research 43

2
Renato Palma Valencia Fixed Income Analytics

Abstract
This document develops the mathematical foundations of Fixed Income Analytics imple-
mented in the companion Python library fixed-income-analytics. Each module corresponds
directly to a Python module in /src.

1 Module 1: Bond Pricing

Python Implementation

Module: src/[Link]
Class: VanillaBond
Notebook: notebooks/01 bond [Link]

1.1 Introduction and Class Hierarchy

This module implements the mathematical foundations of vanilla bond pricing. Starting from
the definition of a bond as a sequence of deterministic cash flows, we derive closed-form pricing
formulas under discrete, generalized and continuous compounding, introduce the yield-to-maturity
as the internal rate of return of the instrument, and develop the first and second-order risk metrics
that form the language of fixed income trading.

src/[Link]
|- VanillaBond
|-- Pricing
| |-- cash_flows()
| |-- bond_price(ytm)
| |-- bond_price_continuous(r)
|-- Yield
| |-- ytm(market_price)
|-- Risk Metrics
| |-- risk_metrics(ytm)
| |-- Macaulay Duration
| |-- Modified Duration
| |-- Convexity
| |-- DV01
|-- Analysis
| |-- price_change(ytm, dy)
| |-- parity(ytm)
|-- Validation
|-- validate_roundtrip(r)

In [Link] the entire module is encapsulated in a single class VanillaBond . A bond is

instantiated with its economic parameters and all pricing, risk and analysis methods are exposed
as instance methods, allowing clean and readable usage:

3
 Renato Palma Valencia Fixed Income Analytics

1 bond = VanillaBond ( coupon_rate = 0.05 ,

2 face_value = 1000 ,
3 maturity = 10 ,
4 frequency = 1)
5 # Pricing
6 P = bond . bond_price ( ytm = 0.04)
7
8 # Risk Metrics
9 metrics = bond . risk_metrics ( ytm = 0.04)
10
11 # Taylor Approximation
12 approx = bond . price_change ( ytm = 0.04 , dy = 0.01)

1.2 Definition and Fair Price

Definition 1.1 (Vanilla Bond). A bond is a fixed income instrument that represents a debt obli-
gation. The issuer commits to pay:
1. Coupons C at periodic dates t1 , t2 , . . . , tn .

2. A principal (Face Value) F at maturity date tn .

The cash flow at each period is formally defined as:
(
C i = 1, 2, . . . , n − 1
CFi =
C +F i=n

Proposition 1.1 (Bond Fair Price). The fair price P of a bond is given by:

C(1 − (1 + r)−n )
P = + F (1 + r)−n
r
with r the constant periodic discount rate and n the number of periods.

Proof. Notice that the price of a bond equals to the present value of all its cash flows:
n n  i
X C F F X 1
P = i
+ n
= n
+C
i=1
(1 + r) (1 + r) (1 + r) i=1
1+r

Pn 1−q n+1
Using the geometric sum formula i=1 qi = 1−q with q = 1/(1 + r):

 n+1 
1
F 1 − 1+r

(1 + r) − (1 + r)−n − r

−n
P = +C − 1 = F (1 + r) + C
 
(1 + r)n 1
1 − 1+r r

1 − (1 + r)−n
 
P =C + F (1 + r)−n
r

4
Renato Palma Valencia Fixed Income Analytics

1.3 Generalization and Examples

r nT


If the periodic discount rate r compounds T times per period, the balance factor is: 1 + T and
the price generalizes to:
Proposition 1.2. The fair price P of a bond is:

1 − (1 + Tr )−nT
   r −nT
P =C· r T +F 1+
(1 + T ) − 1 T

with r the periodic discount rate, n the maturity, T the number of compounds of the discount rate
per period.
Proof.
n n
X C F F X 1
P = r iT + r nT = r nT + C
i=1
(1 + T ) (1 + T ) (1 + T ) i=1
(1 + Tr )iT
  (n+1)T 
n  T !i 1 − 1
F X 1 F 1+ Tr
= +C = r nT + C  − 1
 
(1 + Tr )nT r T
1 + (1 + )

1
i=1 T T 1 − 1+ r
T

r −nT
"
r T
# "
−nT #
1 − 1 + Tr



F 1+ T − 1+ T F
= +C −1 = +C
(1 + Tr )nT (1 + Tr )nT

T
T
1+ r T −1 1+ r
T −1

The general formula unifies several classical instruments as particular cases:

1. Zero-Coupon Bond (C = 0): When the bond pays no coupons, the price reduces to:
 r −nT
P =F 1+
T
The price is simply the present value of the face value. This is the purest form of a fixed
income instrument and plays a central role in bootstrapping, since each zero-coupon bond
defines exactly one point on the discount curve. Note also that the Macaulay Duration of
a zero coupon bond equals exactly its maturity: Dmac = n, which is the maximum possible
duration for a bond of that maturity.
2. Continuous Compounding (T → ∞): Taking T → ∞ and using that limT →∞ (1+ Tr )nT =
ern , we get the following price of the bond:
1 − e−rn
 
P =C· + F e−rn
er − 1
this is the standard continuous compounding formula used throughout quantitative finance
and is the natural limit of the general formula. Consider the case of a bond with a coupon
rate C/F = 5%, a face value F = $1000, a maturity n = 10 and a discount rate r = 0.04.
Notice in Figure 1 how the price converges as there are more compounds T per year, this is,
as T → ∞:

5
Renato Palma Valencia Fixed Income Analytics

Figure 1: Convergence of Bond Price as T → ∞

3. Consol Bond (n → ∞): A consol (or perpetuity) pays coupons forever and never returns
the principal. Taking n → ∞ in the price formula and noting that (1 + r)−nT → 0:

C
P =
r T


1+ T −1

The price is simply the coupon divided by the discount balance factor. This result has a
beautiful intuition: A perpetuity is worth exactly the amount you would need to invest at

T
rate 1 + Tr − 1 to generate coupon C forever. Note that the British government issued
consols (Consolidated Annuities) that traded in market for over 200 years.

T
4. Par Bond (P = F ): When the discount balance factor is 1 + Tr −1 = C
F , the price of
the bond is:
1 − (1 + Tr )−nT
   r −nT  r −nT  r −nT
P =C· +F 1+ =F −F 1+ +F 1+ =F
C/F T T T

This bond trades at par. This case connects directly to the YTM parity result in section 1.4

1.4 Yield to Maturity (YTM)

Definition 1.2 (Yield to Maturity). The YTM is the rate y that equates the market price of the
bond with the present value of its cash flows:
n
X CFi
P =
i=1
(1 + Ty )iT

This equation has no closed-form solution and must be solved numerically.

6
Renato Palma Valencia Fixed Income Analytics

Definition 1.3 (Newton Raphson Method). Define the objective function:

n
X CFi
f (y) = −P =0
i=1
(1 + Ty )iT

and its derivative, used for the Newton Step:

n
′
X i · CFi
f (y) = −
i=1
(1 + Ty )iT +1

The Newton-Raphson iteration updates the estimate with the following recursion:
f (yk )
yk+1 = yk −
f ′ (yk )
starting from an initial guess y0 .
The algorithm converges quadratically to the true YTM. Convergence is guaranteed because f (y)
is strictly convex and monotonically decreasing in y, which ensures a unique root. In [Link] this
is implemented via [Link] with analytical derivative for maximum efficiency.
The key price-yield relationship is:
↑ y ⇒↓ P & ↓ y ⇒↑ P

1.5 Bond Parity

Proposition 1.3. Let c = C/F be the coupon rate. Then:
 y T
P ≶ F ⇐⇒ 1 + −1≷c
T

T
Proof. Substituting 1 + Ty − 1 = c and C = cF :

1 − (1 + Ty )−nT
   y −nT  y −nT  y −nT
P = cF · +F 1+ =F −F 1+ +F 1+ =F
c T T T

T
Since P is strictly decreasing in y (f ′ (y) < 0) and 1 + Ty − 1 > 0 then the result follows
directly.
When there is one capitalization per year (T = 1) the condition collapses to:
P ≶ F ⇐⇒ y ≷ c
The economic intuition for this case is clear: when the market demands a return y > c, the bond’s
coupons are insufficient, so the market discounts the price below par to compensate with a capital
gain at maturity. The opposite holds when y < c. To see this graphically, consider the case of a
bond with coupon rate c = C/F = 5%, a face value of $1000, a maturity n = 10 and a discrete
capitalization T = 1. The Figure 2 plots the price of the bond P against the yield y and shows the
areas where the bond is at premium or discount.
Regardless of whether a bond starts at premium or discount, its price will converge to its face
value F as it approaches maturity (t → n), a phenomenon known as Pull to Par.

