risk aversion, information acquisition, and technology ... · increasing adoption policies...
TRANSCRIPT
Risk Aversion, Information Acquisition,and Technology Adoption
Canan Ulu Jim SmithMcDonough School of Business Fuqua School of Business
Georgetown University Duke University
SAMSI Games and Decisions in Reliability and Risk WorkshopMay 2016
0
Problem: Should Jim buy a Tesla? Or should hewait and learn more?
Wait
Adopt
Quit
Wait
Adopt
Quit
Value of te
chno
logy
Hear about it See 99 rating in consumer reports
Read about cars catching fire after a crash
Wait
Adopt
Quit
103 rating in consumer reports
Wait
Adopt
Quit
Other examples: farmer planting a new variety of soybean, utilitybuilding a power plant based on a new technology, or doctors changingtreatments
1
We study a DP model of information acquisitionin technology adoption decisions.
We build on McCardle (1985) and Ulu and Smith (2009), addingrisk aversion.
In each period, the consumer can adopt the technology, gatherinformation about the technology, or quit.
• State variables: probability distribution on technology benefitswealth
• Beliefs are updated over time using Bayes’ Rule.
• Arbitrary distributions are allowed.
• Information gathering is costly.
We focus on structural properties of the model:
• Properties of the value function (increasing, convex, . . . )
• Monotonicity properties of the optimal policies
• Effects of risk aversion
2
Modeling learning: Notation
k-1
Observe signal x
θ
k
Value of technology
Periods to go
Prior Signal Distribution
Posterior
π(θ) f(x;π) =∫θ L(x|θ)π(θ) dθ Π(θ;π, x) = L(x|θ)π(θ)
f(x;π)
π f(π) Π(π, x)
3
Decision Tree Example: Risk Neutral
Initial Wealth Wealth25 .05 Low (-20)
5.000Neg Pos Adopt .35 Med (-10)
Low (-20) 0.80 0.20 15.000Med (-10) 0.70 0.30 35.500High (25) 0.15 0.85 .60 High (25)
50.000.11 Low (-20)
2.000Adopt .65 Med (-10)
12.00019.333
.38 Negative .24 High (25)2 47.000
22.000Reject
1 22.00035.5
Wait (-3) .02 Low (-20)2.000
33.500Adopt .17 Med (-10)
12.00040.400
.63 Positive .82 High (25)1 47.000
40.400Reject
22.000Reject
25.00025.000
Likelihood
4
Decision Tree Example: Risk Averse
Initial Wealth Wealth Utility25 .05 Low (-20)
5.000 1.609Neg Pos Adopt .35 Med (-10)
Low (-20) 0.80 0.20 15.000 2.708Med (-10) 0.70 0.30 3.376High (25) 0.15 0.85 .60 High (25)
50.000 3.912.11 Low (-20)
2.000 0.693Adopt .65 Med (-10)
12.000 2.4852.621
.38 Negative .24 High (25)2 47.000 3.850
3.091Reject
2 22.000 3.0913.390563
Wait (-3) .02 Low (-20)2.000 0.693
3.391Adopt .17 Med (-10)
12.000 2.4853.570
.63 Positive .82 High (25)1 47.000 3.850
3.570Reject
22.000 3.091Reject
25.000 3.2193.219
Likelihood
5
The value function:
c = cost of waiting (c > 0)
u(w) = DM’s utility for wealth w
Value (or derived utility) function with k periods remaining:
U0(w, π) = u(w)
Uk(w, π) = max
E[u(w + θ̃) |π ] (adopt)
E[Uk−1(w − c,Π(π, x̃)) | f(π) ] (wait)
u(w) (quit)
where
E[u(w + θ̃) |π ] =
∫θ
u(w + θ)π(θ) dθ
E[Uk−1(w,Π(π, x̃)) | f(π) ] =
∫x
Uk−1(w,Π(π, x))f(x;π) dx
6
Illustrative example: beta-Bernoulli model
θ = p− 0.5 where π(p) ∝ p(α−1)(1− p)(β−1) with p ∈ [0, 1]
Expected benefit (E[ θ ]) = α/(α+ β)− 0.5; “precision” = (α+ β)
Signals are + or − with probability p or (1− p). Precision increases byone each period. Given prior with (α, β),
+ signal =⇒ (α+ 1, β)− signal =⇒ (α, β + 1)
Example: Start with (α, β) = (2.25, 1.75), observe (−,+,−,+,−,−):
‐0.50 ‐0.25 0.00 0.25 0.50
Benefit of the technology ()
Prior
After ‐
After ‐,+
After ‐,+,‐
After ‐,+,‐,+
After ‐,+,‐,+,‐
After ‐,+,‐,+,‐,‐
7
Illustrative example: risk-neutral results
u(w) = w; initial wealth = 1.06; c = 0.01; long time horizon
‐0.25
‐0.20
‐0.15
‐0.10
‐0.05
0.00
0.05
0.10
0.15
0.20
0.25
0 5 10 15 20 25
Expected Benefit:E[]
= (/(+
‐0.5)
Time/Precision (+)
Adoption Region
RejectionRegion
WaitRegion
LR‐improving
Policy regions
0.90
0.95
1.00
1.05
1.10
1.15
‐0.10 ‐0.05 0.00 0.05 0.10 0.15
Expected Utility
Expected benefit: E[] = (/(+)‐0.5)
Reject
Adopt
Wait
LR‐improving
Value function with α+ β = 10
8
Illustrative example: risk-averse results
u(w) = 1.2− 0.2w(1−γ) where γ = 6; initial wealth = 1.06; c = 0.01
‐0.25
‐0.20
‐0.15
‐0.10
‐0.05
0.00
0.05
0.10
0.15
0.20
0.25
0 5 10 15 20 25
Expected Ben
efit:E[]
= (/(+
‐0.5)
Time/Precision (+)
Adoption Region
WaitRegion
RejectionRegion
LR‐improving
Policy regions
0.90
0.95
1.00
1.05
1.10
1.15
‐0.10 0.00 0.10 0.20 0.30
Expected Utility
Expected benefit: E[] = (/(+)‐0.5)
Reject
Adopt
Wait
LR‐improving
Value function with α+ β = 10
9
General results: Defining “better” priors
Definition: π2 likelihood-ratio (LR) dominates π1 (π2 �LR π1) ifπ2(θ)/π1(θ) is increasing in θ.
