apoorva javadekar's - comments on lewellen shanken
TRANSCRIPT
Learning, Asset Pricing Testsand Market Efficiency
Apoorva JavadekarPresented for EC794
Boston University
May 2, 2012
() Learning & Asset Prices May 2, 2012 1 / 22
Motivation
Large literature on stock returns predictability
D/P ratio (Campbell, Cochrane)Consumption - Wealth Ratio (Latteu, Ludvigson)Aggregate Market Liquidity (Adrian-Shin)
Main Concern- Predictability in the data 6= predictability by investors
Why? Parameter uncertainty: Agents learn about nature ofuncertainty over time and prices react to learning
Learning ⇒ future returns are a function of past prices and dividends⇒ Predictability observed in the data
But, this predictability can not be exploited by investor: He is unsureif he overestimates or underestimates the parameters
() Learning & Asset Prices May 2, 2012 2 / 22
Motivating Example
suppose dividends are i.i.d over time with mean δ and variance σ2
known variance but unknown mean
After each realized dividend dt , investor updates his belief aboutmean δ.
Price of a stock is an increasing function of belief about δ
With i.i.d nature, best guess δ̃t = dt at t
If dt > δ, then price > fundamental value.
But by i.i.d nature, average dividends in the future are lower than δ⇒ negative price change in future accompanied by lower returns
Result: Higher prices predict lower returns in future and it isobserved in the data. But this predictability can not be exploitedby investor. (He is simply not sure if observed dividend is aboveor below true mean)
() Learning & Asset Prices May 2, 2012 3 / 22
Few Early Remarks
Predictability arises solely from learning (Not from persistence inthe dividend process)
Predictability in data - But it can’t be exploited by investor
Predictable equilibrium returns even when investors are rational
Prices are functions of beliefs at any point in time. Beliefs areinfluenced by true DGP in long run and hence prices converge tofundamental value in LR (only if parameters are time invariant)
Conclusion: Parameter Uncertainty creates wedge betweendistribution of returns in data and distribution perceived by investors
() Learning & Asset Prices May 2, 2012 4 / 22
Model - Set - up and Pricing
Riskless asset paying net return of r every period
N risky assets with i.i.d joint dividend process given by
dt v MVN[δ,Σ] (1)
Investors: Risk Neutral (we want to focus on effects of learning on toasset pricing.)
General pricing result: price is equal to expected discounted valueof future dividends. (only mean matters, not the variance)
Perfect Information: Investor knows both δ and Σ
pt = p =δ
r(2)
Price is constant
⇒ No predictability in returns (dividends are i.i.d and price isconstant)
() Learning & Asset Prices May 2, 2012 5 / 22
Pricing - Parameter Uncertainty
Known Σ and Unknown δ; Investor has a diffuse (Normal) prior over δ
Posterior Distribution: after observing d1, d2,...,dt , subjectivedistribution of δ is given by
δ vs MVN
[dt ,
1
tΣ
](3)
Predictive distribution of dt+1 at time t
dt+1 vs MVN
[dt ,
t + 1
tΣ
](4)
Pricing: applying same logic
pt =1
rdt (5)
() Learning & Asset Prices May 2, 2012 6 / 22
Price properties - Basic
pt =1
rdt (6)
Result 1: Equilibrium price is a function of beliefs
Result 2: Price change if average of dividend realization is differentthan current average of dividends.
Current ’best’ belief about mean level of dividend is average ofdividends observed till the date (given i.i.d nature)
Changes in price:
pt+1 − pt =1
r(dt+1 − dt) (7)
() Learning & Asset Prices May 2, 2012 7 / 22
Predictability - Price Changes
Result 3: Prices are martingale under subjective distribution
E st (pt+1 − pt) =
1
rE st (dt+1 − dt) =
1
r(dt − dt) = 0 (8)
Result 4: Past dividends predict future prices under true distribution
Expected price changes under true distribution: Depends uponexpectation of dt+1 under true distribution
Et(pt+1 − pt) =1
rEt(dt+1 − dt)
=1
rEt
(δ + tdtt + 1
− dt
)=
1
r(t + 1)(δ − dt)
() Learning & Asset Prices May 2, 2012 8 / 22
Predictability - Returns
Result 5: Past dividends predict future returns in the data.
