NAEC Workshop - Complexity of the Economy: Research and Policy Implications (26-27 October 2015, Paris)

1. Network reconstruction, systemic risk,
and early-warning signals
Diego Garlaschelli
Lorentz Institute for Theoretical Physics
University of Leiden
garlaschelli@lorentz.leidenuniv.nl

2. Interbank networks:
crucial for “systemic” risk:
collapse of entire network
Public vs hidden information in networks
Public:
each bank’s total exposure towards the aggregate of all other banks
Hidden:
each bank’s individual exposure towards each single bank

3. Original network, O(N2)
(unknown/hidden)
Local properties, O(N)
(known/public)
?
Can we statistically reconstruct
the original structure in such a way that:
1) privacy is protected
2) higher-order effects are correctly predicted
The challenge:

4. New J. Phys. 16 (2014) 043022 R. Mastrandrea, T. Squartini, G. Fagiolo, D. Garlaschelli
Traditional
approach
(from “strengths” only)
Realized performance
Perfect performance
True (hidden) value
Reconstructedvalue
Average exposure of neighboring banks:

5. New J. Phys. 16 (2014) 043022 R. Mastrandrea, T. Squartini, G. Fagiolo, D. Garlaschelli
+
Traditional
approach
(from “strengths” only)
Our ‘enhanced’
reconstruction
(from “strengths” and “degrees”)
True (hidden) value
Reconstructedvalue
True (hidden) value
Reconstructedvalue
Average exposure of neighboring banks:
Realized performance
Perfect performance
Realized performance
Perfect performance

6. Arms Coffee & Tea
Plastics Optical inst.
Nucl.react. Top 14
Traditional
macroeconomics
(“Gravity” model)
Red: data - Blue: prediction

7. Arms Coffee & Tea
Plastics Optical inst.
Nucl.react. Top 14
Arms Coffee & Tea
Plastics
Optical inst.
Nucl.react. Top 14
Traditional
macroeconomics
(“Gravity” model)
Our network
model
(topology-enriched)
Red: data - Blue: prediction Red: data - Blue: prediction

8. Reducing the info
nontrivial inference from strengths and total number of links
G. Cimini, T. Squartini, A. Gabrielli, D. Garlaschelli, Phys. Rev. E 92, 040802(R) (2015)

9. Percolation
(relative size of
giant component vs
occupation probability p)
E-midWTW Reconstructing systemic risk estimators
Path length
(distribution of
shortest distances λ
among pairs of nodes)
Group DebtRank
(total devaluation induced
by an initial devaluation Φ)
[Battiston et al. 2012]
G. Cimini, T. Squartini, D. Garlaschelli, A. Gabrielli (arXiv:1411.7613, to appear on Sci.Rep.)

10. Early-warning signals in the Dutch Interbank Network
(in collaboration with De Nederlandsche Bank)
Does it always work?
T. Squartini, I. van Lelyveld, D. Garlaschelli, Sci. Rep. 3 (2013) 3357

11. 1998|1999|2000|2001|2002|2003|2004|2005|2006|2007|2008
1998|1999|2000|2001|2002|2003|2004|2005|2006|2007|2008
1998|1999|2000|2001|2002|2003|2004|2005|2006|2007|2008
… seeing
the crisis?
C = Lhomogeneous benchmark
z-score:amountofdeviation

12. 1998|1999|2000|2001|2002|2003|2004|2005|2006|2007|2008
1998|1999|2000|2001|2002|2003|2004|2005|2006|2007|2008
1998|1999|2000|2001|2002|2003|2004|2005|2006|2007|2008
1998|1999|2000|2001|2002|2003|2004|2005|2006|2007|2008
1998|1999|2000|2001|2002|2003|2004|2005|2006|2007|2008
1998|1999|2000|2001|2002|2003|2004|2005|2006|2007|2008
Foreseeing
the crisis?
C = L

C ={ ki
in
,ki
out
}
z-score:amountofdeviation
homogeneous benchmark heterogeneous benchmark

13. Heterogeneity matters!
As seen from a homogeneous benchmark,
the collapse appears sudden (abrupt transition)
As seen from a heterogeneous benchmark,
the collapse appears gradual (continuous transition)
T. Squartini, I. van Lelyveld, D. Garlaschelli, Sci. Rep. 3 (2013) 3357

14. Take-home messages
- Network reconstruction is significantly enhanced by fixing
the degrees in addition to the strengths (role of topology!)
- Far from critical events, maximum-entropy reconstruction
of interbank networks is efficient
- As crises approach, network reconstruction is inefficient
and actually prevents from detecting early-warning signals
- Still, maximum-entropy network reconstruction appears
crucial to create the early-warning signal itself.