Transcript Slide 1

Paris, March 2015
Alexander Denev

As of today stress testing scenarios used by
institutions are
 Often lacking a coherent story behind
 Run through backward-looking models calibrated
on historical data
 Insight free as based purely on macroeconomic
aggregates
 Excessively complex to run
 Yielding results difficult to interpret and to
understand
A silo approach to Stress Testing is still predominant
in institutions
 The models used for Stress Testing are based on
hundreds of variables built on different datasets


The outputs of these models are sometimes
aggregated at a firm-wide level, generally without
accounting for diversification/amplification between
them
Stress Test Result
=
Output Model 1
+
Output Model 2
+
Output Model 3
1.
Gives rise to transparent structures where all the
components interact in a visible manner and can be
understood by the Board/ Senior Management
1.
2.
Gives rise to transparent structures where all the
components interact in a visible manner and can be
understood by the Board/ Senior Management
Can aggregate different sources of information e.g.
historical datasets, market prices, expert opinions etc
1.
2.
3.
Gives rise to transparent structures where all the
components interact in a visible manner and can be
understood by the Board/ Senior Management
Can aggregate different sources of information e.g.
historical datasets, market prices, expert opinions etc
Can be easily updated in the light of new information
1.
2.
3.
4.
Gives rise to transparent structures where all the
components interact in a visible manner and can be
understood by the Board/ Senior Management
Can aggregate different sources of information e.g.
historical datasets, market prices, expert opinions etc
Can be easily updated in the light of new information
Has a rigorous mathematical framework behind it
1.
2.
3.
4.
5.
Gives rise to transparent structures where all the
components interact in a visible manner and can be
understood by the Board/ Senior Management
Can aggregate different sources of information e.g.
historical datasets, market prices, expert opinions etc
Can be easily updated in the light of new information
Has a rigorous mathematical framework behind it
Allows to model holistically different risk types e.g.
credit, market and liquidity risks
Probabilistic Graphical Models (PGM)
HPI Index
Fall
GDP
RMBS
losses
PD RE
PD F
PD TR
PD S
Corporate
defaults
UE
Bank
defaults
GDP Fall
Variables
PD Farmers
PD Builders
PD Developers
PD Drivers
Probabilistic Relationship
CMBS
losses

Probabilistic Graphical Models (PGM) allow:
 To represent visually systems of regression equations
 Understand how the different equations interact through
simple topological rules and hence unveil inconsistencies
PGM as carriers of regression relationships

A simple topological rule allows to read independencies
between error terms i.e. how the different equations interact
GDP
Unemployment
PD Mortgages
PD Real Estate
Companies
The error terms between these two variables are independent but is it realistic?

Probabilistic Graphical Models (PGM) allow to
incorporate for some of the variables forwardlooking information not purely based on historical
regressions and thus be used for scenario analyses
e.g.
 Market implied distributions
 Statistical surveys
 Expert opinions
PGM as carriers of non-regression based causal relationships
Event or Factor
Default of a
Bank
Causal Relationship
Major
readjustment
of risk premia
Restriction of
Credit to the
Economy
Conditional Event or
Variable
House prices
fall
PGM as a tool to distil complex economic narratives
We think that if the vote is for independence Scotland will take 50% oil reserves,
25% of the debt (though could be less) and will have to establish its own currency.
It will have to establish new regulators and because the Scottish government will
be weaker we expect a lot of deposits to leave Scotland for England. We think
that the cost of funding in Scotland will be driven largely by the Oil price, the
share of reserves they take, and the debt they assume. We think that the taxes
the government charge will also be affected by the oil prices, oil reserves they
have, debt assumed and whether their new currency is pegged. Whether their
new current is pegged will depend on the debt they have to service and the oil
share they take.
 The equity of the Scottish banks will depend on the approach of the new
regulators and the degree of deposit flight they experience – and also the cost of
funds for Scotland has a whole.
 Scottish unemployment will be affected by the taxes the new government levy
and the cost of funds in Scotland.
 Scottish GDP will depend on the cost of Funds for Scotland and the taxes that the
new government can raise.





We think that if the vote is for independence Scotland will take 50% oil reserves, 25% of the debt (though could be less) and will have to establish its own currency. It will have to establish new
regulators and because the Scottish government will be weaker we expect a lot of deposits to leave Scotland for England. We think that the cost of funding in Scotland will be driven largely by
the Oil price, the share of reserves they take, and the debt they assume. We think that the taxes the government charge will also be affected by the oil prices, oil reserves they have, debt
assumed and whether their new currency is pegged. Whether their new current is pegged will depend on the debt they have to service and the oil share they take.
The equity of the Scottish banks will depend on the approach of the new regulators and the degree of deposit flight they experience – and also the cost of funds for Scotland has a whole.
Scottish unemployment will be affected by the taxes the new government levy and the cost of funds in Scotland.
Scottish GDP will depend on the cost of Funds for Scotland and the taxes that the new government can raise.
Liquidity Risk
Credit Risk
Market Risk
Cognitive Ease / Coherent
Check the board recognises the way we think this risk
happens.
PDF
Produce a PDF for key factors based on
aggregating hard and soft data.
ST
What if scenarios … conditioning on events that we
suppose might happen - but now able to place a
probability on them!
Reverse ST
Condition on a bad event happening
and see what the world would need
to look like to make it so.
Portfolio Management under Stress
A Bayesian Net Approach to Coherent Asset Allocation
Riccardo Rebonato, Alexander Denev
Probabilistic Graphical Models in Finance
Alexander Denev – May 2015
http://www.globalgraphanalytics.co.uk/