Et(Rt+1) = Et(dt+1 + pt+1 − pt)
= δ +1
r(t + 1)(δ − dt)
Higher past dividends above mean predict lower future returns belowmean
Result 6: Forecast Errors are correlated with past cash flows
Define URt+1 = Rt+1 - E st (Rt+1)
Et(URt+1) =
(1 +
1
r(t + 1)
)(δ − dt) (9)
Market Efficiency ⇒ Forecast errors not predictable. But we haverationality + predictable forecast errors
() Learning & Asset Prices May 2, 2012 9 / 22
Autocorrelation and Variance - Returns
Result 7: Returns are negatively autocorrelated
cov(Rt ,Rt+1) = − 1
r(t)(t + 1)σ2 (10)
Why? Higher returns today imply more likely negative surprisetomorrow which pulls down the price
⇒ Prices will be mean reverting in observed data ex - post
Volatility:
vart(pt+1) =
(1
r(t + 1)
)2
σ2 (11)
() Learning & Asset Prices May 2, 2012 10 / 22
Learning - Third Factor Explaining Time Variation
Prices are backward looking
Price volatility arises as a result of learning solely: Not because ofchanges in discount rates or changes in mean growth rate of dividends
Wrong Conclusion from data: Looking back at the data,econometrician perceives changes in prices as driven by changes inexpected returns. (Econometric tests will tell that there was nochange in mean rate of dividends at any point in time)
Hence, Learning is a completely new factor explaining pricemovements (different from Campbell - Shiller decomposition)
() Learning & Asset Prices May 2, 2012 11 / 22
Simulated Path of Prices
0 200 400 600 800 1000 1200 1400 1600 1800 2000−0.5
0
0.5
1
1.5
2
2.5
3Price Through Time
Time
Pric
es a
nd R
etur
ns
Student Version of MATLAB
() Learning & Asset Prices May 2, 2012 12 / 22
Asset Pricing: Cross Sectional Properties
Consider a multi asset model with risk averse agent (CARA Utilityfunction)
Given his subjective expectations about returns, he chooses a M-Vefficient portfolio (Under normality of prices and returns)
Conditional Variance of Returns
vart(Rt+1) =
[1 +
1
r(t + 1)
]2Σ (12)
All the Variances and covariances are scaled up because of parameteruncertainty (No asymmetry)
⇒ β are same as in no uncertainty case
() Learning & Asset Prices May 2, 2012 13 / 22
Cross Sectional Properties II
Parameter Uncertainty does not imply MV inefficiency:Investors still hold MV efficient portfolios
CAPM holds under subjective distributions
E st (Rt+1) = rpt + β [E s
t (RM,t+1 − rpM,t)] (13)
Expected ex-post deviations from CAPM under subjective beliefs iszero
But, it is not zero under true distribution: Lagged prices anddividends explain cross sectional deviation from CAPM
Why? Simply because δ is unknown
() Learning & Asset Prices May 2, 2012 14 / 22
Results Till Now and Direction Ahead
Result: Past prices and dividends predict future returns under truedistribution
Assumptions: i.i.d returns, constant parameters (δ, σ), diffuse priors
Extensions:Time varying parameters for dividends: Learning is never completeInformative Priors: Agents have some information about changes inparametersGeneral Dividend processes
Goal: How does Asset prices behave in Long run?
() Learning & Asset Prices May 2, 2012 15 / 22
Impact of Informative Priors
Suppose investor has prior about δ: mean of δ∗ and variance of σ2/h.
Higher the h, higher the accuracy of prior
Predictive Distribution: Average of prior and realizations of dividends
dt+1 vs N
[h
t + hδ∗ +
t
t + hdt ,
t + h + 1
t + hσ2]
(14)
Variance converges to true variance in the LR (as previously)
Impact of prior diminishes over time, but never disappears
Pricing: Under risk neutrality
pt =h
t+hδ∗ + t
t+hdt
r(15)
⇒ Prior acts exactly like observing dividends for h periods with meanof δ∗
All the results carry over
() Learning & Asset Prices May 2, 2012 16 / 22
Time Varying Parameters: Set Up
Why? Otherwise uncertainty diminishes over the period ⇒ no LReffects
How?Variance of dividends is constant through timeMean: True mean is redrawn from N(δ∗, σ
s ) after every K periodsSequence of short ”regimes”
What Agent Learns in the LR? True Distribution of mean ofdividend ⇒ In the LR, prior at the start of each regime is that δ vN(δ∗, σ
s )
More complicated Assumptions ⇒ even true distribution may not beknown in the LR
Reduced Form Version: Assume that prior is that δ v N(δ∗,σh ): hcould be different than s
() Learning & Asset Prices May 2, 2012 17 / 22
Time Varying Parameters: Implications
Predictive Distribution: The predictive conditional mean for nextperiod dividend after t periods in the regime
mt =h
t + hδ∗ +
t
t + hdt (16)
where dt is the average dividends in the regime
Pricing of stock
pt =1
(1 + r)K−t
(δ∗
r
)+
K−t+1∑i=1
mt
(1 + r)i(17)
Price= PV(dividends in this regime expected at mt ) + PV(perpetualstream of δ∗)
() Learning & Asset Prices May 2, 2012 18 / 22
Time Varying Parameters: Challenges
Econometrics: Difficult to tell whether conditional or unconditionalcovariance is influential for empirical test
Consider a model with only one regime
Covariance of prices and future returns is conditional on current (theonly) short term mean
For multiple regime: Econometrician implicitly conditions on severalvalues of mean dividend drawn from the distribution
A better ideas: Perform simulation
() Learning & Asset Prices May 2, 2012 19 / 22
Results: Simulation
Rt+1 = α + β (Dt/Pt) + εt (18)
Table : Slope Coefficients: Returns on Lagged DY
Regimes s h=16 h=25 h=49
K =38 16 0.74 0.36 -0.3725 0.95 0.70 0.0949 1.28 1.04 0.68
K =19 16 0.42 -0.31 -2.0425 1.01 0.50 -0.9049 1.64 1.30 0.44
() Learning & Asset Prices May 2, 2012 20 / 22
Results: Simulation
Increase in K ⇒ increase in slope coefficients: Longer the regime,better the learning
As h increase (for a given s), accuracy of prior increase andpredictability increase
As s increase (for a given h), dispersion of the draws of meanreduces and predictability improves
h > s ⇒ economy more volatile than expected ⇒ slow learning
Future returns are decreasing function of current s/h ratio
() Learning & Asset Prices May 2, 2012 21 / 22
Comments
Elegant idea
But difficult to take to data
Ideally, we would like to isolate the impact of learning on price variation(Tractability of Shiller’s approximation)Paper provides no guidance on this front.
Heterogeneous Priors:Will prices reflect ”best” (more accurate) available prior? (Recall fromMicro: Prices could be revealing information)In that case, how much influence learning of ”Naive” agent can haveon prices? Probably not much ⇒ Quantitatively may not be important
() Learning & Asset Prices May 2, 2012 22 / 22