Term Structures of Asset Prices and Returns
David Backus (NYU), Nina Boyarchenko (NY Fed∗ ), & Mikhail Chernov (UCLA) (With thanks to Ian Martin and Stan Zin)
Stockholm School of Economics | June 4, 2015
∗ Fed
disclaimer | June 1, 2015
Look for entropy
Paul Samuelson (“Gibbs in economics,” 1989): I have limited tolerance for the perpetual attempts to fabricate for economics concepts of “entropy.”
1 / 58
Look for logarithms I
Think about log mt,t+1 rather than mt,t+1
I
Sums more user-friendly than products log(mt,t+1 rt,t+1 ) = log mt,t+1 + log rt,t+1 more user-friendly than mt,t+1 rt,t+1 log(mt,t+1 mt+1,t+2 ) = log mt,t+1 + log mt+1,t+2 more user-friendly than mt,t+1 mt+1,t+2
2 / 58
Excess returns
1 Monthly excess log returns in dollars: log rt,t+1 − log rt,t+1
Asset S&P 500 Fama-French (small, low) Fama-French (small, high) Pound Sterling 5-year bond 10-year bond
Mean 0.0040 −0.0030 0.0090 0.0035 0.0015 0.0019
Standard Deviation
Skewness
Excess Kurtosis
0.0556 0.1140 0.0894 0.0316 0.0190
−0.40 0.28 1.00 −0.50 0.10
7.90 9.40 12.80 1.50 4.87
3 / 58
Term structure data: US nominal
Mean Yields and Forward Rates
7.5 7.0 forwards 6.5 6.0
yields
5.5 5.0
0
20
40
60 80 Maturity in Months
100
120
4 / 58
Mean Yield Difference
Mean Yields
Term structure data: US nominal and real 5 4 3 2 1 0 1 2 0 4.0 3.5 3.0 2.5 2.0 1.5 0
nominal real 20
40
60
80
100
120
60 80 Maturity in Months
100
120
nominal minus real 20
40
5 / 58
Term structure data: other assets relative to US nominal −3
2
x 10
1.5 1
Risk premium
0.5 0 −0.5 −1 −1.5 SEK NZD Inflation Dividends
−2 −2.5 −3 0
20
40
60 Horizon
80
100
120
6 / 58
Where we’re headed
What makes these term structures different? Plan of attack I
Entropy: dispersion in the pricing kernel
I
Risk premiums: entropy bound
I
Term structure: pricing kernel dynamics
I
Coentropy: risk premiums revisited
I
Other term structures: cash flow dynamics, more coentropy
7 / 58
Entropy
8 / 58
Entropy I
Entropy is a measure of dispersion: for rv x > 0 L(x) ≡ log E (x) − E (log x) ≥ 0
I
Invariant to scale: L(αx) = L(x) for α > 0
I
Lognormal example: if log x ∼ N (κ1 , κ2 ), then log E (x) = κ1 + κ2 /2 E (log x) = κ1 L(x) = (κ1 + κ2 /2) − κ1 = κ2 /2
9 / 58
Cumulants I
Cumulant generating function (cgf) of rv y ∞ X k(s; y ) = log E e sy = κj s j /j! | {z } j=1 mgf
I
Cumulants are close relatives of moments mean = κ1 variance = κ2 3/2
skewness = κ3 /κ2
excess kurtosis = κ4 /κ22 I
If y is normal: k(s; y ) = κ1 s + κ2 s 2 /2
10 / 58
Entropy and cumulants I
Cumulants of y = log x k(s; log x) = log E (e
s log x
) =
∞ X
κj (log x)s j /j!
j=1 I
Entropy and cumulants (set s = 1) L(x) = k(1; log x) − E (log x) =
∞ X
κj (log x)/j!
j=2
=
κ2 (log x)/2! + κ3 (log x)/3! + κ4 (log x)/4! + · · · | {z } | {z }
(log)normal term
high-order cumulants
11 / 58
Entropy of a stationary stochastic process
I
Conditional entropy defined for conditional distribution Lt (xt+1 ) = log Et (xt+1 ) − Et (log xt+1 )
I
We define entropy as the mean E [Lt (xt+1 )]
I
Connected to entropy for unconditional distribution L(xt+1 ) = E [Lt (xt+1 )] +L[Et (xt+1 )] {z } | entropy
12 / 58
Risk premiums: the entropy bound
13 / 58
Entropy bound (Alvarez-Jermann) I
Returns satisfy the pricing relation Et (mt,t+1 rt,t+1 ) = 1
I
1 Entropy bound: maximize Et (log rt,t+1 − log rt,t+1 )
I
Maximization leads to the bound risk premium?