7
Renato Palma Valencia Fixed Income Analytics

Figure 2: Price-Yield Relationship

Condition Name Intuition

P = F ⇐⇒ y = c Bond at Par Market rate equals coupon rate
P > F ⇐⇒ y < c Bond above Par (Premium Bond) Generous coupons, price rises
P < F ⇐⇒ y > c Bond below Par (Discount Bond) Insufficient coupons, price drops

Table 1: Bonds Parity

1.6 Duration & DV01

Definition 1.4 (Macaulay Duration). Macaulay Duration is the weighted average time of cash
flows of the bond, where weights are present values:
n
1 X CFi
Dmac = i·
P i=1 (1 + Ty )iT

where CFi is the cash flow at period i and T the number of coupons periods per year. This is the
average time for the investor to recover the investment.
As we said in section 1.3, the zero-coupon bond has a Macaulay Duration Dmac = n, now we

−nT
can prove it by applying the definition and recalling that Pzero = F 1 + Ty :
 y nT  y −nT
Dmac = F −1 1 + · nF 1 + =n
T T
Definition 1.5 (Modified Duration). It measures the price sensitivity to yield changes:
Dmac
Dmod =
1 + Ty
We can plot these durations against maturity to see how they behave. Consider a bond with
four different coupon rates c = C/F = (0%, 2%, 5%, 10%), face value F = 1000 and yield y = 5%

8
Renato Palma Valencia Fixed Income Analytics

that compounds once per year T = 1. Figure 3 shows durations against maturity and how as
coupon rates increase, duration gets lower.

Figure 3: Durations versus Maturity

Proposition 1.4. Price sensitivity can be written as:

dP
= −Dmod · P
dy
Proof. Remember that:
n n
X C F X CFi
P = y iT + y nT = y iT
i=1
(1 + T ) (1 + T ) i=1
(1 + T)
n n
dP X i · CFi 1 X CFi Dmac
=− y iT +1 = − y i· y iT = −P · = −Dmod · P
dy i=1
(1 + T ) 1+ T i=1 (1 + T ) 1 + Ty

In continuous compounding (T → ∞):

 y 1 dP
Dmac = lim Dmod 1 + = Dmod = −
T →∞ T P dy

Definition 1.6 (Dollar Value of a Basis Point). DV01 is how much the bond price changes when
the yield changes 1 basis point (0.01%):

dP
DV01 = − · 0.0001 = Dmod · P · 0.0001
dy

9
Renato Palma Valencia Fixed Income Analytics

1.7 Convexity and Taylor Approximation

As duration is a first order linear approximation, convexity captures the second order effect:
Definition 1.7 (Bond Convexity). Bond convexity measures the non-linear relationship between
price and yield:
n
1 d2 P 1 X i(iT + 1) CFi
C= = ·
P dy 2 P i=1 T (1 + Ty )iT +2

Combining Duration and Convexity gives the second-order Taylor approximation of price change:
dP 1 d2 P
P (y) = P (y0 ) + (y − y0 ) + (y − y0 )2 + · · ·
dy 2 dy 2
1
P (y) − P (y0 ) ≈ −Dmod · P (y − y0 ) + P · C(y − y0 )2
 2 
1 2
∆P ≈ P −Dmod · ∆y + C · (∆y)
2
Convexity is always positive for vanilla bonds, meaning the price-yield curve is convex: gains from
falling yields exceed losses from equal rising yields. This is a desirable property for the bond-holder
and explains why higher convexity bonds command a premium in the market.
We can visualize the second-order Taylor approximation and the gains of adding the second-order
term (convexity) instead of just approximating with the first order term for a bond with the same
characteristics as the examples above.

Figure 4: First and Second-Order Approximation of a Bond Price

At changes of the yield of 100 bp, the approximation error with the first-order term only is
$3.4751, adding the second-order term we get an absolute error of $0.1256.

1.8 Conclusions and Key Takeaways

• The bond price is a strictly convex, decreasing function of the yield: Duration captures the
slope and Convexity captures the curvature of this relationship.

10
Renato Palma Valencia Fixed Income Analytics

• The YTM has no closed-form solution but converges quadratically under Newton-Raphson,
since the objective function is strictly convex and monotone.
• Convexity is always positive for vanilla bonds: the second-order Taylor approximation mate-
rially improves the price estimate for yield moves above 50 bps.

The bond pricing framework of this module rests on a critical simplification: the discount rate r
is assumed constant across all maturities. In practice, this is never true. A 1-year investment and a
10-year investment command different rates, and the relationship between rates and maturities (the
term structure of interest rates) is one of the most important objects in fixed income. Formally,
replacing the constant r with a maturity dependent function r(t) transforms the bond price from:
n
X CFi
P =
i=1
(1 + Tr )−iT

into:
n n
X CFi In Continuous Capitalization X
P = r(ti ) −ti T
⇒ P = CFi e−r(ti )·ti
i=1 (1 + T ) i=1

where each cash flow is discounted at its own maturity-specific rate. This requires building the entire
function r(t) from observed market prices, the problem of yield curve construction, which is the
subject of module 2.

11
Renato Palma Valencia Fixed Income Analytics

2 Module 2: Yield Curve

Python Implementation

Module: src/[Link]
Classes: DiscountCurve , Bootstrapper and CurveInterpolator
Notebook: notebooks/02 yield [Link]

2.1 Introduction and Class Hierarchy

This module extends the single-rate bond pricing of Module 1 to the full term structure setting.
Rather than discounting all cash flows at a constant rate r, we build a continuous function r(t) from
observable market prices, interpolate it across arbitrary maturities, and fit a parametric model that
captures its economic structure. The three classes in src/[Link] handle these three tasks in
sequence.

src/[Link]
|
|-- DiscountCurve
| |
| |-- Attributes
| | |-- maturities Observed tenors
| | |-- spot_rates Continuously compounded
| |
| |-- Core
| | |-- discount_factor(t) Z(t) = e^(-r(t)*t)
| | |-- discount_factors() Z vector
| | |-- spot_rate(t) Linear interp
| |
| |-- Rates
| |-- forward_rate(t1, t2) f(t1,t2)
| |-- instantaneous_forward(t) f(t) numerical
| |-- forward_curve(t_grid) f(t) over grid
|
|-- Bootstrapper
| |-- bootstrap(prices, cash_flows, Sequential
| | maturities) extraction
| |-- from_yields(maturities, yields) FRED/market data
|
|-- CurveInterpolator
|-- linear(t) Linear on zero rates
|-- cubic_spline(t) Smooth, C2 continuous
|-- cubic_spline_forward(t) f(t) via spline deriv
|-- nss_spot(t, beta) NSS spot formula
|-- nss_forward(t, beta) NSS forward formula
|-- fit_nss() L-BFGS-B calibration

12
 Renato Palma Valencia Fixed Income Analytics

|-- nss_summary(beta) Economic interpretation

|-- compare(t_grid, beta) All three methods

The three classes are designed to be used in sequence. DiscountCurve is the fundamental data
structure, holding observed spot rates and computing discount factors. Bootstrapper constructs a
DiscountCurve from raw market data. CurveInterpolator takes a DiscountCurve and extends
it to arbitrary maturities via three methods of increasing sophistication:
1 # Step 1: Build curve from FRED yields
2 curve = Bootstrapper . from_yields ( maturities , yields )
3
4 # Step 2: Interpolate at arbitrary maturities
5 interp = C u r v e I n t e r p o l a t o r ( curve )
6
7 # Step 3: Fit NSS parametric model
8 beta = interp . fit_nss ()
9 r_nss = C u r v e I n t e r p o l a t o r . nss_spot ( t_grid , beta )

2.2 Term Structure of Interest Rates and Discount Factor

In the bond pricing of Module 1 we assumed a constant discount rate r. In reality, rates vary with
maturity: a 1-year investment typically commands a different rate than a 10-year one.
Definition 2.1 (Term Structure of Interest Rates). The term structure of interest rates is a function
r(t) that assigns a spot rate to each maturity t.
Assuming that r(t) is continuously differentiable, the bond price generalizes to:
n
X
P = CFi · e−r(ti )ti
i=1

where each cash flow is discounted at its own maturity-specific rate. This is the foundation of all
modern fixed income pricing.
Definition 2.2 (Discount Factor). The continuous discount factor for maturity t is:

Z(t) = e−r(t)·t

It represents the present value today of $1 to be received at time t. The bond price can be
written compactly as:
n
X
P = CFi · Z(ti )
i=1

Discount factors have three fundamental properties. First Z(0) = 1: a dollar today is worth a
dollar. Second, Z(t) ∈ (0, 1), ∀t > 0 and positive rates: future cash flows are always worth less than
their face value today. Third, Z(t) is decreasing in t for positive rates: the further in the future,
the less a dollar is worth today.
The spot rate can always be recovered from the discount factor:
ln Z(t)
r(t) = −
t

13
Renato Palma Valencia Fixed Income Analytics

2.3 Forward Rates and Instantaneous Forward Rates

Definition 2.3 (Forward Rate). The forward rate f (t1 , t2 ) is the implicit rate agreed today for an
investment between two future dates t1 and t2 . It is defined by the no-arbitrage condition:

er(t2 )·t2 = er(t1 )·t1 · ef (t1 ,t2 )·(t2 −t1 )

