2014.05.19 - oecd-eclac workshop_session 1_sheri markose
TRANSCRIPT
OECD – ECLAC WORKSHOP
New tools and methods for policy making
19 May 2014
Session 1:
Multi-Agent Financial Network Models
And Global Macro-net Models:
New Tools for Macro-Prudential Policy
Sheri MARKOSE ([email protected]) Economics Dept. University of Essex,
The software used in network modelling was developed by Sheri Markose with
Simone Giansante and Ali Rais Shaghaghi
Roadmap 1 : Why MAFNs/Macro-Nets for Macro-
prudential policy ?
• Two methodological problems of financial contagion and systemic risk : (i)Paradox of Volatility and the pitfalls of market price data based systemic risk measures hence structural bilateral data based networks modeling needed (ii) Non- trivial Negative Externalities problem → the need for holistic visualization
• Some Applications: Systemic Risk From Global Financial Derivatives Modelled Using Network Analysis of Contagion and Its Mitigation With Super-Spreader Tax
• Global Macro-nets integrated with Real Side Sectoral Flow of Funds to assess extent of economic imbalance from financialization and also from global flows
• Some insights from Indian Financial System and Bilateral Data based Network Modeling: Pioneering first full bilateral digital map of financial system (Brazil and Mexico also mandating bilateral data)
Macro-prudential: Systemic Risk
Management With Networks • Holistic Visualization to overcome fallacy of composition type errors
• ICT data base driven multi-agent financial networks: glorified data
visualizations → Andrew Haldane/ Mark Buchanan Star Trek Vision
• Causal Connections v Statistical Analysis
• Minimum three elements are needed bilateral financial data
:Contracts, Counterparties, Maturity Buckets
• Integration and automation of financial data bases in a MAFN
framework aims to transform the data from a document or record
view of the world to an object-centric view (see Balakrishnan et. al.
2010), where multiple facts about the same real-world financial
entity are accessed to give a composite visualization of their
interactions with other such entities in a scalable way.
Roadmap 2: Systemic Risk Analytics
• Stability of Networks and Eigen-Pair Analysis: Markose et. al. (2012)
• 3 main questions of macro-prudential regulation :
(i)Is financial system more or less stable?
(ii)Who contributes to Systemic Risk ?
(iii)How to internalize costs of systemic risk of ‘super-spreaders’ using Pigou tax based on eigenvector centrality: Management of moral hazard, Bail in vs Bail out : How to Stabilize system using EIG Algorithm ?
• Generalization to multi-layer networks ● Conclusions
Mark ( (i )Market Data based Statistical Models of Systemic
Risk : A Case of Steadfast Refusal to Face Facts ( á la
Goodhart, 2009) ? Absence of Early Warning Signals
Major drawback of market price based systemic risk measures: they
suffer from paradox of volatility(Borio and Drehman ,2009) or paradox
of financial stability issues first addressed by Hyman Minsky (1982).
Market based statistical proxies for systemic risk (eg Segoviano and
Goodhart (2009) banking stability index, Contingent Claim Analysis
Distance to Distress Index of Castern and Kavounis (2011)) at best
contemporaneous with the crisis in markets, at worst they spike
after crisis. Laura Kodres et al IMF WP /2013/115 now call market
based systemic risk indices Coincident and Near Coincident
Systemic Risk Measures: conceded absence of early warning
capabilities.
As credit growth boosts asset prices, CDS spreads and VIX indices
which are inversely related to asset prices are at their lowest precisely
before the crash when asset prices peak → Also procyclicality of
leverage Adrian and Shin (2010, 2011).
Banking Stability Index (Segoviano, Goodhart 09/04) v
Market VIX and V-FTSE Indexes : Sadly market data based
indices spike contemporaneously with crisis ; devoid of requisite info for
Early Warning System
“Paradox of stability” : Stock Index and Volatility Index Paradox of
Volatility (Borio and Drehman(2009); Minsky (1982)) Volatility low
during boom and at local minimum before market tanks : hence misled
regulators “great moderation”
IMF WP /2013/115 Market Data Based Systemic Risk Measures:
Coincident or Near Coincident Devoid of Early Warning Few Weeks at
Best
Arsov et. al. (2013) design IMF Systemic Financial Stress (SFS, black above)
index which records the extreme negative returns at 5 percentile of the (left) tail
for the joint distribution of returns of a selected sample of large US and Eurozone
FIs (Ibid Figure 4 ) Backtesting of popular systemic risk metrics (Red, above)
(ii) Fallacy of Composition In the Generation of
Systemic Risk/Negative Externalities: Holistic
Visualization Needed
Systemic risk refers to the larger threats to the financial system as a whole that
arise from domino effects of the failed entity on others.
At the level of the individual user micro-prudential schemes appear
plausible but at the macro-level may lead to systemically unsustainable
outcomes.
Example 1 : Risk sharing in advanced economies uses O-T-C derivatives.
Success of risk sharing at a system level depends on who is providing
insurance and structural interconnections involved in the provision of
guarantees.