z }| { 1 Et (log rt,t+1 − log rt,t+1 ) ≤ Lt (mt,t+1 ) 1 E (log rt,t+1 − log rt,t+1 ) ≤ E [Lt (mt,t+1 )] | {z } entropy
I
High return is log rt,t+1 = − log mt,t+1
14 / 58
Hansen-Jagannathan bound
I
HJ bound: maximize Sharpe ratio
I
Maximization leads to the bound SRt
1 1 ≡ Et (rt,t+1 − rt,t+1 )/Vart (rt,t+1 − rt,t+1 )1/2
≤ Vart (mt,t+1 )1/2 /Et (mt,t+1 ) I
High return is rt,t+1 =
1 + Vart (mt,t+1 )1/2 mt,t+1 − Et (mt,t+1 ) − Et (mt,t+1 ) Vart (mt,t+1 )1/2
15 / 58
Stan’s “never a dull moment” machine I
Entropy of pricing kernel L(m) = log E e log m − E (log m) = k(1; log m) − E (log m) =
∞ X
κj (log m)/j!
j=2 I
Stan’s entropy machine (but ask about Lukacs) L(m) =
κ2 (log m)/2! + κ3 (log m)/3! + κ4 (log m)/4! + · · · {z } | {z } |
(log)normal term I
high-order cumulants
Kraus and Litzenberger revisited?
16 / 58
Why is this entropy? I
Humpy Dumpty (in “Through the Looking Glass”) “When I use a word,” Humpty Dumpty said, “it means just what I choose it to mean — neither more nor less.”
I
Hans-Otto Georgii (quoted by Hansen and Sargent): When Shannon had invented his quantity and consulted von Neumann on what to call it, von Neumann replied: “Call it entropy. It is already in use under that name and, besides, it will give you a great edge in debates because nobody knows what entropy is anyway.”
17 / 58
Why is this entropy? I
Notation: states z have (true) probabilities π(z)
I
Risk-neutral probabilities π ∗ π ∗ (z) = π(z)m(z)/p 1 m(z) = p 1 π ∗ (z)/π(z) p 1 = E (m) (1-period bond price)
I
Entropy (aka “relative entropy” or “Kullback-Leibler divergence”) L(m) = L(π ∗ /π) = E log(π/π ∗ ) (π ∗ = π ⇒ L(m) = 0, risk premiums = 0)
18 / 58
Vasicek model I
Pricing kernel log mt,t+1 = log β + xt + λwt+1 with {wt } iid, mean zero, variance one, and cgf k(s)
I
Conditional entropy Lt (mt,t+1 ) = k(λ) = λ2 κ2 /2! + λ3 κ3 /3! + λ4 κ4 /4! + · · ·
I
Entropy: the same (maximum risk premium is constant)
19 / 58
State-dependent price of risk I
Pricing kernel log mt,t+1 = log β + xt + (λ0 + λ1 xt )wt+1 with {wt } ∼ NID(0, 1), k(s) = s 2 /2
I
Conditional entropy Lt (mt,t+1 ) = k(λ0 + λ1 xt ) = (λ0 + λ1 xt )2 /2
I
Entropy E [Lt (mt,t+1 )] = E [(λ0 + λ1 xt )2 /2]
20 / 58
Power utility I
Consumption growth gt,t+1 = ct+1 /ct iid
I
Pricing kernel log mt,t+1 = log β − α log gt,t+1 (Vasicek with xt = 0, λ = −α, and wt+1 = log gt,t+1 )
I
Yaron’s bazooka Lt (mt,t+1 ) = k(−α) = (−α)2 κ2 /2! + (−α)3 κ3 /3! + (−α)4 κ4 /4! + · · ·
21 / 58
Power utility: the bazooka −3
Cumulant
2
x 10
0 −2
2
3
4
5
6
7
8
Contribution
Contribution
−3
4
x 10
risk aversion α=2
2 0
2
3
4
5
6
0.1
7
8
risk aversion α=10
0.05 0
2
3
4
5 Order j
6
7
8
22 / 58
Power utility: the bazooka 0.4
Entropy of Pricing Kernel L(m)
0.35 0.3 0.25 disasters
0.2 0.15 0.1 0.05 0 0
normal
equity premium
booms 2
4
6 Risk Aversion α
8
10
12
23 / 58
Term structure: pricing kernel dynamics
24 / 58
The idea
I
In an iid world, entropy is proportional to time interval
I
Deviations from proportionality reflect pricing kernel dynamics
I
Detectable from mean yields and forward rates
25 / 58
Bond prices, yields, and forward rates I
Bond price: ptn is price at t of a claim to one (dollar?) at t + n
I
Bond yield: ytn = −n−1 log ptn
I
Forward rate: ftn = log(ptn /ptn+1 ) ⇒ ytn =