Applying logarithm at both sides:

r(t2 )t2 = r(t1 )t1 + f (t1 , t2 )(t2 − t1 )

r(t2 )t2 − r(t1 )t1
f (t1 , t2 ) =
t2 − t 1
The economic intuition is clear: investing from 0 to t2 at spot rate r(t2 ) must yield the same as
investing from 0 to t1 at r(t1 ) and then rolling over from t1 to t2 at the forward rate f (t1 , t2 ).
Otherwise there would be an arbitrage opportunity.
In terms of the discount factors, the forward rate has an equivalent and elegant expression:

ln Z(t1 ) − ln Z(t2 )
f (t1 , t2 ) =
t2 − t1
Proposition 2.1. The instantaneous forward rate is given by:

∂r(t)
f (t) = r(t) + t ·
∂t
Proof. Setting t1 = t and taking the limit t2 → t:

r(t2 ) · t2 − r(t) · t d ∂r(t)

f (t) = lim f (t, t2 ) = lim = [r(t) · t] = r(t) + t ·
t2 →t t2 →t t2 − t dt ∂t

The two terms have a clear interpretation. The first term r(t) is simply the spot rate at maturity
t. The second term t · ∂r(t)
∂t is a correction for the slope of the yield curve: if the curve is upward
sloping, the instantaneous forward rate lies above the spot rate, and vice versa. When the yield
curve is flat ( ∂r(t)
∂t = 0), the instantaneous forward rate equals the spot rate exactly.
The inverse relationship allows recovering the spot rate from the forward curve:
Z t Z t
1
t · r(t) = f (s)ds ⇒ r(t) = f (s)ds
0 t 0

The spot rate is the average of instantaneous forward rates up to maturity t. This is one of the
most interesting results in fixed income: the entire spot curve is encoded in the forward curve and
vice versa.

14
Renato Palma Valencia Fixed Income Analytics

2.4 Bootstrapping
The yield curve is not directly observable. What we observe in the market are prices of traded bonds
at specific maturities. Bootstrapping is the iterative procedure that extracts the unobservable spot
rates from those observable prices.

Definition 2.4 (Bootstrapping). Bootstrapping is the process of extracting spot rates r(t1 ), . . . , r(tn )
sequentially from observed bond prices, using already extracted rates to solve for the next one.
Proposition 2.2. Given the observed bond price P , cash flows CFi and already extracted discount
factors Z(t1 ), . . . , Z(tn−1 ), the spot rate at maturity n is:
!
1 CFn
r(n) = · ln Pn−1
n P − i=1 CFi · Z(i)

Proof. Starting from the bond price equation under continuous compounding:
n
X n−1
X
P = CFi · e−ir(i) = CFi · Z(i) + CFn · Z(n)
i=1 i=1

Isolating Z(n) and using that Z(n) = e−r(n)·n :

Pn−1
P− CFi · Z(i)
Z(n) = i=1
= e−r(n)·n
CFn
Taking the logarithm and solving for r(n):
!
1 CFn
r(n) = · ln Pn−1
n P− i=1 CFi · Z(i)

The procedure is sequential: r(t1 ) is extracted first (typically from a short-term instrument
like a T-Bill), then r(t2 ) using the already known Z(t1 ) and so on. Each step uses all previously
extracted rates, which is why the method is called bootstrapping: each rate ”pulls up” the next
one.

2.5 Curve Interpolation

Bootstrapping gives us spot rates only at the maturities of traded bonds, typically {0.25, 0.5, 1, 2, 3, 5,
7, 10, 20, 30} years for the U.S. Treasuries. To price instruments at arbitrary maturities, we need to
interpolate between these points.

Definition 2.5 (Linear Interpolation on Zero Rates). For t ∈ [ti , ti+1 ], the linearly interpolated
spot rate is:

r(ti+1 ) − r(ti )
r(t) = r(ti ) + (t − ti )
ti+1 − ti

15
Renato Palma Valencia Fixed Income Analytics

This is the simplest method: the spot rate moves linearly between observed nodes. It is easy to
implement and understand, but has an important drawback: the implied forward rates are piecewise
constant and exhibit discontinuous jumps at each node. In practice this means the forward curve
has ”kinks” at every tenor point, which is financially unrealistic.
Definition 2.6 (Cubic Spline Interpolation). A cubic spline fits a piecewise cubic polynomial S(t)
such that on each interval [ti , ti+1 ]:

S(t) = ai + bi (t − ti ) + ci (t − ti )2 + di (t − ti )3

with the conditions that S(t), S ′ (t) and S ′′ (t) are all continuous at every interior node. This gives a
system of 4(n − 1) unknowns with 4(n − 1) equations (matching values, first and second derivatives
at each node, plus two boundary conditions).
The key advantage over linear interpolation is that the cubic spline produces a smooth forward
curve. Since f (t) = r(t) + t · ∂r(t) ∂r(t) ∂r(t)
∂t , the forward rate depends on ∂t : a discontinuous ∂t from
linear interpolation generates a discontinuous forward curve, while the cubic spline guarantees ∂r(t)
∂t
is continuous everywhere.
The two standard boundary conditions used in practice are:
1. Natural Spline: S ′′ (t1 ) = S ′′ (tn ) = 0, meaning no curvature at the endpoints. Simple and
the most common choice.
2. Not-a-Knot: forces S ′′′ (t) to be continuous at the second and second-to-last nodes. Gener-
ally produces a better behavior at the boundaries and is the default in [Link]
.CubicSpline.
Proposition 2.3. Linear interpolation on zero rates implies piecewise constant forward rates with
jumps at nodes. Cubic spline interpolation implies continuous forward rates everywhere.
Proof. Recall the instantaneous forward rate formula f (t) = r(t)+t·r′ (t). Under linear interpolation
we have that:
r(ti+1 ) − r(ti )
r′ (t) =
ti+1 − ti
is constant on each interval but jumps at nodes, so f (t) also jumps. Under cubic spline, r′ (t) is
continuous by construction, so f (t) is continuous everywhere.
In [Link] both methods are implemented and compared. The cubic spline is the default for
all downstream calculations.
Both linear interpolation and cubic splines are non-parametric methods: they fit the observed
data without imposing any financial structure on the curve. The Nelson-Siegel-Svensson model takes
a fundamentally different approach: it parametrizes the entire yield curve with a small number of
economically interpretable parameters.
Definition 2.7 (Nelson-Siegel-Svensson Model). The model starts with the instantaneous forward
rate, which is specified as:
t −t/λ1 t
f (t) = β0 + β1 e−t/λ1 + β2 e + β3 e−t/λ2
λ1 λ2

16
Renato Palma Valencia Fixed Income Analytics

1
Rt
Integrating r(t) = t 0
f (s)ds (the result from section 2.3), the spot rate curve is:

1 − e−t/λ1 1 − e−t/λ1 1 − e−t/λ2

   
r(t) = β0 + β1 + β2 − e−t/λ1 + β3 − e−t/λ2
t/λ1 t/λ1 t/λ2

with parameters β = (β0 , β1 , β2 , β3 , λ1 , λ2 ).

The first three terms constitute the original Nelson and Siegel (1987) model; the fourth term
is the Svensson (1994) extension that adds a second hump to accommodate more complex curve
shapes.
The parameters have direct economic interpretations. Taking limits:
lim r(t) = β0 lim r(t) = β0 + β1
t→∞ t→0

so β0 is the long-run level of interest rates and β0 + β1 is the short-run level. The parameters β2
and β3 control the magnitude of the humps at intermediate maturities, with λ1 and λ2 controlling
where those humps are located. For a normal upward-sloping curve, β1 < 0 (short rates below long
rates) and β2 > 0 (a positive hump in the medium term).
The parameters are estimated by minimizing the sum of squared errors between model rates
and observed rates:
Xn
β̂ = arg min (robs (ti ) − r(ti ; β))2
β
i=1

subject to β0 > 0 and λ1 , λ2 > 0. This is a nonlinear least squares problem solved numerically via
L-BFGS-B optimization in [Link] .
In the notebook, we use synthetic data of spot rates to compare the three interpolation methods and
its respective forward curves. In Figure 5 we can see the smoothness, discontinuity and flattening
of forward rates mentioned above.

Figure 5: Three Interpolation Methods

Taking 4 key-dates (Pre-COVID, COVID shock, Fed Hikes Peak Dec-2022, and Jan-2025), we
see plotted the NSS yield curve for spot and forward rates. Figure 6 shows how sloping spot curve
implies that the market expects higher rates in the future.

17
Renato Palma Valencia Fixed Income Analytics

Figure 6: NSS Historical Comparison

The three Nelson-Siegel factors (level β0 , slope β1 and curvature β2 ) are not arbitrary. Litterman
et al. (1991) showed empirically that the three first components of the yield curve changes explains
over 99% of total variance and correspond exactly to parallel shifts (level), tilts (slope) and humps
(curvature). This connection between the parametric NSS model and the statistical PCA of module
4 is one of the deepest results in fixed income and provides a powerful cross-validation: a well-
calibrated NSS model should produce factors that align closely with the PCA components.

