2. Financial Computing and Analytics Group
T Aste UCL Sept 2018
Tomaso Aste
Philip TreleavenDenise Gorse
Fabio CaccioliChristopher ClackGuido Germano Giacomo Livan Daniel Fricke Julius Bonart Robert Elliott Smith
Tiziana Di Matteo Paolo Tasca Ariane Chapelle
http://fincomp.cs.ucl.ac.uk/
Nick Firoozye
Daniel Hulme
Jessica James
• Centre for Doctoral Training
• in Financial Computing & Analytics
• Financial Risk Management MSc
• Computational Finance MSc
• Business Analytics MSc
• Centre for Bolckchain Technologies
Mircea Zloteanu
3. • A lot of them
• Interrelated
• Complex
• Non Stationary
• Unstructured
• Inconsistent
Ø High Dimensions,
hard to parametrize
Ø Multivariate models
Ø Heterogeneous: hard to put
together
Ø Small reliable observation set
Ø Hard to make predictions
and validate
Ø Hard to establish the “truth”
Artificial Intelligence
Sparse multivariate probabilistic
predictive models adaptable and
trainable from few training sets
Networks
Complex multilayered networks
to filter information for sparse
modeling with heterogeneous
data
Blockchain
Community consensus to
establish data consistency
Data Filtering Model & Prediction
T Aste
4. Machine Learning
Machine learning can be described as a set of
methodologies to automatize the process of
observation/modeling/predicting/testing
To discover laws automatically form (large amount of) data
If something
happens here
What are the
consequences
there?
Prediction
Parsimony
Artificial intelligence aims to find automatically
maps between sets of variables.T Aste UCL Sept 2018
5. A BW
Predictive modeling
In complex systems these maps are conditional
probabilities.
Prediction is the estimation of the conditional
probability ! "# "$, "& of a (future) event given
the available information about other (past) events
T Aste UCL Sept 2018
6. Predictive modeling
The conditional probability
is a tool for:
- test hypothesis
- quantify risk
- stress testing
- analyze scenarios
XA
− XB
p(XB
| XA
,XW
)
T Aste UCL Sept 2018
7. Predictive modeling
The predicted value is the expectation value
E[XB
| XA
−
,XW
−
]= XB
XB
∑ p(XB
| XA
−
,XW
−
)
The predicted uncertainty is the conditional entropy
H(XB
|XA
−
,XW
−
)=− p(XB
|XA
−
,XW
−
)logp(XB
|XA
−
,XW
−
)
XA ,XB
∑
This is the regression and for linear models (multivariate
Gaussian) this is the linear regression formula
T Aste UCL Sept 2018
8. Predictive modeling
Causality
Is the reduction of uncertainty on variables XB given
the knowledge of the past of variables X
-
A and X
-
W
discounting for the past X-
B
H(XB
|XB
−
,XW
−
)− H(XB
|XA
−
,XB
−
,XW
−
)=TE(XA
→ XB
|XW
−
)
B
B-
A-
W-
Thisis the tranfer entropy that for liner models
(multivariate Gaussians) coincides with
Granger causality
T Aste UCL Sept 2018
9. Predictive modeling
Big Data – High Dimensions
! "# "$, "& =
! "#, "$, "&
! "$, "&
We must estimate from data the most likely joint
probability ! "#, "$, "& distribution of the system
of events
T Aste UCL Sept 2018
10. To estimate probability we must estimate a
number of quantities between
High dimensional problem!
N2
2
≤ Quantities to Estimate ≤ N!
Information increases linearly with observations
But model-parameters increase at least with the
square!