Only 5% of world OTC derivatives is for hedging purposes
Credit Risk Transfer in Basel 2 gave capital reductions from 8% to 1.6% capital
charge if banks got CDS guarantees from ‘AAA’ providers
Structure of Global Financial Derivatives Market:Modern
Risk Management based on Fragile Topology that Mitigates
Social Usefulness (2009,Q4 204 participants): Green(Interest Rate),
Blue (Forex), Maroon ( Equity); Red (CDS); Yellow (Commodity); Circle 16
Broker Dealers in all markets (Source Markose IMF W, 2012)
Granular Banks and Non Bank Financial Intermediaries (Dec 2012)- Note that
insurance companies (H codes) mutual funds(G codes) and not banks (A-D
codes) are net liquidity providers: Fact can be missed in banks only models !
A Financial Intermediary is member of multiple financial
markets (multi-layer networks) How to calculate its
centrality across the different networks it is present ?Joint
Eigen-vector centrality
Multi-Layer Network with
common nodes in some
layers ; m markets
Single network
Castren and Racan (ECB 2012 WP) Phenomenal Global Macro-net
Model With National Sectoral Flow of Funds To Track Global Financial
Contagion! Only Problem- the Castren-Racan Systemic Risk Analytics
Fail to have Early Warning Capabilities
The circle in the center represents banking systems that are exposed to the cross
border liabilities of sectors (household, non bank corporate, public etc) within
countries. The latter with sectoral flow of funds are given in the outer circle
Castren-Racan (2012) Loss Multiplier Systemic Risk Measure(blue
lines) vs. Markose et al (2012, 2013) Maximum Eigenvalue of Matrix of
Net Liabilities Relative to Tier 1 Capital (green line)
Castren-Racan loss multiplier (blue lines), unfortunately, peaks well after crisis has
started and asset side of FIs is considerably weakened. Markose (2012,2013) direct
measure of maximum eigenvalue, (green line)of matrix of liabilities of countries relative to
Tier1 capital of the exposed national banking systems, will capture growing instability of
network relative to distribution of capital buffers well ahead of actual crisis.
Global Macro-net plagued by within country
sectoral imbalances and global imbalances
only 8% of total lending for real investment
0%
20%
40%
60%
80%
100%
120%
140%
160%
180%Other Financial Institutions
Secured Household (banks)
Secured Household (BSocs)
Unsecured Household
LBO targets
Commercial Real Estate
'core' PNFC
Network Stability and Systemic Risk Measure:
Why Does Network Structure Matter to Stability ?
lmax = 𝑁𝐶 s < 1 Formula for network stability
• My work influenced by Robert May (1972, 1974)
• Stability of a network system based on the
maximum eigenvalue lmax of an appropriate
dynamical system
• May gave a closed form solution for lmax in terms
of 3 network parameters , C : Connectivity , number
of nodes N and s Std Deviation of Node Strength :
lmax = 𝑁𝐶 s All 3 network statistics cannot
grow and the network remain stable. Eg a highly
asymmetric network with high s such as core periphery,
its connectivity has to be very low for it to be stable
From Epidemiology : Failure of i at q+1 determined
by the criteria that losses exceed a predetermined
buffer ratio, r, of Tier 1 capital
𝑢 iq+1 = (1 - r) uiq + (𝑥𝑗𝑖−𝑥𝑖𝑗)
𝐶𝑖0
+
𝑢𝑗𝑞1
𝑗 (2)
(i)First term i’s own survival probability given by the
capital Ciq it has remaining at q relative to initial capital
Ci0 , r is common cure rate and (1 - r) is rate of not
surviving in the worst case scenario .
(ii) The sum of ‘infection rates’= sum of net liabilities of
its j failed counterparties relative to its own capital is
given by the term (𝑥𝑗𝑖
−𝑥𝑖𝑗
)
𝐶𝑖0
+
𝑗
Stability of the dynamical network
system : Eigen Pair (λmax , v)
In matrix algebra dynamics of bank failures given by:
Ut +1 = [´ + (1- r)I] Ut = Q Ut (3)
I is identity matrix and r is the % buffer
The system stability of (2) will be evaluated on the basis
of the power iteration of the matrix Q=[(1-r)I+Θ´]. From
(3), Uq takes the form:
Uq= Qq U0
Stability Condition lmax(´) < r
After q iterations
λmax is maximum
eigenvalue of Θ
Solvency Contagion and Stability of Matrix Θ’
: Netted impact of i on j relative to j’s capital
)2(
0...)(
....)(
.0........
)(...0....
)(.........0..
)(........
