Bayesian Methods in Finance

Nick Polson
University of Chicago

1
 Overview
1. Finance Models
Diffusions, Black-Scholes, Options Pricing
Stochastic Volatility and Jump-Diffusions

2. MCMC Financial Econometrics

Applications: Portfolio Choice and Option Pricing

3. Filtering Methods
Sequential filtering and Parameter learning: Particle Filtering

2
 Reading
• Bayesian Analysis of Stochastic Volatility Models (1994, JPR)

• MCMC methods for Financial Econonmetrics (2003, JP)

• The Impact of Jumps in Volatility and Returns (2003, EJP)

• Bayesian Portfolio Selection: the SP500 from 1970-1996 (2000, PT)

• Bayesian Inference for Derivative Prices (2004, PS)

• Sequential Parameter Estimation for Stochastic Volatility models

(2005, JPS)

• Optimal Filtering of Jump-Diffusions: Extracting Latent States from

Asset Prices (2005, JPS)

3
 1: Finance Models

1. Finance Models
Diffusion models with Stochastic Volatility and Jumps
Johannes and Polson (2004) MCMC Financial Econometrics
2. Applications
• Option Pricing
Black-Scholes and Stochastic Volatility
• Portfolio Allocation
SP500 and Nasdaq NDX100 Equity Indicies
Many other puzzles
1. Equity Premium Puzzle
Why did stocks outperform bonds and cash?
2. Historical versus Implied Volatility
Why is option volatility different from historical?

4
 Models and Data
• Diffusion models
Brownian motion (Bachelier, 1900, BS, 1973)
Stochastic Volatility (Taylor, 1986, JPR, 1994)
Jump-Diffusion models (Merton, 1976, DPS, 2000)

• Data: Returns on Assets (Equities, Interest rates, Exchange rates)

The two standard US equity indicies are: SP500 and Nasdaq NDX100
Let’s look at the type of financial time series we wish to analyse

5
 NDX: 1990−2004
5000
Price

0 2000

1990−01−02 1997−04−18 2004−08−02

Time
6

SP500: 1990−2004
1500
Price

0 500

1990−01−02 1997−04−18 2004−08−02

Time
• Brownian Model (Bachelier, 1900)
Prices St :
dSt
= µdt + σdWtP
St
Inference: Estimate expected returns µ and volatility σ

• Basis for Black-Scholes: Option Pricing

• Caveat: Model mis-specification

7
• Stochastic Volatility (Taylor, 1986)
Assume that asset prices St follow a diffusion model of the form
dSt p
= µdt + Vt dWts
St
p
dVt = κv (θv − Vt ) + σv Vt dWtv

This allows for volatility to be stochastic. The parameter σv governs

the volatility of volatility
Leverage: leverage effect ρ is a negative correlation between shocks to
stock prices and volatility
Inference: Estimate Volatility Vt and Θ = (µ, κv , θv , σv ) (JPR, 1994)
Sequential Inference: estimate volatility and parameters in real-time
as data arrives

8
• Jump Diffusion Models (Merton, 1976)
Assume that we can have a jump to stock prices.
 
Nt (P)
X
P
dSt = µSt dt + σSt dWt + d  Sτj− (eZj (P) − 1)
j=1

Inference: Extract volatility, jumps and parameters (EJP, 2003)

We can also extend this to the double jump model where we allow
jumps to volatility as well. (DPS, 2000)

9
 Stochastic Volatility Jump Models
• Later we’ll use an SVJ model:

 
Nt (P)  
p X
(P)
dSt = µPt St dt s
+ St Vt dWt (P) + d  Z
Sτ j − e j − 1 
j=1
p
dVt = κv (θv − Vt ) dt + σv Vt dWtv (P)

• Here Wts (P) and Wtv (P) are Brownian motions, Nt (P) counts the
number of jump times, τj , prior to time t, µt is the equity risk
premium, µPt St is the jump compensator.

10
 Multivariate Merton’s Model
• Merton’s (1976) jump-diffusion model.
 
Nt (P)
X
dSt = µSt dt + σSt dWt (P) + d  Sτj− (eZj (P) − 1)
j=1

• The MCMC algorithm samples from p (Θ, X|Y ) , where

Θ = (µ, Σ, λ, µJ , ΣJ ) and X = (J, Z), where J, Z, and Y are vectors
of jump times, jump sizes and observed prices.

