You are on page 1of 34

FRAGILITY HEURISTICS

Nassim Nicholas Taleb


(NYU and Real World Risk Institute LLC)
Feb 18, 2016
Bank of England

SOME STATISTICAL
RIGOR

Main Problem
uTypical Response: Nothing new, we know

a lot about fat tails

Q: Then why do you keep blowing up from fat-tailed


events?
Q: Why isnt econometrics more effective than
astrology?
Q: Then why do you use Markowicz mean-variance?
u It

is not changing the color of the dress.


Scientific Rigor: Entire statistical methodology,
decision theory need to be revised/adapted.

Law of Large Numbers

Convergence (temporal & cross sectional)


MAD

20

40

60

80

100

1 day in >10,000 provides the bulk of 4th


moment

Inequality Measures
u Under

Fat Tails, measures of concentrations are


highly biased and super-additive
GINI
Centiles

Much of the recent attributed growth in inequality


could be an optical illusion/ mismeasurement
We need more scrutiny

GINI and Pikettism

EVT

Sorry, but Central Limit fails for power-law


distributed finite variance processes
Tails never

become thin,
even under
finite variance
Central limit
never reaches
asymptote

Social Science Statistical Toolkit

Good News: we know where the bias is!


Probability

Pareto 80/20
For a negatively (positively)
skewed distribution > 93% of
observations fall above
(below) the mean.
Unseen rare events

-140

-120

-100

-80

-60

-40

-20

Outcomes
%>(<)mean
1.00
0.95
0.90
0.85
0.80
0.75

1.5

2.0

2.5

3.0

The left graph shows hormesis for an organism (similar to Figure 19): we
can see a stage of benefits as the dose increases (initially convex) slowing down into
a phase of harm as we increase the dose a bit further (initially concave); then we see
things flattening out at the level of maximum harm (beyond a certain point, the organism is dead so there is such a thing as a bounded and known worst case scenario
in biology). To the right, a wrong graph of hormesis in medical textbooks showing
initial concavity, with a beginning that looks linear or slightly concave.

FIGURE 34.

Asymmetry

THE INVERSE TURKEY PROBLEM


Probability
Probability
Probability

Probability
Probability

Unseen
events
Unseen
rarerare
events
Unseen
rare events

Unseen rare rare


events events
Unseen

EXTREME RISK INITIATIVE NYU SCHOOL OF ENGINEERIN


20

FIGURE 35.

40

60

80

100

120

140

Outcomes
Outcomes

140 140 120 120 100 100 80

80

60

60

40

40

20

20

Outc Outcomes
Outcomes

Antifragile, Inverse Turkey Problem: The unseen rare event is positive.

4) Difference
betweenskewed
sample
mean
likeWhen
you look at a positively
(antifragile)
timeand
series maximum
and make inferences
about themean:
unseen, you
the good
stuff and underestimate
benefits
(the Pisano,
lihood
: miss
Table
I shows
the true the
mean
using
the
2006a, 2006b, mistake). On the right, the other Harvard problem, that of Froot
parametrization
ofcorresponds
the Pareto
distribution
above
inverting
(2001). The filled area
to what
we do not tend
to seeand
in small
samples,
insufficiency of points.
Interestingly
the shaded
area increases
with or
model
error.
thefrom
transformation
back
to compact
support.
"True"
maxiThe more technical sections call this zone (turkey) and (inverse turkey).
mum likelihood, or "statistical" mean is between 3 and 4 times
observed
mean.
DIFFERENCE
BETWEEN POINT ESTIMATES AND DISTRIBUTIONS
This means the "journalistic" observation of the mean, aside
Let us apply this analysis to how planners make the mistakes they make, and why
deficits
tend
to be worse than
planned:of relying on sample mean, unfrom
the
conceptual
mistake
derestimates the true mean by at least 3 times and higher future
observations would not allow the conlusion that violence has
"risen".
B

E STIMATES ( AND STAN


D ISTRIBUTION PARAME
F OR BOTH ACTUAL AND

NUMBER OF EVENTS LYI


OF

Data
Raw Data

Naive Rescaling
Log-rescaling

SHADOW MEAN
TABLE I

S AMPLE MEANS AND ESTIMATED MAXIMUM LIKELIHOOD MEAN ACROSS


MINIMUM VALUES L R ESCALED DATA .