Examples of LR improvements:
• Beta: Increasing α while holding the precision (α+ β) constant• Normal: Increasing the mean while holding the variance constant
LR-dominance implies FOSD, but the reverse is not true.
0 1
π1
π2
Not a LR-improvement
0 1
π1
π2
A LR-improvement
The LR-order survives Bayesian updating: given a signal x
π2 �LR π1 ⇔ Π(π2, x) �LR Π(π1, x), for all x ∈ X.
10
General results: Ordered signal processes
Definition: The signal process L(x|θ) satisfies the monotone-likelihood-ratio (MLR) property if the signal space X is ordered and
L(x|θ2) �LR L(x|θ1) for all θ2 ≥ θ1 .
Examples: Bernoulli signals; normal signals
If the signal process satisfies theMLR property, then:
• π2 �LR π1 ⇒ f(π2) �LR f(π1)
• For any prior π, x2 ≥ x1 ⇔ Π(π, x2) �LR Π(π, x1).
11
General results: Increasing value functions
Definition: V (π) is LR-increasing if V (π2) ≥ V (π1) wheneverπ2 �LR π1.
Proposition [Increasing]: Suppose the DM’s utility function u(w) isincreasing in w and the signal process satisfies the MLR property. Then,for all k and w, the value function Uk(w, π) is LR-increasing in π.
12
General results: Increasing value functionsValue (or derived utility) function with k periods remaining:
U0(w, π) = u(w)
Uk(w, π) = max
E[u(w + θ̃) |π ] (adopt)
E[Uk−1(w − c,Π(π, x̃)) | f(π) ] (wait)
u(w) (quit)
0.90
0.95
1.00
1.05
1.10
1.15
‐0.10 0.00 0.10 0.20 0.30
Expected Utility
Expected benefit: E[] = (/(+)‐0.5)
Reject
Adopt
Wait
LR‐improving
Value function with α+ β = 10
13
General results: Increasing policies
Proposition: Suppose the DM’s utility function is increasing and thesignal process satisfies the MLR property.
• Rejection: If it is optimal to reject with prior π2, it is also optimalto reject with any prior π1 such that π2 �LR π1.
Proof: Follows from LR-increasing value functions.
• Adoption: Suppose the DM is risk neutral (or risk seeking). If it isoptimal to adopt with prior π1, then it is also optimal to adopt withany prior π2 such that π2 �LR π1.
Proof: Utility difference between adoption and waiting is LR-increasing.
With risk neutrality, policies “increase” from quit to wait to adopt as πLR-improves.
14
Illustrative example revisited
Policies are LR-increasing.
‐0.25
‐0.20
‐0.15
‐0.10
‐0.05
0.00
0.05
0.10
0.15
0.20
0.25
0 5 10 15 20 25
Expected Benefit:E[]
= (/(+
‐0.5)
Time/Precision (+)
Adoption Region
RejectionRegion
WaitRegion
LR‐improving
Policy regions
0.90
0.95
1.00
1.05
1.10
1.15
‐0.10 ‐0.05 0.00 0.05 0.10 0.15
Expected Utility
Expected benefit: E[] = (/(+)‐0.5)
Reject
Adopt
Wait
LR‐improving
Value function with α+ β = 10
15
Illustrative example revisited
Policies are LR-increasing.