11
 Vasicek with Jumps
• Vasicek’s (1977) model incorporating jumps assumes the short rate
follows:
 P 
  N t
X
P P r
drt = ar − br rt dt + σr dWt (P) + d  ZjP 
j=1

P
where N P and
t is a Poisson  process with intensity λ r
2
ZjP ∼ N µPJ , σJP


.
• Affine risk premiums for the diffusive risks and otherwise constant
risk premiums leads to
 Q 
  N t
X
drt = aQ Q r
r − br rt dt + σr dWt (Q) + d
 ZjQ 
j=1

is a standard Brownian motion, NtQ has intensity λQ ,

• Here Wtr (Q) 
 2
and Zj ∼ N µJ , σJQ
Q Q

12
 Portfolio Choice
• How to allocate assets between risky assets?
1. Mean-Variance Analysis
de Finetti (1941) “On a Problem of Re-insurance”
Markowitz (1951) “Portfolio Selection”
2. Simple Case: Cash or stocks?
dSt
Samuelson (1969), Merton (1971): St = µdt + σdWtP
Allocation to stocks
1µ−r
ω? =
γ σ2
γ is individuals risk aversion
How do we pick risk aversion? What’s a reasonable value?

• General Case: µt and Vt and hedging demands (JPS, 2002).

13
 Historical Return Data
• Here the returns for US stocks etc
Period Stocks Bonds Bills Gold Inflation
1802-1998 7.0 3.5 2.9 - 0.1 1.3
1802-1870 7.0 4.8 5.1 0.2 0.1
1871-1925 6.6 3.7 3.2 -0.8 0.6
1926-1998 7.4 2.2 0.7 0.2 3.1
1946-1998 7.8 1.3 0.6 -0.7 4.2
• Which portfolio allocation do you choose?

• What does the Samuelson/Merton rule yield?

can we make this dynamic?

14
SP500 Data

• Market Timing: time-varying mean-reverting risk premia model

No real gains

• Volatility Timing: SV and leverage

Strategy Mean SD Sharpe
S&P500 13.6 14.9 0.49
SV: γ = 2 16.1 17 0.53
SV: γ = 4 11.2 10.2 0.68
SV: γ = 6 9.6 7.1 0.94

15
(*E-5) αµ αv
10 0.020
0.015
9
0.010
•• •• •••••••••• 0.005 ••• •
8 •••••••••••••••• ••• •• • ••••••••• • • •• • •••••••••••••••••••••••••••••••••••••••••••• • ••
• • •• • • ••••• •••••••••••••••••••••••••• • •
••
0.0 ••
7 ••••
••••••••••••• • •••••••••••••••••••••••• • ••• ••
-0.005 •• •••• • ••••• ••••••••• •••••
•••• ••••• •••••••••••••••••••• •••••••• ••••••••••••
• • •
6 -0.010

1976 1980 1984 1988 1992 1996 2000 1976 1980 1984 1988 1992 1996 2000

(*E-4) σµ βv
1.00
30
•••••
0.99 •• •••••••••••••••••••••••••••••••••• •••••••
• ••• ••••••••
•• ••••••••••• ••••• • ••••••••••••••••••••••••••••••
• • •• •• ••••••• ••••
20 0.98 ••• • •
• ••• •
••• • ••
• ••••••••••••••••• •• • ••••• •• 0.97
• • •• • ••• ••••••••• • ••••••••••••• •• • • • • • •
10 •• •• • • • • ••••• • ••• ••••••••• •••••••••••••••••• •••••••••••••••
•
• 0.96

0 0.95

1976 1980 1984 1988 1992 1996 2000 1976 1980 1984 1988 1992 1996 2000

ρ σv
0.5 0.18
0.16
0.0 •• •
• • ••• • • •• • •• • • • • • • • 0.14 •
• • • • • •••• ••••• • •••• • •••• • ••• •••••• • •••• ••• ••• ••••
•••• ••••••• • • •••••••••••• • • • •• • • ••
• • • •• •• • •••• •• ••• ••• 0.12 ••••• ••••••••••• •• ••••• • •••••• • • ••••••••••••••••••••••
••• • •• • •••• • • •• ••••••••••• • • • •• •
-0.5 ••••
0.10 •
•••• •••• • ••••••• • • •••••• • •••••
••• • • • •• •• • ••••••••••••
0.08
-1.0
0.06
1976 1980 1984 1988 1992 1996 2000 1976 1980 1984 1988 1992 1996 2000