L
Sample Mean
10K
9.079 106
25K
9.82 106
Tale_9781400067824_3p_all_r1.indd 444
50K
1.12 107
100K
1.34 107
200K
1.66 107
500K
2.48 107

ML Mean
3.11 107
3.62 107
4.11 107
4.74 107
6.31 107
8.26 107

Ratio
3.43
3.69
3.67
3.53
3.79
3.31

B. Conclusion
History as seen from tail analysis is far more risky, and conflicts far more violent than acknowledged by naive observation

III. M ET

A. Rescaling method

We remove the c
laws as follows (see
for number of incide
a naive rescaling of
population at period
of Ht .
Next, with today
naively rescaled mi
introduce a smooth
satisfying:

10/10/12 8:44 AM

i ' is "smooth": '


1

17

Regular is predic.ve of regular


M@t+1D
0.030
0.025
0.020
0.015
0.010
0.005
0.005

0.010

0.015

0.020

0.025

M@tD
0.030

18

.but not the tails (sorry, again,


nonparametric stat cant do anything for you
there)
M@t+1D
0.0004

Concentration of tail events


without predecessors

0.0003
Concentration of tail events
without successors

0.0002

0.0001

0.0001

0.0002

0.0003

0.0004

0.0005

M@tD

Fake Empiricism/Public
(MIS)understanding of Science

Unconditional Dislike of Volatility/Disorder


Cluster

Fragility Logic
A- All fragile objects dislike disorder/the disorder
cluster [by definition]. If you accept definition/
formulation, then
" Fragility can be expressed as aversion to volatility
B- Everything that is averse to uncertainty in tails
has to have a nonlinear concave response North of
the breaking point [by theorem, with R. Douady 2013]
E.g. 10% drop in the market harms more than 2 x a 5% drop

" Fragility requires asymmetry


" Fragility is detectable, even measurable
" Fragility is dose dependent

In probability of payoff

Sensitivity to s (TBD) below K

DISORDER CLUSTER

i) uncertainty, ii) variability, iii)


imperfect, incomplete
knowledge, iv) chance, v)
chaos, vi) volatility, vii) disorder,
viii) entropy, ix) time, x) the
unknown, xi) randomness, xii)
turmoil, xiii) stressor, xiv) error

Disorder Cluster(nonparametric s)

s = | x | p(x)dx

s = | x | p(x)dx
+

s = | x | p(x)dx

Transfer Theorems


Convexity Incr. Right tail Likes vola.lity
&
Concavity Incr. LeH tail Dislikes vola.lity (or disorder
brothers)

x and f(x) are dierent animals bridged by a


convex func.on

NONLINEARITY AND
NECESSITY

Convexity Eects
For the coee cup,

shocks bring higher and


higher harm as their
intensity increases (up to
the point of breaking).

Jumping 10 .mes 1

meter harms less than 1


.me 10 meter
Times

meters
1

10.00

5.00

10

1.00

100

0.10

1,000

0.01

10,000

0.001

100,000

0.0001

Zoom

Harm
fHXL

Event Size
X

A matching linear
exposure would be
very volatile for small
variations.

Degrees of Nonlinearity
Response
3.0
Broken glass !
2.5
2.0
1.5
1.0
0.5

10

15

Dose

20

1.0
Extremely high
: Heaviside

0.8

Higher

0.6

(more Sigmoidal)
0.4
Lower
(more linear)