‐0.25
‐0.20
‐0.15
‐0.10
‐0.05
0.00
0.05
0.10
0.15
0.20
0.25
0 5 10 15 20 25
Expected Ben
efit:E[]
= (/(+
‐0.5)
Time/Precision (+)
Adoption Region
WaitRegion
RejectionRegion
LR‐improving
Policy regions
0.90
0.95
1.00
1.05
1.10
1.15
‐0.10 0.00 0.10 0.20 0.30
Expected Utility
Expected benefit: E[] = (/(+)‐0.5)
Reject
Adopt
Wait
LR‐improving
Value function with α+ β = 10
16
Illustrative example revisited
0.90
0.95
1.00
1.05
1.10
1.15
‐0.10 ‐0.05 0.00 0.05 0.10 0.15
Expected Utility
Expected benefit: E[] = (/(+)‐0.5)
Reject
Adopt
Wait
LR‐improving
Risk neutral
0.90
0.95
1.00
1.05
1.10
1.15
‐0.10 0.00 0.10 0.20 0.30Expected Utility
Expected benefit: E[] = (/(+)‐0.5)
Reject
Adopt
Wait
LR‐improving
Risk averse
With risk aversion, can adopt and wait cross twice as weLR-improve π?
17
If the DM is risk averse, adoption policies maynot be monotonic in π.
Example: log utility; three technology values; signals satisfy MLR property
Wealth Utility
Initial Wealth=23.002 .08 Low (‐20)
c=3 3.0 1.10
log utility
Adopt .35 Med (‐10.5)
12.5 2.52
.57 High (+25)
48.0 3.87
.15 Low (‐20)
0.002 ‐6.21
Adopt .65 Med (‐10.5)
9.5 2.25
.41 ‐ Signal .19 High (+25)
2 45.0 3.81
Quit
1 20.0 3.003.177209003
Wait (‐3) .03 Low (‐20)
0.002 ‐6.21
Adopt .14 Med (‐10.5)
9.5 2.25
.59 + Signal .83 High (+25)
1 45.0 3.81
Quit
20.0 3.00
Quit
23.0 3.14
Wealth Utility
.00 Low (‐20)
3.0 1.10
Adopt .38 Med (‐10.5)
12.5 2.52
.62 High (+25)
48.0 3.87
.00 Low (‐20)
0.002 ‐6.21
Adopt .77 Med (‐10.5)
9.5 2.25
.38 ‐ Signal .23 High (+25)
2 45.0 3.81
Quit
2 20.0 3.003.362675024
Wait (‐3) .00 Low (‐20)
0.002 ‐6.21
Adopt .14 Med (‐10.5)
9.5 2.25
.62 + Signal .86 High (+25)
1 45.0 3.81
Quit
20.0 3.00
Quit
23.0 3.14
18
Comparing waiting and adopting in this example:
0.00
0.10
0.20
0.30
0.40
0.50
0.60
0.70
0.80
0.90
1.00
0 5 10 15 20 25 30 35 40 45 50 55
Cumulative Probability
Wealth + Technology Value () ‐ Search Costs
Adopt
Wait
‐8.00
‐6.00
‐4.00
‐2.00
0.00
2.00
4.00
0 5 10 15 20 25 30 35 40 45 50 55
Utility
Wealth + Technology Value () ‐ Search Costs
Log utility
With log utility, a bad technology outcome + search costs can becatastrophic if the resulting wealth level is near zero.
• Search costs push the DM “over the edge”
• Can we ensure monotonicity by limiting the degree of risk aversion?
19
Increasing Adoption Policies
Proposition Suppose the DM is risk averse and her utility function u exhibitsdecreasing absolute risk aversion (DARA), i.e., her risk tolerance τu(w) isincreasing. Then, if
τu(w0 + θ − c) ≥ −θ
where w0 = w − kc and θ = min θ (“Not too risk averse”), then adoptionpolicies are monotonic.If u is CARA, no risk tolerance bound is required.
We define a new property: “sLR-increasing”:
• LR-increasing functions are sLR-increasing
• sLR-increasing functions are single-crossing
• Bayesian updating preserves sLR-increasing property
We show utility difference between adoption and waiting is sLR-increasing
• Then, utility difference between adoption and waiting is singlecrossing.
20
Summary:With a DARA utility function that is “not too risk averse,” we have:
‐0.25
‐0.20
‐0.15
‐0.10
‐0.05
0.00
0.05
0.10
0.15
0.20
0.25
0 5 10 15 20 25
Expected Ben
efit:E[]
= (/(+
‐0.5)
Time/Precision (+)
Adoption Region
WaitRegion
RejectionRegion
LR‐improving
The same structural properties as in the risk-neutral model(existence of thresholds, etc.), but risk aversion leads to quittingsooner and adopting later (if CARA).
21
Generalizations:
Model with discounting, u(w + NPV of costs/benefits)
• Delay is costly; also risk reducing
• Results and proofs follow the same pattern
Other applications of s-increasing: Monotonic policies in DPs
• Submodularity (increasing differences) conditions (Topkis (1979),Lovejoy (1987a,b)) are sometimes hard to establish
• Single-crossing conditions (e.g., Milgrom and Shannon (1994), Quahand Strulovici (2012),. . . ) are hard to use in DPs
22
Thank you!
23