I
One-period return:
j−1 j=1 ft
Pn
n+1 n log rt,t+1 = log(pt+1 /ptn+1 ) ⇒ E (log r n+1 ) = E (f n )
I
Cross sections reflect pricing kernel dynamics
26 / 58
Bond pricing fundamentals I
Markov environment with state variable x
I
Bond pricing is recursive p n (xt ) = Et m(xt , xt+1 )p n−1 (xt+1 ) starting with p 0 = 1
I
Equivalent to p n (xt ) = Et m(xt , xt+1 )m(xt+1 , xt+2 ) · · · m(xt+n−1 , xt+n )
I
Definitions give us yields y n (xt ) and forward rates f n (xt )
27 / 58
Entropy and the term structure I
Entropy over n periods mt,t+n = mt,t+1 mt+1,t+2 · · · mt+n−1,t+n Lt (mt,t+n ) = log Et (mt,t+n ) −Et (log mt,t+n ) {z } | log ptn =−nytn
L(n) ≡ E [Lt (mt,t+n )] = −nE (y n ) − nE (log mt,t+1 ) I
Two measures of horizon dependence H(n) ≡
n−1 L(n) | {z }
avg over n periods
− L(1) |{z}
= −E (ytn − yt1 )
one period
F (n) ≡ L(n + 1) − L(n) − L(1) = −E (ftn − ft0 )
28 / 58
Entropy and the term structure
I
The iid benchmark: {mt,t+1 } iid ⇒ L(n) = n L(1) H(n) = 0 F (n) = 0
I
Also: yields and forwards constant, same at all maturities
I
Anything different from this reflects dynamics in m
29 / 58
Vasicek model: dynamic structure I
Pricing kernel log mt,t+1 = log β + xt + λwt+1 xt+1 = ϕxt + σwt+1 x is (persistent or long-run) risk, λ is price of risk
I
Moving average representation log mt,t+1 = log β + xt + λwt+1 = log β + λwt+1 + σwt + σϕwt−1 + · · · | {z } xt
30 / 58
Vasicek model: entropy
I
Pricing kernel dynamics inherited from x
I
Term structure of entropy L(1) = k(λ) L(2) = k(λ) + k(λ + σ) L(3) = k(λ) + k(λ + σ) + k(λ + σ + σϕ)
I
What makes this non-iid?
31 / 58
Vasicek model: parameter values I
Short rate yt1 = ft0 = − log Et (mt,t+1 ) = −[log β + k(λ)] − xt
I
I
Choose I
(ϕ, σ) = (0.98, −0.006) match variance and autocorrelation
I
w normal ⇒ k(s) = s 2 /2
I
λ = 0.088 matches mean forward spread E (f n − f 0 ) (ask how this works)
Features I
σ and λ must have opposite signs for curve to slope up
I
λ much greater than σ in absolute value
32 / 58
Vasicek model: mean forward spreads 2.5 Vasicek model
Mean Forward Spread
2.0 1.5 1.0 0.5 0.0
0
dots are data 20
40
60 80 Maturity in Months
100
120
33 / 58
Vasicek model: moving average coefficients 0.003 Moving Average Coefficients
= 0.088 0.002 0.001 0.000 0.001
0
2
4 Lag
6
8
34 / 58
Vasicek model: entropy 0.5
Entropy
0.4 0.3 iid benchmark 0.2 Vasicek model
0.1 0.0
0
20
40 60 80 Time Horizon in Months
100
120
35 / 58
Coentropy: risk premiums revisited
36 / 58
The idea
I
Expected excess returns differ across assets
I
Reflects dependence of pricing kernel and cash flows
I
We measure dependence with coentropy
I
Extend shortly to long time horizons
37 / 58
Coentropy I
Coentropy is a measure of dependence: for x1 , x2 > 0 C (x1 , x2 ) ≡ L(x1 x2 ) − L(x1 ) − L(x2 )
I
I
Features I
Invariant to scaling
I
Equals zero if x1 and x2 are independent
Related to (joint) cgf k(s1 , s2 ) = log E (e s1 log x1 +s2 log s2 ) C (x1 , x2 ) = k(1, 1) − k(1, 0) − k(0, 1) | {z } | {z } | {z } x1 x2
x1
x2
38 / 58
Coentropy (continued) I
If log x = (log x1 , log x2 ) is normal, coentropy = covariance
I
Can also be much different
I
Example: Poisson mixture (“jump process”)
I
I
Poisson jumps: probability e −ω ω j /j! of j = 0, 1, 2, . . ..