-10

0.2

-5

10

Either Broken or Intact


40

Fully

Fully

Broken

Intact

Item

low

produces an

Item

Arcsine distribution

30

High values of

produce

a bell shaped distribution


20

10

0.2

0.4

0.6

0.8

1.0

In words
All that has bimodal distribution of type above is

necessarily nonlinear
The fragile is necessarily locally indifferent to small
variations
Banks that blow up from unexpected shocks are necessarily

indifferent to the small variations of same variables

Tail Options and Tail Events


5

10

15

L = 10
-2000

-4000

-6000

-8000

L = 20

L=0

20

Explosions of Vol

ATM
L=5
L=10
L=15
L=20

2
2
5
27
302
7686

3
3
10
79
1486
72 741

SIMPLE HEURISTIC

4
4
16
143
3298
208 429

The Heuris.c
Compute model or risk model at parameter p, p + x%, p-x%
(where x% is the average devia.on)
Model or Measure:
-Stress test ( p is the stress test, say -15%, -20%,-25%)
-ShorXall (modied VAR) by varying either vola.lity or tail level K
- Any parameter entering equa.on

Errors in ruler

BARBELL

Barbell
(Extremely Conservative)
+ (1-) Very Aggressive

> Medium Strategy


01

Convex Transformation
Probability distribution of x

Probability distribution of fHxL


Low Uncertainty

Convex transformation f(x)


High Uncertainty

The Problem of Exposure


DOWNSIDE
Convex
Convex
Concave
Concave

UPSIDE
Convex
Concave
Concave
Convex

Antifragile
Robust
Fragile
Fragile

Convex Losses, Concave Gains


Thin Tails Robust
(S curve in nature)

Convex Exposure Right Skew


70

0.6

60

0.5

50

0.4

40

0.3
30

0.2
20

0.1

10
-2

10

Fence Fat Tails Fragile

Generalizing the S Curve


1.0

Antifragile

0.5
Fragile Zone

-5

-0.5

-1.0

10

15

20
Harm
Slowing Down

Size Effects

52

Too big to fail

HOW MODELS FALL


APART

Model Error/Risk Exposures




ERROR= Error1 +Error2+Error3


RISK = Risk1 +Risk2 +Risk3
Error/Risk1: Linear rare (shows early)
Error/Risk2: Missed signicant variable rare, small eect in
nance/economics
Error/Risk3:Missed 2nd (+) Order Eect [PROCRUSTEAN BEDS]
common
CENTRAL IDEA Model error detec.on For Error3 detects Risk3
RISK = MODEL ERROR
Tail mispricing (IMF paper, 2012)= 2nd (+) Order eects

Convexity Eects
(Expecta.on)

Error = Average Health of Grandmother


across temperatures Health of
Grandmother (given an average
temperature)

t + t2
H 1

H (t1 ) + H (t 2 )
2

v/s

Jensens Inequality
Response f

f Hx + DxL + f Hx - DxL
2

H
fHxL
Dose

Model Error (Negative)


A

B (K )

) ( )

f x d dx

) ( )

( )

f (x d )dx

f x d dx

( )

f (x d )dx

K
1
f ( x | + ) + f ( x | ) dx f (x | ) dx
2

B (K )

B ( X ) =

1
f ( X | + ) + f ( X | ) f ( X | )
2
K

B (K ) = B (x) dx

Example: Why Decits Are (Almost) Always


Worse than Promised
POINT ESTIMATE:
Unemployment 9% Decit 200 billion

Unempl 8% Decit $75 billion (improves 125 bil)
Unempl 10% Decit 550 bil (worsens 350)
Projec.on Error -> at least underes.mated by 225

Convexity, Convexity Bias


Deficit bil
1400
1200
1000
800
600
400
200
8

10

11

Unemployment

Bias

Fragility unseen left tail

1000

800

600

400

Prediction Problem

200

Deficit

Model Error (Positive)


Robust in the Left Tail and

C (K )

f x ( ) d dx f (x ( ) d )dx

Missing the point


Missing the Point

Markowitz mean variance


Ricardo compara.ve
advantage
Debt
Samuelson op.miza.on
Small Probabili.es (worse in
the tails)

Vulnerability

Under-diversica.on
Overspecializa.on

Projec.on error
Everywhere
Fukushima, Crisis 2007-20??,
etc.

Unrealis.c but coherent

[Protagoras problem]
Unrealis.c & incoherent

65

66

Even a blind man can see

Probability

Hidden Iatrogenics

Benefits

Outcomes

You might also like