)(00
0.....0.)()(
0
1
11
1
11
33
3
3223
3
3113
2
2112
jt
jNNj
t
NN
Nt
NiiN
t
ii
Nt
NN
t
tt
C
xx
C
xx
C
xx
C
xx
C
xx
C
xx
C
xx
C
xx
Financial network models to date have yielded mixed
results : None about propagators of 2007 crisis (C:
Core; P Periphery (see Fricke and Lux (2012))
Some Networks: A graphical representation of random
graph (left) and small world graph with hubs, Markose
et. al. 2004
Contagion when JP Morgan Demises in Clustered CDS Network 2008
Q4 ( Left 4 banks fail in first step and crisis contained) v
In Random Graph (Right 22 banks fail !! Over many steps)
Innoculate some key players v Innoculate all ( Data Q4 08)
Eigenvector Centrality (EVC)
Centrality: a measure of the relative importance of a node
within a network
Eigenvector centrality Based on the idea that the centrality vi of a node should be proportional to
the sum of the centralities of the neighbors
l is maximum
eigenvalue of Θ
A variant is used in the Page Ranking algorithm used by Google
The vector v, containing centrality values of all nodes is obtained by solving the
eigenvalue equation Θ 𝒗𝟏 = λmax 𝒗𝟏 .
λmax is a real positive number and the eigenvector 𝒗𝟏 associated with the largest
eigenvalue has non-negative components by the Perron-Frobenius theorem (see
Meyer (2000))
Right Eigenvector Centrality : Systemic Risk Index Tax using Right EVC
Left Eigenvector centrality Leads to vulnerability Index
Mitigation of Systemic Risk Impact of
Network Central Banks: How to
stabilize ?
To date the problem of how to have banks internalize
their systemic risk costs to others (and tax payer) from
failure has not been adequately solved
In particular, penalty for being too interconnected has
not been dealt with from direct bilateral network data
Tax according to right eigenvector centrality
There are 5 ways in which stability of the
financial network can be achieved
(i)Constrain the bilateral exposure of financial
intermediaries (Ad hoc constraints do not work) Serafin
Martinez implemented these in Mexico
(ii) Ad hocly increase the threshold rho in (11),
(iii) Change the topology of the network
(iv) Directly deduct a eigenvector centrality based
prefunded buffer in matrix
Si # = 𝜃𝑗𝑖
#𝑗 =
1
𝐶𝑖 ( (𝑥𝑗𝑖𝑗 − 𝑥𝑖𝑗)
+ − vi 𝐶𝑖) .
(i) & (ii) do not price in negative externalities and systemic
risk of failure of highly network central nodes. Network
topologies emerge endogenously and are hard to manipulate
exogenously.
How to stabilize: Superspreader tax
quantified : tax using Eigen Vector Centrality
of each bank vi or vi
^2 to reduce max
eigenvalue of matrix to 6%
Superspreader tax rate
• Too interconnected to fail addressed only if systemic risk from individual banks can be rectified with a price or tax reflecting the negative externalities of their connectivity
• Lessons to be learnt : Disease Transmission in scale free networks (May and Lloyd (1998), Barthelemy et. al : With higher probabilities that a node is connected to highly connected nodes means disease spread follows a hierarchical order. Knowledge of financial interconnectivity essential for targeted interventions
• Highly connected nodes become infected first and epidemic dying out fast and often contained in first two tiers
• Innoculate a few rather than whole population; Strengthen hub; Reduce variance of node strength in maximum eigenvalue formula
Conclusion : Regulators and Systemic Risk Researchers must face up
to limits of data mining market price data for early warning signals and
instead mandate structural bilateral financial data and digitally map the
macro-financial network system
• Changes in eigenvector centrality of FIs can give early warning of instability causing banks
• EVC basis of Bail in Escrow fund/Capital Surcharge
• These banks with high Eigen vector centrality will, like Northern Rock, be winning bank of the year awards ; however potentially destabilizing from macro-prudential perspective Capital for CCPs to secure system stability can use same calculations
• Beware gross aggregation and netting across product classes for which there is no multi-lateral clearing ; more systemic risk and hence more, capital/collateral needed for stabilization
• Single network v multi-layer networks
Other Concluding Remarks
References:
(1) Markose, S.M (2013) "Systemic Risk Analytics: A Data
Driven Multi-Agent Financial Network (MAFN) Approach",
Journal of Banking Regulation , Vol.14, 3/4, p.285-305, Special
Issue on Regulatory Data and Systemic Risk Analytics. (2) Markose, S. (2012) “Systemic Risk From Global Financial
Derivatives: A Network Analysis of Contagion and Its Mitigation With
Super-Spreader Tax”, November, IMF Working Paper No. 12/282,
(3) Markose, S., S. Giansante, and A. Shaghaghi, (2012), “Too
Interconnected To Fail Financial Network of U.S. CDS Market:
Topological Fragility and Systemic Risk”, Journal of Economic
Behavior and Organization, Volume 83, Issue 3, August 2012, P627-
646
http://www.sciencedirect.com/science/article/pii/S0167268112001254
(3)Multi-Agent Financial Network (MAFN) Model of US Collateralized
Debt Obligations (CDO): Regulatory Capital Arbitrage, Negative CDS
Carry Trade and Systemic Risk Analysis, Sheri M. Markose, Bewaji
Oluwasegun and Simone Giansante, Chapter in Simulation in
Computational Finance and Economics: Tools and Emerging
Applications Editor(s): Alexandrova-Kabadjova B., S. Martinez-
Jaramillo, A. L. Garcia-Almanza, E. Tsang, IGI Global, August 2012.
http://www.acefinmod.com/CDS1.html