2.6 Conclusions and Key Takeaways

• The discount factor Z(t) = e−r(t)·t is the fundamental object of the yield curve: all pricing,
forward rates and bootstrapping results follow from it.

• The instantaneous forward rate f (t) = r(t) + t · r′ (t) encodes the slope of the spot curve and
is continuous only when r′ (t) is continuous, a property guaranteed by the cubic spline but
not by linear interpolation.
• NSS imposes an economic structure on the curve with six interpretable parameters: long-run
level β0 , short-run adjustment β1 , two curvature components β2 , β3 and two decay parameters
λ1 , λ2 . It is the standard model of central banks worldwide precisely because its factors
correspond to the three dominant modes of yield curve variation.

The yield curve r(t) and its discount factors Z(ti ) are now complete building blocks. With them
we can price any fixed-rate instrument by discounting its cash flows at maturity-specific rates rather
than a single constant. However, having a precise price is only the first step. The next question
is how sensitive that price is to changes in the curve: if rates rise by 5 basis points at the 5-year
node, how much does a bond portfolio lose? Does the exposure come from the short end or the long
end of the curve? These questions define interest rate risk, and answering them requires moving
from a single DV01 number, which assumes a parallel shift of the entire curve, to a full vector of
Key Rate Durations that decomposes the sensitivity node by node.
Formally, the parallel DV01 of Module 1 measured:

18
Renato Palma Valencia Fixed Income Analytics

dP
DV01 = − · 0.0001
dy
∂r(ti )
under the assumption ∂y = 1 for all i. Module 3 relaxes this assumption and measures
instead:

1 ∂P
KRDk = −
P ∂r(tk )

for each node tk independently, yielding a complete picture of where interest rate risk lives along
the curve.

19
Renato Palma Valencia Fixed Income Analytics

3 Module 3: Sensitivity Analysis

Python Implementation

Module: src/[Link]
Classes: BondSensitivity and PortfolioSensitivity
Notebook: notebooks/03 [Link]

3.1 Introduction and Class Hierarchy

Module 2 delivered a complete yield curve r(t) and its discount factors Z(ti ). This module asks:
how sensitive is the value of a bond or portfolio to movements in that curve? Module 1 answered
this with a single number, the parallel DV01, under the restrictive assumption that all rates move
by the same amount. Here we relax that assumption entirely. The key insight is that interest rate
risk is not scalar but vectorial: a portfolio has a different sensitivity at each point of the curve, and
understanding that vector is what separates a risk manager who can hedge from one who cannot.
The two classes in src/[Link] implement this decomposition for a single bond and for a
portfolio.

src/[Link]
|
|-- BondSensitivity
| |
| |-- Inputs
| | |-- bond VanillaBond instance
| | |-- curve DiscountCurve instance
| |
| |-- Pricing
| | |-- _price(curve) P = sum CF_i * Z(t_i)
| |
| |-- Parallel Risk
| | |-- dv01_parallel()
| | |-- modified_duration_curve()
| |
| |-- Key Rate Risk
| | |-- key_rate_durations() KRD_k vector
| | |-- dv01_by_key_rate() DV01_k vector
| |
| |-- Summary
| |-- summary() All metrics + additivity check
|
|-- PortfolioSensitivity
|
|-- Inputs
| |-- bonds List of VanillaBond
| |-- curve DiscountCurve

20
 Renato Palma Valencia Fixed Income Analytics

| |-- weights Units held per bond

| |-- labels Bond identifiers
|
|--| Aggregation
| |-- portfolio_value()
| |-- dv01_parallel()
| |-- key_rate_durations() Value-weighted KRD
| |-- dv01_by_key_rate() Summed DV01_k
|
|--| Applications
| |-- scenario_pnl(curve_shift)
| |-- hedge_ratios(hedges)
|
|--| Reporting
|-- summary_table() One row per bond + total

The design mirrors the mathematical structure exactly. BondSensitivity wraps a VanillaBond
and a DiscountCurve and computes sensitivities by perturbing the curve node by node. Portfolio
Sensitivity aggregate those individual vectors across bonds:
1 # Single bond sensitivity
2 sens = B ondSensi tivty ( bond , curve )
3 krd = sens . k e y _ r a t e _ d u r a t i o n s () # { t_k : KRD_k }
4 dv01 = sens . d v 0 1 _ b y _ k e y _ r a t e () # { t_k : DV01_k }
5 sm = sens . summary () # Additivity check included
6
7 # Portfolio aggregation
8 port = P o r t f o l i o S e n s i t i v i t y (
9 bonds = [ bond_2y , bond_7y , bond_30y ] ,
10 curve = curve ,
11 weights = [10 , 5 , 2]
12 )
13 pnl = port . scenario_pnl ( curve_shift ) # dv ~ - DV01 * dr

3.2 Parallel Shift DV01

In module 1 we defined DV01 as the price sensitivity of a single bond to a 1 bp parallel shift in
its yield. Here we extend this concept to the full yield curve setting, where the discount curve r(t)
replaces the constant yield y.
The price of a bond under the full curve is:
n
X n
X
P = CFi · Z(ti ) = CFi · e−r(ti )·ti
i=1 i=1

Definition 3.1 (Parallel Shift). A parallel shift of ∆y moves every point on the curve by the same
amount

r(t) → r(t) + ∆y, ∀t

21
Renato Palma Valencia Fixed Income Analytics

The DV01 under a parallel shift is then:

∂P P (r(t) − 0.0001) − P (r(t) + 0.0001)

DV01Parallel = − · 0.0001 ≈
∂y 2
where the numerical approximation uses a central finite difference. This is identical to the single-
bond DV01 of module 1 when the curve is flat, but extends naturally to arbitrary curve shapes.
In terms of the Modified Duration:

DV01Parallel = Dmod · P · 0.0001

where Dmod is computed under the full curve rather than a single yield. Figure 7 shows DV01
against maturity and the price impact of a change of 100 bps in the yield. We can see that the
longer the maturity, the longer the DV01.

Figure 7: DV01 against Maturity and Price Impact of ∆y = 1%

The parallel DV01 captures the aggregate interest rate risk of the bond but tells us nothing about
which part of the the curve drives that risk. This limitation motivates the Key Rate Durations.

3.3 Key Rate Durations

As we noticed before, the parallel shift assumption is a significant simplification. In practice, yield
curves do not move in parallel: short rates and long rates respond differently to monetary policy,
economic conditions and supply-demand dynamics. Key Rate durations (KRDs), introduced by
Ho (1992), decompose the total interest rate sensitivity into contributions from each point on the
curve.
Definition 3.2 (Key Rate Durations). The key rate duration at a node tk is the price sensitivity
of the bond to a 1 bp shift at that specific maturity, holding all the other rates constant:
1 ∂P
KRDk = −
P ∂r(tk )

22
Renato Palma Valencia Fixed Income Analytics

In practice this is computed numerically via central finite differences:

P (r(tk ) − ε) − P (r(tk ) + ε)
KRDk ≈
2ε · P
where ε = 0.0001 (1 bp) and all the other rates remain fixed. The result is a vector of sensitivities,
one per key rate node:
KRD = (KRD1 , KRD2 , . . . , KRDm )
KRD is not only a partial derivative, when a node changes, the movement spreads across the
adjacent points: the shape of the perturbation depends of the interpolation method chosen (see for
instance module 2). The KRD vector has two fundamental properties.
1. Additivity: The sum of all KRDs equals the Modified Duration under the full curve:
m
X
KRDk = Dmod
k=1

This provides a natural validation: If the KRDs do not sum to the total Modified Duration,
there is an error in the implementation. Figure 8 shows four bonds (2Y 4%, 7Y 5%, 30Y 5%
and 10Y 0%) and shows the KRD across the key rate nodes, also calculating the modified
duration, confirming numerically the additivity property. We have to be careful, because this
result is true only when there is continuous capitalization of the yield. Additivity is exact
when each node of bootstrapping corresponds exactly yo a cash flow, which is guaranteed
under continuous capitalization. In discrete capitalization there is a small cross term that
makes additivity an approximated property.
2. Locality: KRDk is large when cash flows are concentrated near maturity tk and small oth-
erwise. A zero coupon bond maturing at tk has KRDk = tk and all the other KRDs equal to
zero.
Pm
Proposition 3.1. For a bond price under continuously compounded curve, k=1 KRDk = Dmod .
Proof. A parallel shift ∆y can be decomposed as the sum of individual key rate shifts of ∆y at each
node. Since price is linear in small shifts:
m m
1 ∂P 1 X ∂P X
− =− = KRDk
P ∂y P ∂r(tk )
k=1 k=1

The left-hand side is exactly Dmod by definition.

The dollar sensitivity at each node is:
DV01k = KRDk · P · 0.0001
So the total DV01 satisfies:
m
X
DV01Parallel = DV01k
k=1

This decomposition is directly what risk systems report to risk managers: not a single number but
a vector of sensitivities showing exactly which part of the curve a position is exposed to.