“The more you look the less you see”T Aste UCL Sept 2018
11. Information filtering
We must construct models that
require the estimation of a
smaller number of quantities
T Aste UCL Sept 2018
12. Graphical models
To construct the joint multivariate distribution we can
make use of the structure of conditional dependency
p(XA
,XB
|XW
)≠ p(XA
|XW
)p(XB
|XW
)
XW
= XAll Variables
{XA
,XB
}
A B
dependent
p(XA
,XB
|XW
)= p(XA
|XW
)p(XB
|XW
)
A B
independent
T Aste UCL Sept 2018
13. Graphical models
If these inference networks are chordal (or
decomposable) we then have
p(X) =
p(Xcliques
)
cliques
∏
p(Xseparators
)
ks−1
separators
∏
The joint probability distribution can
be estimated form the probability of
cliques and separators
IF THE NETWORK IS SPARSE
Quantities to estimate scale
linearly with the number of
variables!
S. L. Lauritzen, Graphical Models (Oxford:Clarendon, 1996)T Aste UCL Sept 2018
14. This is great… however to establish conditional
dependency
p(XA
,XB
| XW
) ≠ p(XA
| XW
)p(XB
| XW
)
requires the estimation of the same number of quantities
as computing the entiere joint distribution function!
Building the exact inference network is an impossible
task for a large number of variables
Graphical models
T Aste UCL Sept 2018
15. Information filtering networks
To solve this problem we propose to build the inference
structure for the graphical model from
Information filtering network
• Massara, Guido Previde, Tiziana Di Matteo, and TA. "Network Filtering for Big Data: Triangulated Maximally Filtered Graph"
Journal of Comlex Networks (2016) arXiv preprint arXiv:1505.02445 (2015).
• Nicoló Musmeci,, Tomaso Aste, and Tiziana Di Matteo. "Relation between financial market structure and the real economy:
comparison between clustering methods." PloS one 10.3 (2015): e0116201.
• F. Pozzi, T. Di Matteo, and TA , “Spread of risk across financial markets: better to invest in the peripheries”, Scientific Reports 3
(2013) 1665.
• W.M. Song, T. Di Matteo and T. Aste, “Hierarchical information clustering by means of topologically embedded graphs”, PLoS
ONE, 7 (2012) e31929
- TA, T. Di Matteo and S. T. Hyde, Complex networks on hyperbolic surfaces Physica A 346 (2005) 20-26.
- M. Tumminello, TA, T. Di Matteo, and R. N. Mantegna, “A tool for filtering information in complex systems”
Proceedings of the National Academy of Sciences of the United States of America 102 (2005) 10421 .
T Aste UCL Sept 2018
16. Information filtering networks
Connect the nearest vertices
eucleadean distance = most correlated
hyperbolic distance = mutual information
Keep the graph chordal
clique forests
Add other constraints
max clique size (2 = MST)
planarity (TMFG)
information criteria (e.g. Akaike)
These are fast algorithms < O(N2)
(topological & homological measures, betty numbers, cycles and cliques retrieved from construction)
Massara, Guido Previde, Tiziana Di Matteo, and TA. "Network Filtering for Big Data: Triangulated Maximally Filtered Graph"
Journal of Complex Networks (2016) arXiv preprint arXiv:1505.02445 (2015).T Aste UCL Sept 2018
18. By constraining the model to reproduce observed
moments while maximizing Shannon-Gibbs entropy
(maximum Entropy method), at the second order, we
have that the model must be a multivariate Gaussian:
LoGo
p(X1
,...XN
) =
1
Z
exp(− Xi
Ji,j
Xj
i, j
∑ )
We keep only the significant interactions
and set to zero (Max Ent.) the uncertain
ones: Ji,j = 0 iff Xi , Xj conditionally
independent
Ji,j is sparse and it has the structure given
by the information filtering network
T Aste UCL Sept 2018
19. Ji,j is computed form local inversion of the covariance
matrix over the clique forest
Ji,j
= Σ(C)−1
Cliques C
∑ − (kS
−1)Σ(S)−1
Separators S
∑
We obtain a sparse inverse covariance (our graphical
model) by doing local inversions only
Super-fast algorithm O(N) even O(logN) if parallelized
LoGo
W. Barfuss, GP Massara, T Di Matteo & TA “Parsimonious modeling with Information Filtering Networks” Phys Rev E 94, 062306
(2016). arXiv preprint arXiv:1602.07349 (2016).