I
Conditional on j, log x ∼ N (jθ, j∆)
Properties Cov(log x1 , log x2 ) = ω(θ1 θ2 + δ12 ) C(x1 , x2 ) = E2C2E
39 / 58
Coentropy and covariance 1.8 Coentropy and Covariance
1.6 1.4 1.2 1.0 0.8 covariance
0.6 0.4
coentropy
0.2 0.0
0
1
2 3 Poisson intensity ω
4
5
40 / 58
Coentropy and excess returns I
Consider a claim to the cash flow gt,t+1
I
Return is cash flow over price
I
Invariance to scaling implies Lt (rt,t+1 ) = Lt (gt,t+1 ) Ct (mt,t+1 , rt,t+1 ) = Ct (mt,t+1 , gt,t+1 )
I
Expected excess returns (“risk premiums”)
1 Et (log rt,t+1 − log rt,t+1 ) = −Lt (gt,t+1 ) − Ct (mt,t+1 , gt,t+1 ) 1 E (log rt,t+1 − log rt,t+1 ) = − E [Lt (gt,t+1 )] − E [Ct (mt,t+1 , gt,t+1 )] {z } | {z } | entropy
coentropy
41 / 58
KLV model (streamlined version) I
Add another disturbance to Vasicek log mt,t+1 = log β + xt + λ1 w1t+1 + λ2 w2t+1 xt+1 = ϕxt + σw1t+1 (w1t , w2t ) ∼ NID(0, I )
I
Stir in some cash flow growth log gt,t+1 = log γ + θxt + η1 w1t+1 + η2 w2t+1
I
Entropy and coentropy E [Lt (mt,t+1 )] = (λ21 + λ22 )/2 E [Lt (gt,t+1 )] = (η12 + η22 )/2 E [Ct (mt,t+1 , gt,t+1 )] = λ1 η1 + λ2 η2 42 / 58
KLV model: numerical example I
Choose (ϕ, σ, λ1 ) = (0.98, −0.0006, 0.088) as before to fit yields/forwards
I
Choose (η1 , η2 ) = (−0.005, −0.050) to match I
Variance of excess return on equity (0.05)
I
Correlation of excess returns on equity and bonds (0.1)
I
Choose λ2 = 0.097 to match equity premium
I
Results (monthly): I
Bond premium (10 years): 0.002
I
Equity premium: 0.004 = 0.005 (coentropy) – 0.001 (entropy)
I
Entropy of m: 0.009 (upper bound)
43 / 58
Other term structures: cash flow dynamics
44 / 58
The idea
I
Consider claims to currencies, equity indexes, dividends, ...
I
How do their term structures compare?