23
Renato Palma Valencia Fixed Income Analytics

Figure 8: Key Rate Duration by Bond Type

3.4 Portfolio Sensitivity

The KRD framework extends naturally from a single bond to a portfolio. Let a portfolio consists
of N bonds with prices Pj and KRD vectors KRD(j) , held in quantities wj . The portfolio value is:
N
X
V = wj Pj
j=1

The portfolio KRD at a node tk is the value-weighted average:

PN (j)
portfolio j=1 wj Pj · KRDk
KRDk =
V
and the portfolio DV01 at each node:
N
(j)
X
DV01portfolio
k = wj · DV01k
j=1

This vector representation enables two powerful risk management applications.

24
Renato Palma Valencia Fixed Income Analytics

1. Hedging: Given a target portfolio with DV01 vector dtarget and a set of hedging instruments
with DV01 vectors d(j) , the hedge ratios hj solves:
X
hj d(j) = −dtarget
j

This is a linear system that can be solved exactly if the number of instruments equals the
number of key rate nodes, or approximately via least squares otherwise. A perfect hedge of
KRDs does not only protect against parallel shifts, but against curve shifts (butterfly shifts)
and slope changes (twists).
2. Scenario Analysis: Given a non-parallel curve shift ∆r = (∆r(t1 ), . . . , ∆r(tm )), the ap-
proximate portfolio P &L is:
m
X
∆V ≈ −V KRDk · ∆r(tk ) = −dportfolio · ∆r
k=1

This is the inner product of the DV01 vector with the curve shift vector: a simple but powerful
tool for stress testing.
Consider a portfolio with three bonds: 2Y 4%, 7Y 5%, 30Y 5%. Figure 9 shows the stacked

Figure 9: Stacked KRD of a 3-Bond-Portfolio

key rate duration by nodes, where the highest KRDs concentrate at the maturities of the bonds
conforming the portfolio.

3.5 Conclusions and Key Takeaways

• The parallel DV01 is a useful summary statistic but hides the true structure of risk. Two
portfolios with identical parallel DV01 can have very different exposures to curve steepening,
flattening or butterfly moves, only the KRD vector reveals this.
• The additivity property k KRDk = Dmod is both a mathematical result and a practical
P
validation tool: any implementation that violates it contains an error.

25
Renato Palma Valencia Fixed Income Analytics

• Hedge ratios from the linear system j hj d(j) = −dtarget may require short positions. This
P
is economically correct and standard in practice: a portfolio concentrated at the long end is
hedged by shorting long-dated instruments, not by buying short-dated ones.

The KRD vector KRD ∈ Rm has one entry per curve node (typically 10 or more). Working
with this full vector is informative but computationally expensive and statistically redundant: the
nodes are highly correlated, and most of the movement in the yield curve is driven by a small
number of underlying forces. A natural question is: can we find a lower-dimensional representation
that captures most of the risk without losing important information? The answer is yes, and it
comes from the covariance structure of historical yield curve changes. If Σ is the covariance matrix
of daily rate changes, then the portfolio P&L variance is:

Var(∆V ) = dT Σ d

where d ∈ Rm is the DV01 vector of this module. Evaluating this expression requires estimating
and inverting Σ, a full m × m matrix. Module 4 decomposes Σ via PCA into three dominant
eigenvectors: level, slope and curvature and shows that:
3
X
2
Var(∆V ) ≈ λk vkT d
k=1

reducing the problem from a full m × m matrix to three scalar projections. The DV01 vector d
computed here is the direct input to that calculation.

26
Renato Palma Valencia Fixed Income Analytics

4 Module 4: PCA on Term Structure

Python Implementation

Module: src/[Link]
Classes: YieldCurvePCA and PCARiskManager
Notebook: notebooks/04 [Link]

4.1 Introduction and Class Hierarchy

Modules 2 and 3 together produce a yield curve r(t) and a KRD DV01 vector d ∈ Rm that describes
where risk lives along the curve. The natural next question is: can we compress that m-dimensional
risk vector into something smaller without losing important information? The answer exploits a
remarkable empirical regularity. Daily changes in yield curves across all major markets are highly
correlated: when rates move, they rarely move in arbitrary directions. Instead they move along
a small number of dominant patterns (level, slope and curvature) that account for over 99% of
total variance. PCA discovers these patterns statistically from historical data and transforms the
m×m covariance problem of Module 3 into three scalar projections. The two classes in src/[Link]
implement this decomposition and connect it back to the risk vector d from Module 3.

src/[Link]
|
|-- YieldCurvePCA
| |
| |-- Inputs
| | |-- df_yields Historical yield panel
| | |-- n_factors Components to retain
| |
| |-- Core (_fit)
| | |-- X Daily changes matrix
| | |-- Sigma Covariance matrix
| | |-- eigenvalues lambda_1 >= ... >= lambda_m
| | |-- eigenvectors v_1, ..., v_m
| |
| |-- Variance
| | |-- variance_explained(K) R2_K
| | |-- variance_table() Full table
| |
| |-- Factor Structure
| | |-- loadings(K) V_K matrix
| | |-- scores(K) F = X @ V_K
| | |-- factor_shock(k, sigma) Stress scenario
| |
| |-- Reconstruction
| | |-- reconstruct(K) X_hat = F_K @ V_K.T
| | |-- reconstruction_error(K) MSE analytical

27
 Renato Palma Valencia Fixed Income Analytics

| |
| |-- NSS Alignment
| |-- nss_alignment(lambda) R2 per factor
|-- PCARiskManager
|
|-- Inputs
| |-- pca YieldCurvePCA instance
| |-- dv01_vec KRD DV01 vector from M3
|
|-- Factor Risk
| |-- factor_dv01() DV01_PCk = v_k.T @ d
| |-- total_dv01() Sum of KRD DV01s
| |-- pnl_variance(K) Var(dV) exact + approx
|
|-- Immunization
| |-- immunization_check() DV01_PC1 = 0?
|
|-- Scenarios
|-- scenario_pnl(shocks)
|-- summary()

The two classes are designed to be used in tandem. YieldCurvePCA is fitted once on historical
data and exposes the eigenstructure of the covariance matrix. PCARiskManager takes a fitted PCA
object and a DV01 vector from Module 3 and translates node-level risk into factor-level risk:
1 # Fit PCA on historical yields
2 pca = YieldCurvePCA ( df_yields , n_factors = 3)
3
4 # Connect to portfolio DV01 from Module 3
5 rm = PCARi skManage r ( pca , dv01_vec )
6
7 # Factor DV01 decomposition
8 fdv01 = rm . factor_dv01 ()
9 # { ’ PC1 Level ’: 0.082 , ’ PC2 Slope ’: -0.011 , ...}
10
11 # P & L variance under factor model
12 pnl = rm . pnl_variance ()
13
14 # Immunization check : is DV01_PC1 ~ 0?
15 imm = rm . i m m u n i z a t i o n _ c h e c k ()

4.2 Motivation
The yield curve does not move randomly: when the FED raises rates, the entire curve tends to
shift upward. When the economy slows, short rates fall faster than long rates, steepening the curve.
These patterns suggest that the high-dimensional object we call the yield curve (a vector of rates
at 10 or more maturities) is driven by a small number of underlying factors. Principal Component
Analysis (PCA) formalizes this intuition. Given a historical panel of yield curve changes, PCA
finds the directions of maximum variance in the data: the directions along which the curve moves

28
Renato Palma Valencia Fixed Income Analytics

the most. Litterman et al. (1991) showed empirically that three such directions explains over
99% of total yield curve variance, and that they correspond exactly to parallel shifts (level), tilts
(slope) and humps (curvature). This result has profound implications both for risk management
and for the connection with the NSS model of module 2. To motivate this, notice how different
maturities correlate. Figure 10 shows how maturities co-move and that they are highly correlated

Figure 10: Covariance and Correlation Matrix

at neighboring maturities and less so at the extremes.

4.3 PCA Framework

Let ∆rt ∈ Rm be the vector of daily changes in spot rates at m maturities on day t:
∆rt = (∆r(t1 ), ∆r(t2 ), · · · , ∆r(tm ))T
Over T trading days we observe the data matrix:
X ∈ RT ×m
where row t is the vector of rate changes on day t. After centering (subtracting the column means),
the empirical covariance matrix is:
1
Σ= XT X ∈ Rm×m
T −1
Definition 4.1 (Principal Components). The principal components of X are the eigenvectors of
Σ, obtained via the spectral decomposition:
Σ = VΛVT
where Λ = diag(λ1 , λ2 , . . . , λm ) with λ1 ≥ λ2 ≥ · · · ≥ λm ≥ 0 are the eigenvalues and V =
[v1 |v2 | · · · |vm ] is the orthogonal matrix of eigenvectors, so that VT V = I.
Each eigenvector vk ∈ Rm is a factor loading vector : it describes a pattern of simultaneous rate
movements across all maturities. The corresponding eigenvalue λk equals the variance explained
by that factor.

29
Renato Palma Valencia Fixed Income Analytics

Definition 4.2 (Factor Scores). The projection of the data onto the k-th principal component is
the factor score:

fk = Xvk ∈ RT

It represents the time series of how much each daily curve move was driven by factor k.