T Aste UCL Sept 2018
20. SPARSE CAUSALITY NETWORK RETRIEVAL FROM SHORT TIME SERIES
RidgeSignificant
G-lassoSigificant
LoGoSignificant
0 1 2 3 4 5
p/q
0
0.1
0.2
0.3
0.4
0.5
γ
Ridge
Glasso
LoGo
p=100 variables
Test: Network Retrieval
From simulated time
series with a known
causality structure we
test how LoGo can
discover the causality
links
is our model,
constructed from
observations, able to
discover the ‘real’
causality structure?
T Aste UCL Sept 2018
21. Test: Network Retrieval
From simulated time
series with a known
causality structure we
test how LoGo can
discover the causality
links
p=100 variables
8 SPARSE CAUSALITY NETWORK RETRIEVAL FROM SHORT TIME SERIES
0 2 4 6 8 10
0
0.5
γ
ridge
0
0.2
0.4
0 2 4 6 8 10
0
0.5
γ
Glasso
0
0.2
0.4
0 1 2 3 4 5
p/q
0
0.5
γ
LoGo
0
0.2
0.4
> 0.5
> 0.5
> 0.5
TP/n
TP/n
TP/n
is our model,
constructed from
observations, able to
discover the ‘real’
causality structure?
T Aste UCL Sept 2018
22. p=100 variables
Test: Network Retrieval
From simulated time
series with a known
causality structure we
test how LoGo can
discover the causality
links
is our model,
constructed from
observations, able to
discover the ‘real’
causality structure?
T Aste UCL Sept 2018
24. 10-3
10-2
Uncertanty (= 1/data series length)
650
700
750
800
Prediction(=Log-Likelihood)
LoGo - TMFG
State-of-the-art sparse model (G-lasso)
LoGo - MST
Linear Model
Test: Prediction Likelihood
is our model, constructed from past observations, associated
with a large likelihood for future observations?
T Aste UCL Sept 2018
25. Predictive modeling:
can we predict uncertainty propagation in the global
banking system?
10 Banking Regions
(World Bank Organization)
North America
South America
Africa
Europe
West Asia
Middle East
South Asia
East Asia
Southeast Asia
Australia
4 Industry
classifications
Banks
Diversified
Financial
Insurance
Real Estate
T Aste UCL Sept 2018
26. The predicted volatility on variables B caused by past of
variables A discounting the contribution from past of B
H(XB
|XB
−
)−H(XB
|XB
−
,XA
−
)=TE(XA
→XB
)
Predictive modeling:
can we predict uncertainty propagation in the global
banking system?
TRANSFER ENTROPY
TE(XA
→XB
)=
1
2
log
Var(XB
|XB
−
)
Var(XB
|XB
−
,XA
−
)
For linear modeling it is the contribution of variable A (past) to the
variance of variable B unexplained form past B
T Aste UCL Sept 2018
27. Predictive modeling:
can we predict uncertainty propagation in the global
banking system?
N America
S America
Africa
Europe W Asia
Mid East
S Asia
E Asia
SE Asia
Aus/NZ
N America
S America
1990-1992
T Aste UCL Sept 2018
28. Predictive modeling:
can we predict uncertainty propagation in the global
banking system?
N America
S America
Africa
Europe W Asia
Mid East
S Asia
E Asia
SE Asia
Aus/NZ
N America
S America
1990
N America
S America
Africa
Europe W Asia
Mid East
S Asia
E Asia
SE Asia
Aus/NZ
N America
S America
1995-1997
T Aste UCL Sept 2018
29. Predictive modeling:
can we predict uncertainty propagation in the global
banking system?
N America
S America
Africa
Europe W Asia
Mid East
S Asia
E Asia
SE Asia
Aus/NZ
N America
S America
1990
N America
S America
Africa
Europe W Asia
Mid East
S Asia
E Asia
SE Asia
Aus/NZ
N America
S America
1995
N America
S America
Africa
Europe W Asia
Mid East
S Asia
E Asia
SE Asia
Aus/NZ
N America
S America
2000-2002
T Aste UCL Sept 2018
30. Predictive modeling:
can we predict uncertainty propagation in the global
banking system?