16
 Wealth - SP500 - fixed v - parameter learning
gamma = 2 gamma = 4 gamma = 6
4 4 4
Market Market Market
Leverage Leverage Leverage
3 No Leverage 3 No Leverage 3 No Leverage

2 2 2

1 1 1

0 0 0

1974 1979 1984 1989 1994 1999 1974 1979 1984 1989 1994 1999 1974 1979 1984 1989 1994 1999

2.0 2.0 2.0

1.5 1.5 1.5

1.0 1.0 1.0

0.5 0.5 0.5

0.0 0.0 0.0

1974 1979 1984 1989 1994 1999 1974 1979 1984 1989 1994 1999 1974 1979 1984 1989 1994 1999

17
 Portfolio Selection
Polson and Tew (2000) Bayesian Portfolio Selection for the SP500 index:
1970-1996.
• Markowitz: Mean-variance portfolio
1
min xT Σx − xT µ − λc(x)
x 2

• Bayes

1 T  xT ι = 1
min x V ar(Rt+1 |Dt )x subject to
x 2  xT E(Rt+1 |Dt ) = µP

• Hierarchical Model
(tR̄ + λµ0 )
E(Rt+1 |Dt ) =
t+λ
 
ΣP = C Σ̂ + S0 + λµ0 µT0 0
+ tR̄R̄ − (λ + t)E(Rt+1 |Dt )E(Rt+1 |Dt ) T

 
1 ν0 +t
where C = 1 + t+λ (ν0 +t−2)(t−k−2) .

18
• Different shrinkage factors for different history lengths.

• Portfolio Allocation in the SP500 index

Entry/exit; splits; spin-offs etc. For example, 73 replacements to the
SP500 index in period 1/1/94 to 12/31/96.

• Advantage?: E(α|Dt ) = 0.39, that is 39 bps per month which on an

annual basis is 468 bps. The posterior mean for β is p(β|Dt ) = 0.745

• x̄M = 12.25% and x̄P T = 14.05%.

19
 Date Symbol 6/96 12/89 12/79 12/69
General Electric GE 2.800 2.485 1.640 1.569
Coca Cola KO 2.342 1.126 0.606 1.051
Exxon XON 2.142 2.672 3.439 2.957
ATT T 2.030 2.090 5.197 5.948
Philip Morris MO 1.678 1.649 0.637 *****
Royal Dutch RD 1.636 1.774 1.191 *****
Merck MRK 1.615 1.308 0.773 0.906
Microsoft MSFT 1.436 ***** ***** *****
Johnson/Johnson JNJ 1.320 0.845 0.689 *****
Intel INTC 1.262 ***** ***** *****
Procter and Gamble PG 1.228 1.040 0.871 0.993
Walmart WMT 1.208 1.084 ***** *****
IBM IBM 1.181 2.327 5.341 9.231
Hewlett Packard HWP 1.105 0.477 0.497 *****
Pepsi PEP 1.061 0.719 ***** *****
Pfizer PFE 0.918 0.491 0.408 0.486
Dupont DD 0.910 1.229 0.837 1.101
AIG AIG 0.910 0.723 ***** *****
Mobil MOB 0.906 1.093 1.659 1.040
Bristol Myers Squibb BMY 0.878 1.247 ***** 0.484
GTE GTE 0.849 0.975 0.593 0.705
General Motors GM 0.848 1.086 2.079 4.399
Disney DIS 0.839 0.644 ***** *****
Citicorp CCI 0.831 0.400 0.418 *****
BellSouth BLS 0.822 1.190 ***** *****
Motorola MOT 0.804 ***** ***** *****

20
 Ford F 0.798 0.883 0.485 0.640
Chervon CHV 0.794 0.990 1.370 0.966
Amoco AN 0.733 1.198 1.673 0.758
Eli Lilly LLY 0.720 0.814 ***** *****
Abbott Labs ABT 0.690 0.654 ***** *****
AmerHome Products AHP 0.686 0.716 0.606 0.793
FedNatlMortgage FNM 0.686 ***** ***** *****
McDonald’s MCD 0.686 0.545 ***** *****
Ameritech AIT 0.639 0.782 ***** *****
Cisco Systems CSCO 0.633 ***** ***** *****
CMB CMB 0.621 ***** ***** *****
SBC SBC 0.612 0.819 ***** *****
Boeing BA 0.598 0.584 0.462 *****
MMM MMM 0.581 0.762 0.838 1.331
BankAmerica BAC 0.560 ***** 0.577 *****
Bell Atlantic BEL 0.556 0.946 ***** *****
Gillette G 0.535 ***** ***** *****
Kodak EK 0.524 0.570 1.106 *****
Chrysler C 0.507 ***** ***** 0.367
Home Depot HD 0.497 ***** ***** *****
Colgate COL 0.489 0.499 ***** *****
Wells Fargo WFC 0.478 ***** ***** *****
Nations Bank NB 0.453 ***** ***** *****
Amer Express AXP 0.450 0.621 ***** *****