The total variance in the data is the trace of the covariance matrix:
m
X
Total Variance = tr(Σ) = λk
k=1

The fraction of variance explained by the first K principal components is:

PK
2 λk
RK = Pk=1
m
k=1 λk

In practice for U.S. Treasury yields:

R12 ≈ 80% R22 ≈ 95% R32 ≈ 99%

This means that three numbers (the scores on three factors) capture essentially all the information
in the daily movement of ten or more yield curve points. The remaining factors are noise. Using
synthetic data in the notebook, we found the following results.

Figure 11: Variance Explained by Component

Figure 11 shows the explained variance across components, and it is clear that a ≈ 98% of the
variance is explained by the first 3 components.

4.4 The Three Factors (Level, Slope and Curvature)

The empirical result of Litterman et al. (1991) is that the three dominant eigenvectors have stable
and economically interpretable shapes across markets and time periods:

30
Renato Palma Valencia Fixed Income Analytics

1. PC1 (Level): The first eigenvector v1 has entries of approximately equal sign and magnitude
across all maturities. A unit move along PC1 shifts the entire curve up or down by roughly
the same amount at every maturity. This is, the parallel shift scenario of module 3.
1
v1,j ≈ √ , ∀j
m

2. PC2 (Slope): The second eigenvector v2 has positive entries at short maturities and negative
entries at long maturities (or vice versa depending on sign convention). A unit move along
PC2 steepens or flattens the curve: short rates move in the opposite direction to long rates.
3. PC3 (Curvature): The third eigenvector v3 has entries that are positive at short and long
maturities and negative at intermediate maturities (or vice versa). A unit move along PC3
creates a hump in the middle of the curve, corresponding to the curvature parameter β2 in
the NSS model.
This connection with NSS is not coincidental. Both frameworks are describing the same un-
derlying structure: the yield curve lives in a low-dimensional subspace spanned by level, slope and
curvature. NSS imposes this structure parametrically from the start; PCA discovers it statistically
from the data.

Figure 12: Eigenvectors vs Maturities

We can visualize the eigenvectors associated with the first three components in Figure 12. We
can easily see what we stated, the first eigenvector v1 (level) is relatively flat across all maturities
(parallel shift), the second eigenvector v2 (slope) has positive entries at short and negative entries
at long maturities. The third eigenvector v3 (curvature) has positive entries in the very short
maturities, negative entries at medium maturities and positive again at long maturities.
Figure 13 shows the time series of each factor score fk,t = Xt · vk over the sample period. Three
observations stand out. First, the PC1 score has the largest amplitude by construction: its variance
equals λ1 , the largest eigenvalue, so it is the most volatile series. Second, PC2 and PC3 scores are
progressively quieter, reflecting their smaller eigenvalues. Third, the scores carry economic content:
sustained negative values of the PC2 score correspond to periods of yield curve inversion (short rates
rising relative to long rates), while sharp spikes in the PC3 score correspond to episodes where the
belly of the curve moved independently of the wings, typical during central bank forward guidance
transitions.

31
Renato Palma Valencia Fixed Income Analytics

Figure 13: Factor Scores over Time

4.5 PCA Reconstruction and Risk Management

Given the first K principal components, the yield curve change on day t can be approximated as
K
X
∆rt ≈ fk,t · vk = VK fK,t
k=1

where VK ∈ Rm×K contains the first K eigenvectors and fK,t ∈ RK are the factor scores on day t.
The reconstruction error is:
εt = ∆rt − VK fK,t
and its mean squared error across all days and maturities is exactly:
T m
1X 2
X
MSE = ∥εt ∥ = λk
T t=1
k=K+1

The sum of the discarded eigenvalues. This provides a clean analytical formula for the reconstruction
error without having to compute it empirically. A powerful application of the reconstruction is the

32
Renato Palma Valencia Fixed Income Analytics

definition of stress scenarios. Instead of moving nodes randomly, we can move the factors x standard
deviations and see what happens, generating a plausible market scenario. This reconstruction error

Figure 14: PCA Reconstruction

is visualized in Figure 14, where we can see the clear match between the reconstructed yield curve
and the actual ∆rt .
The PCA decomposition transforms the problem of managing risk at m correlated nodes into the
problem of managing risk along K uncorrelated factors. The DV01 of the bond portfolio along
factor k is:
DV01P Ck = vkT · d
where d ∈ Rm is the vector of key rate DV01 from module 3. Since the factors are orthogonal, the
total variance of the portfolio P &L is:
K
X
Var(∆V ) = (DV01P Ck )2 λk
k=1

This is much simpler than working with the full m × m covariance matrix of key rate sensitivities,
and is exactly the computation performed by risk engines under a factor risk model. If we want for
instance a level neutral portfolio, we search for DV01P C1 = 0, this is more robust than traditional
immunization by duration because PCA knows that the curve does not change exactly parallel.
This is known as Factor Immunization.
PK
Proposition 4.1. Under the factor model ∆r = k=1 fk vk with Var(fk ) = λk and uncorrelated
factors, the portfolio P &L variance is:
K
X
Var(∆V ) = dT Σd ≈ (vkT d)2 λk
k=1
T
Proof. The P&L is ∆V = −d ∆r. Its variance is:
m
X
Var(∆V ) = dT Σd = (vkT d)2 λk
k=1

33
Renato Palma Valencia Fixed Income Analytics

Truncating to the first K factors gives the approximation.

Figure 15: Factor Risk Management

For a portfolio with three bonds (2Y 4%, 7Y 5%, 30Y 5%), Figure 15 plots DV01P Ck against
KRDs, and the total P&L variance explained by each component.
The connection between PCA and NSS runs deeper than a superficial analogy. Define the NSS
loading vectors:

1 − e−tj /λ 1 − e−tj /λ
ϕ0 = 1 ϕ1 (j) = ϕ2 (j) = − e−tj /λ
tj /λ tj /λ
evaluated at each maturity tj . These are the columns of the NSS design matrix. The claim is that the
empirical eigenvectors v1 , v2 , v3 from PCA are well approximated by ϕ0 , ϕ1 , ϕ2 up to normalization
and sign. This can be verified empirically: regress each PCA eigenvector onto the NSS loading
vectors and compute the R2 . In practice, R2 = 0.95 for all three factors, confirming that NSS
and PCA are discovering the same underlying structure from different angles: one parametric, one
statistical. If R2 is low, it means that λ in the NSS model is not well-calibrated to capture the
current historic dynamic of the market. PCA factors are data-dependent while NSS is a fixed
structure, we need to choose between real statistic or an imposed functional form. In the notebook,
we see that factor 1 v1 is well-aligned between PCA and NSS, not the same result yields from the
other factors. Figure 16 compares both eigenvectors against maturities and it is easy to see that
the curves with NSS are excessively smooth. Formalizing the visualization, we have to notice that
NSS has a fixed exponential functional form parametrized by λ1 and λ2 , while PCA captures the
exact empiric form of the historic curve. The excess of smoothness of NSS in PC2 and PC3 is
consequence that one and only one exponential cannot reproduce complex humps that appear in
the data.

4.6 Conclusions and Key Takeaways

• Three factors explain over 99% of daily yield curve variance. The remaining components
are idiosyncratic noise that cannot be hedged systematically and should not drive portfolio
positioning.

34
Renato Palma Valencia Fixed Income Analytics

Figure 16: PCA Eigenvectors vs NSS Loading Vectors

• Factor immunization (DV01P C1 = 0) is strictly more robust than duration matching: it

neutralizes the dominant source of yield curve risk without assuming a perfectly parallel shift,
which never occurs in practice.
• PCA and NSS discover the same underlying structure from different angles: one statistical,
one parametric. When R2 between eigenvectors and NSS loadings is high (> 0.90), the two
models are consistent and the NSS calibration can be trusted. When R2 is low for PC2 or
PC3, it signals that the chosen decay parameter λ is not well-suited to the current market
regime.

The PCA of this module has a direct financial application beyond risk management. The yield
curve r(t) from Module 2 and its discount factors Z(ti ) are the inputs to pricing the most liquid
fixed income derivative in the world: the Interest Rate Swap (IRS). An IRS exchanges a fixed rate
K for a floating rate (SOFR or EURIBOR) on a notional N over a sequence of dates t1 , . . . , tn . Its
fair fixed rate, the swap rate K ∗ , is determined entirely by the discount factors of Module 2:

1 − Z(tn )
K ∗ = Pn
i=1 δi · Z(ti )

The connection with PCA is immediate: the DV01 of a swap is also a vector d ∈ Rm , concentrated
at the payment dates of the fixed leg, and the factor DV01 decomposition DV01P Ck = vk⊤ d applies
without modification. A swap desk that manages its book in factor space (neutralizing PC1, PC2
and PC3 exposures) is running exactly the risk framework developed across Modules 3 and 4.

