The “risk matrix:” Predicting financial risk taking with FMRI risk matrix_english.… · The...
Transcript of The “risk matrix:” Predicting financial risk taking with FMRI risk matrix_english.… · The...
The “risk matrix:”Predicting financial risk taking with FMRI
Brian KnutsonStanford University
Financial Education and Investor Behavior Conference, Rio de Janeiro, December 4, 2014
Overview
• Background• Findings
– Predictors of financial risk taking– Anomalies and aggregation
• Conclusions and implications
Questions
• Which brain mechanisms anticipate good and bad events (affective neuroscience)
• Does their activity influence choice? (neuroeconomics)
correlated activity
predictive activity
output input
Anticipatory Affect
Watson J, Clark LA, Tellegen A (1988). J Pers Soc Psychol, 54, 1063-1070.
Avoidance Arousal
calm
quiet
tired
unhappy
fearful
happy
excited
aroused
Approach
Valence
Positive Arousal
Negative Arousal
Anticipating Gains Versus Losses
Gain Anticipation > Loss Anticipation; n=21 studies
Knutson B, Greer SM (2008). Phil Trans Royal Soc B, 363, 3771-3786.
Anticipatory Affect Model
• Functional targets– Nucleus accumbens (NAcc): Gain anticipation– Anterior insula: Loss anticipation– Medial prefrontal cortex (MPFC): Value integration
Knutson B, Greer SM (2008). Phil Trans Royal Soc B, 363, 3771-3786.
Risk Taking Predictions
• Risks involve uncertain but large gains and losses• NAcc activation -> risk seeking• Anterior insula activation -> risk avoidance
Kuhnen CM, Knutson B (2005). Neuron, 47, 763-770.
AnteriorInsula
NucleusAccumbens
Behavioral Investment Allocation Strategy (BIAS) Task
• 20 blocks of 10 trials• At block outset, one stock is randomly assigned as “good,”
the other is “bad”• Subjects know one stock is better, but don’t know which:
Good stock: Bad stock: +$10 w/ prob. 50% +$10 w/ prob. 25% –$10 w/ prob. 25% –$10 w/ prob. 50% +$0 w/ prob. 25% +$0 w/ prob. 25%
Bond: +$1 w/ prob. 100%
Kuhnen CM, Knutson B (2005). Neuron, 47, 763-770.
BIAS Task Trial
BIAS Task Outcomes
• Choices:– Risk seeking: Choosing a stock– Risk averse: Choosing a bond
• Mistakes:– Violate risk neutrality, prior information– Risk seeking: Choosing a stock when the bond is optimal– Risk averse: Choosing a bond when the stock is optimal
Anticipatory activity predicts choice
• NAcc activity precedes risk seeking shift (bond to stock) • Anterior insula activity precedes risk averse shift (stock to bond)• Effects are stronger for “irrational” choices (mistakes)
...but...
Multiple systems or just one?
Tom S et al (2007). Science, 315, 515.
Anomalies
• Higher order moments (e.g., skewness)• Incidental affect
Skewed Gambles
Wu CC, Bossaerts P, Knutson B (2011) PLoS ONE, 6, e16838
Skewed Gambles
Wu CC, Bossaerts P, Knutson B (2011) PLoS ONE, 6, e16838
Skewed Gambles
Wu CC, Bossaerts P, Knutson B (2011) PLoS ONE, 6, e16838
< <
Skewed Gambles
Wu CC, Bossaerts P, Knutson B (2011) PLoS ONE, 6, e16838
= =
Skewed Gambles
Wu CC, Bossaerts P, Knutson B (2011) PLoS ONE, 6, e16838
Predicting choice with GraphNet
Grosenick L et al (2013). NeuroImage, 72, 304-321.
• Predictive• Interpretable• Generalizable
Neural predictors of gambling
Wu CC, et al (In Prep).
2 s
+$5.25
-$1.75
$0.00
+$5.25
-$1.75
$0.00
R
L
4 s
+ $5.25
Total: $15.25
2 s
2 s
+$5.25
-$1.75
$0.00
+$5.25
-$1.75
$0.00
R
L
4 s
+ $5.25
Total: $15.25
2 s
Neural predictors of gambling
Wu CC, et al (In Prep).