21
 2: MCMC Financial Econometrics
Wide range of applications for estimation, filtering and state variable
extraction have been developed.

• Johannes and Polson (2004): MCMC Financial Econometrics

Handbook of Financial Econometrics provide an overview.

• Filtering Methods (JPS, 2004)

• Applications
1. Stochastic Volatility
2. Jump Diffusions

22
Target Distribution p(Θ, X|Y )

• Bayes rule
p(Θ, X|Y ) ∝ p(Y |X, Θ)p(X|Θ)p(Θ)
where p(X|Θ) gives state dynamics
X = (V, J) volatility and jumps.

• Estimation Risk: Effect of dynamic learning about Θ and pricing

• Pricing: allows you to address pricing issues

• Model Risk: MCMC diagnostics for model mis-specification.

23
Building MCMC Algorithms
p (Θ|X, Y ) and p (X|Θ, Y ) uniquely determine p (Θ, X|Y ).

• Gibbs Sampling
Merton jump-diffusion models, regime switching, . . .

• Metropolis-Hastings
1. Nonlinear State Space Models
Volatility and/or Jumps
2. Parameter inference
leverage ρ-correlation between returns and volatility (JPR ’04).
3. FFBS (CK ’94) for block simulation of p(X|Θ, Y )

24
Option Pricing

• Volatility Puzzle
Why are implied volatilities typically higher than historical
volatilities?
VIX Option-based implied volatility

• SV model explains this using

Time-varying Vt , Leverage effect ρ and Market price of risk λv

• Key: Inference depends on both Derivative and Asset price

information.
Output V = {Vt }Tt=1 and Parameters Θ for option pricing

25
 50
Implied
Historical
40
Volatility (%)

30
26

20

10

1990 1992 1994 1996 1998 2000

Year
Historical and Implied Vols
There are also two types of financial volatilities

1. Historical Volatility: these are volatility estimates arrived at from

looking at the historical path of prices and using a model (maybe
time-varying) to estimate the future path of volatility

2. Implied Volatility: these come from exchange based market measures

explaining the market’s current perception about what average future
volatility will look like. VIX and VXN indices for the S&P500 and
nasdaq indicies, respectively.

27
Option Pricing
Black-Scholes

EQ


• Option Price: e−r(T −t) f (S T )|St , Θ for pay-out f (ST ).

• Risk-Neutral Dynamics Q
dSt
= rdt + σdWtQ
St
Interest rate r.

• Black-Scholes:
Ct = St P1 (σ) − Ke−r(T −t) P2 (σ)
for given probabilities Pi (σ).
Strike price K, current stock price St .

28
Option Pricing: SV and Jumps

• Stochastic Volatility: Heston model (1993)

Ct = St P1v (Vt , Θ) − Ke−r(T −t) P2v (Vt , Θ)

• Jumps: Duffie, Pan and Singleton (2000)

Ct = St P1J (Vt , Θ) − Ke−r(T −t) P2J (Vt , Θ)

“Same” formula.
Caveat: dramatically different prices Ct depending on parameters and
state variables

• MCMC Advantages
Can filter Vt , Jt and learn p(Θ|datat ) dynamically

29
 Application
• PS (2004) “Bayesian inference for Derivative Prices” use option and
stock prices data for the SP500.

• Remember the square-root option pricing model

dSt p
= µdt + Vt dWts
St
p
dVt = κv (θv − Vt ) + σv Vt dWtv

The parameter σv governs the volatility of volatility

• Now we can incorporate the information about parameters that is

also contained in option prices.
This will introduce a market price of volatility risk
Need the risk-neutral dynamics

30
• Heston (1993) considers pricing a call option with stochastic volatility
and leverage:
D Q −rτ