N America
S America
Africa
Europe W Asia
Mid East
S Asia
E Asia
SE Asia
Aus/NZ
N America
S America
1990
N America
S America
Africa
Europe W Asia
Mid East
S Asia
E Asia
SE Asia
Aus/NZ
N America
S America
1995
N America
S America
Africa
Europe W Asia
Mid East
S Asia
E Asia
SE Asia
Aus/NZ
N America
S America
2000
N America
S America
Africa
Europe W Asia
Mid East
S Asia
E Asia
SE Asia
Aus/NZ
N America
S America
2005-2007
T Aste UCL Sept 2018
31. Predictive modeling:
can we predict uncertainty propagation in the global
banking system?
N America
S America
Africa
Europe W Asia
Mid East
S Asia
E Asia
SE Asia
Aus/NZ
N America
S America
1990
N America
S America
Africa
Europe W Asia
Mid East
S Asia
E Asia
SE Asia
Aus/NZ
N America
S America
1995
N America
S America
Africa
Europe W Asia
Mid East
S Asia
E Asia
SE Asia
Aus/NZ
N America
S America
2000
N America
S America
Africa
Europe W Asia
Mid East
S Asia
E Asia
SE Asia
Aus/NZ
N America
S America
2005
N America
S America
Africa
Europe W Asia
Mid East
S Asia
E Asia
SE Asia
Aus/NZ
N America
S America
2010-2012
T Aste UCL Sept 2018
32. Predictive modeling:
can we predict uncertainty propagation in the global
banking system?
N America
S America
Africa
Europe W Asia
Mid East
S Asia
E Asia
SE Asia
Aus/NZ
N America
S America
1990
N America
S America
Africa
Europe W Asia
Mid East
S Asia
E Asia
SE Asia
Aus/NZ
N America
S America
1995
N America
S America
Africa
Europe W Asia
Mid East
S Asia
E Asia
SE Asia
Aus/NZ
N America
S America
2000
N America
S America
Africa
Europe W Asia
Mid East
S Asia
E Asia
SE Asia
Aus/NZ
N America
S America
2005
N America
S America
Africa
Europe W Asia
Mid East
S Asia
E Asia
SE Asia
Aus/NZ
N America
S America
2010
N America
S America
Africa
Europe W Asia
Mid East
S Asia
E Asia
SE Asia
Aus/NZ
N America
S America
2014-2016
T Aste UCL Sept 2018
33. 1995 2000 2005 2010 2015
0
0.1
0.2
0.3
0.4
0.5
0.6
0.7
0.8
0.9
1
NA
Change of influence
North America
Centrality proxy computed
as the average fraction
between strength of
vertices in the regions
outside NA associated
with vertices in NA (black
arcs) with respect to the
total strength associated
with vertices outside the
firm’s regions (black and
red arcs)
NA
T Aste UCL Sept 2018
34. 1995 2000 2005 2010 2015
0
0.1
0.2
0.3
0.4
0.5
0.6
0.7
0.8
0.9
1
EU
NA
Change of influence
Europe
North America
T Aste UCL Sept 2018
35. 1995 2000 2005 2010 2015
0
0.1
0.2
0.3
0.4
0.5
0.6
0.7
0.8
0.9
1
ASE
EU
NA
Change of influence
East Asia
Europe
North America
T Aste UCL Sept 2018
36. 1995 2000 2005 2010 2015
0
0.1
0.2
0.3
0.4
0.5
0.6
0.7
0.8
0.9
1
ASE
ASS
EU
ME
NA
Change of influence
East Asia
Europe
North America
South Asia
Middle East
T Aste UCL Sept 2018
37. 1995 2000 2005 2010 2015
0
0.1
0.2
0.3
0.4
0.5
0.6
0.7
0.8
0.9
1
ASE
ASS
ASSE
EU
ME
NA
Change of influence
East Asia
Europe
North America
South Asia
Middle East
South East Asia