35
Renato Palma Valencia Fixed Income Analytics

5 Module 5: IRS Pricing

Python Implementation

Module: src/[Link]
Classes: VanillaIRS and SwapPortfolio
Notebook: notebooks/05 [Link]

5.1 Introduction and Class Hierarchy

The yield curve of Module 2 and the sensitivity framework of Module 3 converge in this final module
to price the most liquid fixed income derivative in the world: the Interest Rate Swap (IRS). With
over $400 trillion in outstanding notional, the IRS market dwarfs all bond markets combined, and
it is the instrument for which systems like Murex MX.3 were fundamentally designed. The key
insight of this module is that an IRS is not a new mathematical object: it is a portfolio of Forward
Rate Agreements, each of which is priced using exactly the forward rates and discount factors of
Module 2. The swap rate K ∗ is determined entirely by the discount curve, and the swap DV01 is a
vector in Rm that feeds directly into the PCA risk framework of Module 4. The project thus closes
a complete cycle: from a single bond price to a factor-based risk decomposition of a swap portfolio.

src/[Link]
|
|-- VanillaIRS
| |
| |-- Attributes
| | |-- N Notional
| | |-- K Fixed rate
| | |-- tenor Maturity in years
| | |-- freq Payment frequency
| | |-- curve DiscountCurve (Module 2)
| | |-- position ’long’ or ’short’
| | |-- pay_dates Payment schedule
| | |-- Z Discount factors at pay_dates
| |
| |-- Leg Pricing
| | |-- fixed_leg_value(curve)
| | |-- floating_leg_value(curve) = N always
| |
| |-- Swap Pricing
| | |-- swap_rate() K* formula
| | |-- value(curve) MtM = float - fixed
| | |-- inception_value() Validation: must be 0
| |
| |-- FRA Decomposition
| | |-- fra_values() One row per period
| | sum = swap value at K*

36
 Renato Palma Valencia Fixed Income Analytics

| |
| |-- Sensitivity
| | |-- dv01() Parallel shift
| | |-- key_rate_dv01() KRD DV01 vector
| | |-- modified_duration()
| | |-- value_vs_rate() P&L profile
| |
| |-- Summary
| |-- summary() All metrics
|
|-- SwapPortfolio
|
|-- Aggregation
| |-- total_value(curve)
| |-- total_dv01()
| |-- key_rate_dv01()
|
|-- Applications
| |-- scenario_pnl(curve_shift)
| |-- summary_table()
|
|-- Factory
|-- curve_trade(N, short, long, freq, curve)

The two classes mirror the mathematical structure exactly. VanillaIRS prices a single swap and
exposes its complete sensitivity profile, while SwapPortfolio aggregates across multiple contracts
and implements the classic curve trade:
1 # Price a 10 y semi - annual swap at fair rate
2 swap = VanillaIRS (1 _000_000 , 0.0 , 10 , 2 , curve )
3 K_star = swap . swap_rate ()
4 swap = VanillaIRS (1 _000_000 , K_star , 10 , 2 , curve )
5
6 # Validate : inception value must be zero
7 assert abs ( swap . i n ce pt io n _v al ue () ) < 1e -4
8
9 # FRA decomposition : sum must be zero at K = K *
10 fra = swap . fra_values ()
11 assert abs ( fra [ " FRA Value " ]. sum () ) < 1e -4
12
13 # Curve trade : pay fixed 2y , receive fixed 10 y
14 trade = SwapPortfolio . curve_trade (
15 1 _000_000 , 2 , 10 , 2 , curve
16 )

5.2 Interest Rate Swap Definition and Structure

Definition 5.1 (Interest Rate Swap). An interest rate swap (IRS) is a contract between two
counter-parties that agree to exchange cash flows at predetermined future dates t1 , t2 , . . . , tn . In
the most common form, the plain vanilla IRS, one party pays a fixed rate K and receives a floating

37
Renato Palma Valencia Fixed Income Analytics

rate (typically SOFR or EURIBOR) 1 , both applied to a notional amount N . The net cash flow at
each period i is:
Vi = N · (Li−1 − K) · δi
where Li−1 is the floating rate observed at the start of the period i and δi = ti − ti−1 is the day
count fraction. The party paying fixed and receiving floating has a long position; the party paying
floating and receiving fixed has a short position.
The fundamental insight is that a Swap is nothing more than a portfolio of forward rate agree-
ments (FRAs), one for each payment date. Each FRA exchanges a fixed payment against a floating
payment at a single future date, and summing them yields the full swap. This connection, developed
in section 5.4, is the most elegant result of the module.

5.3 The Floating Leg equals Par

Before pricing the swap, we need one foundational result: the present value of the floating leg
always equals the notional N at the start of each reset period. This is the key to the entire pricing
framework.
Proposition 5.1 (Floating Bond = Par). The present value of the floating leg at t = 0 is:
Vfloat = N
Proof. The floating leg pays N · Li−1 · δi at each ti . By the definition of the LIBOR/SOFR rate,
each payments satisfies:
N · Li−1 · δi · Z(ti ) = N · (Z(ti−1 ) − Z(ti ))
This follows because Li−1 is defined precisely as the rate that makes the present value of a unit
investment grow from Z(ti−1 ) to Z(ti ) over the period. Summing telescopically over all periods
and adding the present value of the return of the notional at tn :
n
X
Vfloat = N (Z(ti−1 ) − Z(ti )) + N · Z(tn ) = N (Z(t0 ) − Z(tn )) + N · Z(tn ) = N · Z(0) = N
i=1

The telescopic sum collapses beautifully: the floating leg is worth par regardless of the shape of
the yield curve, because the floating rate resets continuously to the market rate.

5.4 Swap as a Portfolio of FRAs

Definition 5.2 (Forward Rate Agreement). A FRA is a contract where a fixed rate K is agreed
today for a loan over a future period [ti−1 , ti ]. Its value at t = 0 is:
VFRA = N · δi · (f (ti−1 , ti ) − K) · Z(ti )
where f (ti−1 , ti ) is the forward rate defined in section 2.3.
1 In modern post-crisis markets, a multi-curve framework is used: the floating leg is projected using a forwarding

curve (e.g., EURIBOR 6 M), while all cash flows are discounted using a risk-free OIS curve (e.g., SOFR or ESTR). For
simplicity, this module assumes a single-curve environment where Z(t) is used for both projection and discounting.

38
Renato Palma Valencia Fixed Income Analytics

The IRS is then exactly the sum of these FRAs:

n
X n
X
Vswap = VFRA (ti ) = N · δi · (f (ti−1 , ti ) − K) · Z(ti )
i=1 i=1

This decomposition makes the pricing completely transparent: the swap is worth the present value
of the difference between the forward rates the market expects and the fixed rate K the counter-
party pays.

5.5 Swap Rate K*

Definition 5.3 (Swap Rate K ∗ ). The swap rate K ∗ is the fixed rate that makes the present value
of the swap zero at inception:

Vswap (K ∗ ) = 0

Proposition 5.2. The swap rate is given by:

1 − Z(tn )
K ∗ = Pn
i=1 δi · Z(ti )

Proof. The swap can be decomposed as the difference between a floating bond and a fixed coupon
bond:
n
!
X
Vswap = Vfloat − Vfixed = N − N K · δi · Z(ti ) + Z(tn )
i=1

setting Vswap = 0 and dividing by N :

n
X
1 = K∗ · δi · Z(ti ) + Z(tn )
i=1
1 − Z(tn )
K ∗ = Pn
i=1 δi · Z(ti )

The numerator 1 − Z(tn ) is the present value of a basis point (almost): it measures how much
the discount factor has decayed from today to maturity. The denominator is the present value of
an annuity of δi at each payment date and it represents the sensitivity of the swap value to a 1 bp
change in the swap rate. The swap is therefore the rate of annuity whose present value matches
the decay of the discount factor to maturity. In the notebook, we use the following vanilla Swap:
Notional N = 1000000, fair swap rate K ∗ = 5.4939%, tenor = 10Y, frequency = 2, position = long.
Making a FRA decomposition, we get the FRA values by period showed in Figure 17 and validate
that the sum of FRAs has to be zero.

39
Renato Palma Valencia Fixed Income Analytics

Figure 17: FRA Decomposition

5.6 Swap Value at t

Once the swap is agreed at rate K, if rates change the swap acquires positive or negative value.
For the fixed payer (long position), the value at time t is:

Vt = Vfloat (t) − Vfixed (t)

Using the result that the floating leg always resets to par at each reset date, and recomputing the
fixed leg under the new discount curve:
" #
X
Vt = N 1 − Zt (tn ) − K δi · Zt (ti )
i:ti >t

where Zt (·) denotes discount factors under the curve at time t. Equivalently this can be written as:
X
Vt = (Kt∗ − K) δi · Zt (ti )
i:ti >t

where Kt∗ is the current swap rate for the remaining tenor. If rates have risen (Kt∗ > Kt ), the fixed
payer benefits because they are playing below-market fixed rate. If rates have fallen (Kt∗ < Kt ) the
fixed payer losses. In Figure 18 we can visualize how the swap curve behaves against the spot and
forward curve, noting that the spot curve r(t) is above the swap curve K ∗ (t) for long tenors and
very similar for short tenors.