2 s
+$5.25
-$1.75
$0.00
+$5.25
-$1.75
$0.00
R
L
4 s
+ $5.25
Total: $15.25
2 s
LOSO CV = 65%
Neural predictors of gambling
Wu CC, et al (In Prep).
Neural predictors of gambling
Volumes of interest
Graphnet
Volumes of interest
Wu CC, Sacchet MD, Knutson B (2012). Frontiers in Neuroscience, 8, 159. see also Mohr PNC et al (2010) J Neurosci, 30, 6613-6619 .
n = 10 studiesn = 21 studies
Graphnet
Predictors vary by gamble type
DV: Risky Choice Coefficient (SEM) Z-scoreTrial number -0.008 (0.005) -1.73
Total earnings -0.079 (0.019) -4.24***Previous outcome -0.093 (0.033) -2.79**
Gamble 0.036 (0.259) 0.14NAcc (bilateral) 2.311 (0.316) 7.30***
Ant Ins (bilateral) -0.732 (0.239) -3.07*NAcc x Gamble 0.737 (0.233) 3.16*Ant Ins x Gamble -0.125 (0.293) -0.43
Constant 1.451 (0.468) 3.10*
# of observations: 2736Pseudo-R2: 0.194
*p < 0.05**p <0.01
***p < 0.001
Positive skew preference
Wu CC, et al (In Prep).
• Low and high magnitudes• Gain and loss frames• Hypothetical and actual
Incidental affect: Pictures
Knutson B et al (2008) NeuroReport, 19, 509-513.
• Incidental affective cues can influence financial risk taking
Incidental affect: Pictures
Knutson B et al (2008) NeuroReport, 19, 509-513.
• With partial mediation by brain activity
inhale
6 s
choose
choose
4 s
6 s+
4-8 s
Incidental affect: Odors
+ $2.65
- $2.65
+ $3.41
- $2.05
+ $4.58
- $1.53
+ $7.00
- $1.00
+ $16.52
- $0.42
- $2.65
+ $2.65
- $3.41
+ $2.05
- $4.58
+ $1.53
- $7.00
+ $1.00
- $16.52
+ $0.42
Positive Skew Negative Skew
Incidental affect: Odors
F(1, 27) = 4.27; p < .05
*
positive neutral negative
Pos
itive
-Ske
w P
refe
renc
e ±
SE
M
-0.3
-0.15
0
0.15
0.3
Smells
Incidental affect: Odors
positive neutral negative
Neg
ativ
e-S
kew
Pre
fere
nce
± S
EM
-0.3
-0.15
0
0.15
0.3
SmellsF(1, 27) = 1.571; p = .221
Incidental affect: Odors
Scaling to the aggregate?
Cohn A., Fehr E., Marechal MA (In Prep)
Scaling to the aggregate?
Andrade EA, Odean TA, Lin S (In Prep)
Challenge
• “Provide an example of a novel economic model derived originally from neuroeconomic research that improves our measurement of the causal relationship between a standard exogenous environmental condition -- one with which economists have been historically concerned -- and a standard economic choice.” (Bernheim)
Bernheim BD (2008). In Neuroeconomics: Decision Making and the Brain, 115-125.
Conclusions
• Two brain systems* predict risk taking– NAcc activity predicts financial risk seeking – Anterior insula activity predicts financial risk avoidance
• Neurally-informed models generate broader explanations and predictions than traditional models (e.g., mean-variance)
– Higher order moments– Incidental affect
*or more
Implications
• Do other circuits (e.g., habitual, symbolic, reflective) modulate or overshadow the influence of affective circuits (and if so, when)?
• Do individual predictions aggregate to the group level (and if so, when)?
• Can this information improve investor behavior?• Can the information help predict market events?
Thanks:
• Charlene Wu• Peter Bossaerts• Spanlab• McClure Lab• National Science Foundation• You