T Aste UCL Sept 2018
38. 1995 2000 2005 2010 2015
0
0.1
0.2
0.3
0.4
0.5
0.6
0.7
0.8
0.9
1
AF
ASE
ASS
ASSE
AUS
EU
ME
NA
RU
SA
Change of influence
East Asia
Europe
North America
South Asia
Middle East
South East Asia
T Aste UCL Sept 2018
39. 2006 2007 2008 2009 2010 2011 2012 2013 2014 2015 2016
0
0.002
0.004
0.006
0.008
0.01
0.012
0.014
0.016
Average TE Region -> Bank: LoGo, w 300 3 lagsSubprimemortgagecrisis
Securitizationmarketclosedown
Globalstockmarketsharpfall
USnear-recorddeficit$410billion
NationalizationofNorthernRock
FannieMae/FreddieMacrescue
LehmanBrothersfiledforbankruptcy
BankofEnglandssurprisedinterestratecut
RescueofRBS,Lloyds,andHBOS
IMFapproved$2.1bnloanforIceland
USgovtgaveBankofAmerica$20bnaid
RBSreported£2.1bnloss
12.5%economiccontractioninJapanEuropeansovereigndebtcrisis
UScreditdowngradefromA+toALIBORscandal
CityofDetroitbankruptcyUkranian/Syrian/EgyptunrestEbolaepidemic
BearSternscollapse
IndyMaccollapse
CreditdowngradeofFranceandAustriafromAAAtoAA+
PortugalreceivedbailoutLiquidityinjectionbyUS,CAN,JAP,UK,Swisscentralbanks
NA
EU
AS
Dynamic impact
T Aste UCL Sept 2018
40. Influence by activity
1995 2000 2005 2010 2015 2020
0
0.2
0.4
0.6
0.8
1
Banks
Diversified Financials
Insurance
real estate
NA
1995 2000 2005 2010 2015 2020
0
0.1
0.2
0.3
0.4
0.5
Banks
Diversified Financials
Insurance
real estate
EU
NA EU
T Aste UCL Sept 2018
41. Influence by activity
1995 2000 2005 2010 2015 2020
0
0.02
0.04
0.06
0.08
0.1
0.12
Banks
Diversified Financials
Insurance
real estate
ME
1995 2000 2005 2010 2015 2020
0
0.02
0.04
0.06
0.08
0.1
0.12
0.14
0.16
Banks
Diversified Financials
Insurance
real estate
ASS
Middle East South Asia
T Aste UCL Sept 2018
42. Networks for Prediction
With p(XB|XA-) we can quantify probability of future events
With p(XB|XA) we can predict impact of unobserved scenarios
and test hypothesis
p(XA,XB) can be constructed from local
probability estimations over an
information filtering network (low dimension problem)
Nodes and edges can be added or removed with local moves
only
Aggregation of risk is straightforward
LoGo works better than state-of-the-art sparse graphical
models and it is faster
T Aste UCL Sept 2018
43. Si l’ordre satisfait la raison, le désordre fait les délices de l’imagination
http://www.cs.ucl.ac.uk/staff/tomaso_aste/
Thank YOU!
http://fincomp.cs.ucl.ac.uk/
W. Barfuss, GP Massara, T Di Matteo & TA
“Parsimonious modeling with Information Filtering Networks” Phys Rev E 94,
062306 (2016) arXiv preprint arXiv:1602.07349 (2016).
Massara, Guido Previde, Tiziana Di Matteo, and TA. "Network Filtering for Big
Data: Triangulated Maximally Filtered Graph" Journal of Comlex Networks
(2016) arXiv preprint arXiv:1505.02445 (2015).
http://blockchain.cs.ucl.ac.uk/
Tiziana Di Matteo, and TA. ”Sparse causality network retrieval from short time
series” Complexity (2017).
T Aste UCL Sept 2018