5.7 Swap DV01 and Applications

The DV01 swap under a parallel shift of the yield curve is:

DV01swap = DV01fixed leg − DV01floating leg

Since the floating leg has very low duration (it resets to par at every payment date, so its effective
duration is approximately δ1 /2, half of the first period), the swap DV01 is dominated by the fixed

40
Renato Palma Valencia Fixed Income Analytics

Figure 18: Swap Curve vs Spot Curve

leg:
fixed
DV01swap ≈ Dmod · Vfixed · 0.0001

The key rate DV01 profile of the swap is concentrated at intermediate maturities, peaking near the
midpoint of the swap tenor. This contrasts sharply with a zero coupon bond, which concentrates
all risk at maturity, and makes swaps particularly useful instruments for targeting specific parts
of the yield curve. The swap has no convexity from the floating leg, all convexity comes from the
fixed leg only. In figure 19 we can visualize a swap P&L profile against parallel shifts of the yield

Figure 19: Swap Sensitivity Decomposition

and KRD of a DV01. We can notice that the key rate nodes gets more relevant at long tenors near
to maturity and that positive changes in level of the yield increases the value of the long position
swap and when rates rise, the fixed payer benefits because they locked in a below-market fixed rate.
Swaps are the most liquid fixed income derivatives in the world, with outstanding notional exceeding
$400 trillion. Their three main uses are:

41
Renato Palma Valencia Fixed Income Analytics

1. Debt transformation: A corporate issuer with floating rate debt (paying LIBOR + spread)
can enter a swap paying fixed and receiving floating, effectively converting the floating obli-
gation to a fixed rate. The swap eliminates the uncertainty of future interest payments.
2. Duration management: A pension fund or insurance company with long-dated liabilities
wants to increase the duration of its asset portfolio without buying bonds. By receiving fixed
in a long-dated IRS, the fund increases duration efficiently and at low transaction cost, since
the swap requires no upfront payment.
3. Yield curve positioning: A portfolio manager who believes the yield curve will flatten
can receive fixed in a long-dated swap and pay fixed in a short-dated swap simultaneously,
creating a position that profits from flattening regardless of the absolute level of rates. This
is the classic curve trade in fixed income.
4. Economic Advantage: IRS allow the ”decoupling” of credit risk and interest rate risk.
This enables a more efficient capital allocation in the economy, as companies can choose their
funding structure based on their comparative advantage (credit) and then adjust their interest
rate exposure through derivatives.

5.8 Conclusions and Key Takeaways

• The floating leg always equals par regardless of the shape of the yield curve. This is not
an approximation but an exact result from the telescoping property of discount factors: all
pricing complexity of the swap lives in the fixed leg alone.

• The swap rate K ∗ is the weighted average of the forward rates implied by the discount curve,
with weights proportional to δi · Z(ti ). In an upward-sloping curve, short forward rates pull
K ∗ below the long-maturity spot rate, explaining why the swap curve lies below the spot
curve for long tenors.

• The FRA decomposition exposes the internal structure of the swap: periods where the forward
rate exceeds K contribute positive value to the fixed payer, and vice versa. At K = K ∗ these
contributions cancel exactly, confirming the fair pricing condition. Note that the forward rates
in each FRA must be computed under simple compounding, consistent with how LIBOR and
SOFR rates are quoted in practice.

• The curve trade (pay fixed short, receive fixed long) is P&L neutral to parallel shifts when
DV01s are matched. Profit and loss arise purely from changes in the slope of the yield curve,
making it the canonical instrument for expressing a view on curve shape without directional
rate exposure.

42
Renato Palma Valencia Fixed Income Analytics

6 Project Summary and Further Research

What This Project Built
This project developed a complete, self-contained fixed income analytics library in Python, grounded
in formal mathematical derivations and validated against real market data from the Federal Reserve.
The five modules form a coherent chain:
1. Module 1 established the mathematical foundations of bond pricing, from closed-form for-
mulas to Newton-Raphson yield extraction and second-order risk approximations.
2. Module 2 replaced the constant discount rate with a full term structure, building the
yield curve from observable prices via bootstrapping and extending it to arbitrary matu-
rities through three interpolation methods of increasing sophistication: linear, cubic spline
and Nelson-Siegel-Svensson.
3. Module 3 decomposed interest rate risk from a single parallel DV01 into a full Key Rate Du-
ration vector, enabling granular hedging and non-parallel scenario analysis across a portfolio
of bonds.
4. Module 4 reduced the m-dimensional KRD vector to three orthogonal factor exposures via
PCA, confirmed the Litterman-Scheinkman result empirically, and connected it to the NSS
parametric model, showing that both frameworks discover the same underlying structure.

5. Module 5 closed the cycle by pricing vanilla Interest Rate Swaps using the discount factors
of Module 2, decomposing them into FRA portfolios, and exposing their sensitivity profile as
a KRD DV01 vector that feeds directly into the risk framework of Modules 3 and 4.
The result is a project where each module is both mathematically derived in this document and
numerically implemented in a corresponding Python class, with interactive notebooks that validate
every result against real or realistic market data.

Virtues of the Approach

• Theory-code correspondence: every formula in this document has a direct implementation
in src/, with identical notation. The document and the code are designed to be read side by
side.
• Validation by construction: roundtrip tests (price → YTM → price), the KRD additivity
property, the floating leg = par result and the FRA sum = 0 condition are all checked
programmatically in every notebook run. The library cannot silently produce wrong results.

• Real data: all yield curves are built from U.S. Treasury data sourced from FRED, ensuring
that the results reflect actual market conditions rather than stylized examples.
• Modularity: each class is self-contained and interoperable. A DiscountCurve built in Mod-
ule 2 is consumed by BondSensitivity in Module 3, PCARiskManager in Module 4 and
VanillaIRS in Module 5 without modification.

43
Renato Palma Valencia Fixed Income Analytics

Limitations
• Single-curve framework: All modules assume a single discount curve for both projection
and discounting. Post-2008 markets require a multi-curve framework where the floating leg
is projected using a tenor-specific forward curve (e.g., EURIBOR 6M) and discounted using
an OIS curve (SOFR or ESTR). The mathematical extension is straightforward but requires
a second DiscountCurve object and a modified fra values() method.
• Vanilla instruments only: The library prices plain vanilla bonds and swaps. It does not
cover options, swaptions, caps and floors, callable bonds or any instrument with optionality,
which require stochastic rate models (Hull-White, Black-Karasinski) beyond the deterministic
curve framework of this project.
• Deterministic curve: The yield curve is treated as a deterministic function of time. A
stochastic extension would model r(t) as a stochastic process (e.g., Vasicek, CIR or HJM),
enabling the pricing of interest rate derivatives with early exercise features.
• No credit risk: All instruments are assumed default-free. Incorporating credit spreads,
survival probabilities and recovery rates would extend the framework to corporate bonds and
credit default swaps (CDS).
• Static PCA: The PCA of Module 4 is estimated on a fixed historical window. A dynamic
extension using rolling or exponentially weighted covariance matrices would allow the factor
structure to adapt to changing market regimes, which is particularly important around central
bank policy shifts.

Natural Extensions
Each limitation above is also a natural next step, ordered from most to least immediate:
1. Multi-curve framework (Module 2 extension): Implement separate projection and discount
curves, calibrate basis spreads and reprice the swap portfolio under the post-2008 standard.
2. Interest rate options (Module 6): Implement caps, floors and European swaptions under
Black’s model, connecting the volatility surface to the implied forward rate distribution.
3. Stochastic rate models (Module 7): Implement Vasicek and Hull-White one-factor models,
calibrate them to the observed yield curve and price Bermudan swaptions via trinomial trees
or Monte Carlo simulation.
4. Credit risk (Module 8): Extend the discount curve to incorporate credit spreads, implement
the Merton structural model and price Credit Default Swaps as a natural complement to the
IRS framework.
5. Dynamic risk management (Module 9): Replace the static PCA with a Dynamic Nelson-
Siegel model, implement Kalman filtering to track factor evolution in real time and back-test
the factor immunization strategy across historical rate cycles.
The framework developed here is the foundation on which all these extensions build. Every new
instrument will consume a DiscountCurve , expose a KRD DV01 vector and connect to the PCA
risk engine, because that is how the fixed income world is organized, from the simplest Treasury
bill to the most complex structured product.

44
Renato Palma Valencia Fixed Income Analytics

References
Ho, T. S. (1992). Key rate durations: Measures of interest rate risks. The Journal of Fixed Income,
2(2):29–44.
Litterman, R. B., Scheinkman, J., and Weiss, L. (1991). Volatility and the yield curve. The Journal
of Fixed Income, 1(1):49–53.
Nelson, C. R. and Siegel, A. F. (1987). Parsimonious modeling of yield curves. Journal of business,
pages 473–489.
Svensson, L. E. (1994). Estimating and interpreting forward interest rates: Sweden 1992-1994.

45