Yt = E e (ST − K)+ |Vt , Θ, Λ

• Risk-neutral dynamics (E Q )
dSt p
= rdt + Vt dWts
St
p
dVt = κ (θ − Vt ) + σv Vt dWtv
∗ ∗

• Leverage effect ρ = corr (dWts , dWtv ) (negative)

• Market price of risk λv adjusts drift (κ, θ) → (κ∗ , θ∗ ):

κ
κ∗ = κ + λ v and θ∗ = θ
κ + λv

31
Full Conditionals:

1. Asset price drift: p(µ|rest) (Gibbs)

2. Pricing variance: 2 |rest)
p(σD (Gibbs)
3. Volatility states: p({Vt }Tt=1 |rest) (Single-state Metropolis)
4. Volatility drift: p(κ, θ|rest) (Random Walk Metropolis)
5. Leverage, vol of vol: p(ρ, σv |rest) (Random Walk Metropolis)
6. Market price of risk: p(λv |rest) (Random Walk Metropolis)

32
Posterior Estimates (Annualized)

Parameter Y S only Y S and Y D

σD = 5% σD unknown
√
E( Vt ) 21.60% 21.90% 21.82%
(1.46) (1.15) (0.86)
κ 0.045 0.046 0.055
(0.012) (0.005) (0.005)
σv 0.047 0.048 0.055
(0.004) (0.002) (0.001)
ρ -0.70 -0.78 -0.74
(0.08) (0.02) (0.01)
λv – -0.010 -0.017
– (0.005) (0.005)

33
Option Pricing Implications
SV Jump Equity Models
DPS (2000) and EJP (2003): double jump model
 
Nt
X
p s
d log (St ) = µdt + Vt− dWt + d  ξτsj 
j=1
 
Nt
X
p
dVt = κv (θv − Vt ) dt + σv Vt− dWtv + d  ξτvj 
j=1

where ξ s = µs + ρJ ξ v + σs ε, ε ∼ N (0, 1), ξ v ∼ exp (µv ) and Nt ∼ P oi(λt).

• Stochastic Volatility (SV) with λ = 0

• Jumps in returns (SVJ) with µv = 0

• Double Jump model (SVCJ) with possibly correlated jumps between

volatility and returns.

34
Data: Two periods of market stress: October 20, 1987 and October 27,
1997. Typical day January 20, 1988.

• Filtered Volatilities to compute Black-Scholes implied volatilities for

the three models SV, SVJ, SVCJ

• Findings
1. SV can only attribute moves to volatility which cannot move fast.
2. SVCJ uses a jump in returns and a large contemporaneous jump
in volatility. Hence better around a crash.
Volatility Curve? High spot volatility and a flat implied.

35
 Option Pricing
October 20, 1987 January 20, 1988
0.8 0.4
SV SV
0.7 SVJ SVJ
SVCJ 0.35 SVCJ
0.6
Implied Volatility

0.5 0.3

0.4 0.25
0.3
0.2
0.2

0.1 0.15
0.8 0.9 1 1.1 1.2 0.8 0.9 1 1.1 1.2

October 27, 1997 Average Volatility

0.5
SV SV
0.45 SVJ SVJ
SVCJ SVCJ
0.4 0.25
Implied Volatility

0.35

0.3 0.2

0.25

0.2 0.15

0.15
0.8 0.9 1 1.1 1.2 0.85 0.9 0.95 1 1.05 1.1 1.15

36
 October 20, 1987 January 20, 1988
0.8 0.4
SV SV
0.7 SVJ SVJ
SVCJ 0.35 SVCJ
0.6
Implied Volatility

0.5 0.3

0.4 0.25
0.3
0.2
0.2

0.1 0.15
0.8 0.9 1 1.1 1.2 0.8 0.9 1 1.1 1.2
37

October 27, 1997 Average Volatility

0.5
SV SV
SVJ 0.28 SVJ
0.45
SVCJ SVCJ
0.26
0.4
Implied Volatility

0.24
0.35 0.22

0.3 0.2
0.18
0.25
0.16
0.2
0.14
0.15
0.8 0.9 1 1.1 1.2 0.9 1 1.1
 Discussion

• MCMC useful for practical implementation of finance models

• Nonlinear state dynamics, sharper parameter estimates and model

mis-specification diagnostics

• Implications for financial decision making

• Further issues: Estimation Risk, Sequential Parameter Learning and

Pricing

38