I
The time horizon of coentropy
I
Explorations with the KLV model
45 / 58
Prices, yields, and forward rates I
Let pbtn be price at t of claim to cash flow growth gt,t+n
I
Term structure ybtn = −n−1 log pbtn fbtn = log(b ptn /b ptn+1 ) n+1 n log rbt,t+1 = log pbt+1 − log pbtn+1 + log gt,t+1
E (log rbn+1 ) = E (f n + log g ) I
Forward price qtn = pbtn /ptn
⇒
n−1 log qtn = ytn − ybtn log q n+1 − log q n = f n − fbn t
t
t
t
46 / 58
Pricing fundamentals I
Pricing is recursive pbn (xt ) = Et m(xt , xt+1 )g (xt , xt+1 )b p n−1 (xt+1 ) b t , xt+1 )b = Et m(x p n−1 (xt+1 ) b t , xt+1 ) = m(xt , xt+1 )g (xt , xt+1 ) with pb0 = 1 and m(x
I
Definitions give us yields ybn (xt ) and forward rates fbn (xt )
I
Think of this as a change of units: dollars to yen
I
Empirical strategy changes: we observe g
47 / 58
Entropy, coentropy, and term structures
I
Entropy and coentropy b t,t+n ) = Lt (mt,t+n gt,t+n ) Lt (m = Ct (mt,t+n , gt,t+n ) + Lt (mt,t+n ) + Lt (gt,t+n ) Lmb (n) = Lm (n) + E [Ct (mt,t+n , gt,t+n )] + E [Lt (gt,t+n )] = Lm (n) + Cmg (n) + Lg (n) | {z } | {z } coentropy
I
entropy
Same connection to mean yields and forward rates as before
48 / 58
Term structure data: other assets relative to US nominal −3
2
x 10
1.5 1
Risk premium
0.5 0 −0.5 −1 −1.5 SEK NZD Inflation Dividends
−2 −2.5 −3 0
20
40
60 Horizon
80
100
120
49 / 58
KLV model I
Model log mt,t+1 = log β + xt + λ1 w1t+1 + λ2 w2t+1 xt+1 = ϕxt + σw1t+1 log gt,t+1 = log γ + θxt + η1 w1t+1 + η2 w2t+1
I
Transformed pricing kernel long-run risk
b t,t+1 log m
z }| { = (log β + log γ) + (1 + θ) xt + (λ1 + η1 ) w1t+1 + (λ2 + η2 ) w2t+1 | {z } | {z } price of lr risk
I
price of iid risk
Roles of: η1 , λ2 , η2 , θ
50 / 58
KLV model: currencies I
Model log mt,t+1 = log β + xt + λ1 w1t+1 + λ2 w2t+1 xt+1 = ϕxt + σw1t+1 log gt,t+1 = log γ + θxt + η1 w1t+1 + η2 w2t+1
I
Currencies: θ ≈ 0, η12 + η22 ≈ 0.033 , λ2 ≈ 0 (for now)
I
Then (η1 , η2 ) control coentropy I
(η1 , η2 ) = (0, 0.3): coentropy is zero
I
(η1 , η2 ) = (0.3, 0): coentropy is positive
I
(η1 , η2 ) = (−0.3, 0): coentropy is negative
51 / 58
KLV model: currencies 10 9 Mean Forward Rates
8
negative coentropy
USD benchmark
7 6 5
zero coentropy
4 3 2 1 0
positive coentropy 20
40
60 80 Maturity in Months
100
120
52 / 58
KLV model: currencies 3.5
Mean Forward Spreads
3.0 2.5
negative coentropy
USD benchmark
2.0 1.5 1.0
positive coentropy
0.5 0.0
0
20
40
60 80 Maturity in Months
100
120
53 / 58
KLV model: equity I
Model log mt,t+1 = log β + xt + λ1 w1t+1 + λ2 w2t+1 xt+1 = ϕxt + σw1t+1 log gt,t+1 = log γ + θxt + η1 w1t+1 + η2 w2t+1
I
Reminder: transformed pricing kernel long-run risk
b t,t+1 log m
z }| { = (log β + log γ) + (1 + θ) xt + (λ1 + η1 ) w1t+1 + (λ2 + η2 ) w2t+1 | {z } | {z } price of lr risk
I
price of iid risk
Roles of: λ2 + η2 , η1 (−0.005), θ (±0.25)
54 / 58
KLV model: equity 3.0
Mean Forward Spreads
2.5 theta = 0.25
2.0
theta = 0
1.5 theta = -0.25
1.0 0.5 0.0
0
20
40
60 80 Maturity in Months
100
120
55 / 58
KLV model: equity
Mean Forward Differentials
4.8 theta = 0.25
5.0 5.2
theta = 0
5.4 5.6 5.8 0
theta = -0.25 20
40
60 80 Maturity in Months
100
120
56 / 58
Where were we?
57 / 58
Summary and open questions
Summary I
Significant variation in average term structures across assets
I
Connected to entropy and coentropy
I
Large quantitative effects in simple models (still no bazooka!)
Open questions I
What would you do with these ingredients?
I
Would predictability interest you?
58 / 58