1. Dynamic Interpretation of Emerging Systemic Risks
Kathleen Weiss Hanley1 and Gerard Hoberg2
1Lehigh University
2University of Southern California
MFM Conference March 2017
Hanley and Hoberg (2016)
Dynamic Emerging Systemic Risks

2. Special Thanks: National Science Foundation
This project was made feasible through NSF grant #
Grant was funded through CIFRAM program. A special call for projects that might benefit the Office of Financial Research (OFR).
We still know little about crises build, or how to predict and preempt them. Huge ramifications if progress can be made.
Hanley and Hoberg (2016)
Dynamic Emerging Systemic Risks

3. Special Thanks: metaHeuristica
Analytics made possible using metaHeuristica Software
Hanley and Hoberg (2016)
Dynamic Emerging Systemic Risks

4. Hypothesized Informational Environment
Suppose 3 states of the world: 1
Non-crisis periods. No information production predicted. Transition periods (we propose): Some info production. Crisis periods. Extensive information production. 2 3
Central Premise: Information producers in transition period will trade and their actions might be detectable.
Implication: Interpretable early warning system possible.
Hanley and Hoberg (2016)
Dynamic Emerging Systemic Risks

5. Theoretical Motivation
Detecting information about banks is challenging.
Efficient debt contracting “requires that no agent finds it profitable to produce costly information about the bank’s loans.” [Dang, Gorton, Holstrom, and Ordonez (2016)]
Reasons: Costly information, loan size incentives ...
Implication: Optimal bank opacity. Expect no signal in normal times.
Implication: Even in transition periods, expect weak signal. Need strong power to overcome noise. Use big data.
Hanley and Hoberg (2016)
Dynamic Emerging Systemic Risks

6. Properties of ideal predictive systemic risk model
Automated and free of researcher bias. Interpretable without ambiguity.
Can detect risks dynamically that did not appear in earlier periods.
Permits flexibility to delve deeper into topics of interest. Detects risk factors well in advance of panics.
Our approach makes significant headway on all 5 dimensions.
Hanley and Hoberg (2016)
Dynamic Emerging Systemic Risks

7. Methods: See Paper for Details
RESULT: A firm-year panel database with 18 thematic scores for each observation.
Hanley and Hoberg (2016)
Dynamic Emerging Systemic Risks

8. Most Novel Innovation: Semantic Vector Analysis
LDA alone is popular but difficult to interpret. Yet it can pick up “systemic” content.
A second stage SVA model solves the interpretability problem.
See Mikolov, Chen, Corrado, and Dean (2013) and Mikolov, Sutskever, Chen, Corrado, and Dean (2013).
We are not aware of other finance papers using this technology.
Hanley and Hoberg (2016)
Dynamic Emerging Systemic Risks

9. Semantic themes
Important: Selected themes must be present in verbal LDA factor analysis (i.e., have a strong verbal factor structure in risk disclosures). Ensures only systematically relevant risks make the list.
Hanley and Hoberg (2016)
Dynamic Emerging Systemic Risks

10. Examples of Semantic Vectors
Mortgage Risk
Capital Requirements
Row
Word
Cosine Dist 1
mortgages
capital
0.789 2
mortgage
0.7974
requirements 3
impac alt
0.7148
meet
0.5369 4
residential mortgage
0.7085
regulatory
0.4508 5
originated
0.6939
additional
0.4422 6
residential mortgages
0.6922
capital expenditure
0.4404 7
adjustable rate
0.6726
minimum
0.4278 8
collateralizing
0.6372
expenditures
0.4273 9
originations
0.6363
requirement
0.4228 10
fhlmc
0.6303
iubfsb
0.4166 11
fnma
0.6271
fund
0.4096 12
fannie mae
0.6231
liquidity
0.407 13
single family
0.6174
comply
0.4004 14
freddie mac
0.6156
ratios
0.3963 15
mbs
0.6142
regulations
0.3939 16
originate
0.6095
satisfy
0.39 17
newly originated
0.6069
required
0.3864 18
association fnma
0.606
guidelines
0.3836 19
mortgage backed
0.6052
regulators
0.3798 20
loan originations
0.6049
needs
0.3781
Hanley and Hoberg (2016)
Dynamic Emerging Systemic Risks

11. Result of Computational Linguistics: Large bank-year panel database
Note: This database is used to construct a network of bank pairwise common exposures.
Hanley and Hoberg (2016)
Dynamic Emerging Systemic Risks

12. Stock Return Covariance Matrix is Stochastic
Question: Can we use big data to examine when perturbation is likely systemic risk?
Hanley and Hoberg (2016)
Dynamic Emerging Systemic Risks

13. Does Risk Factor Network Explain Covariance Matrix Network?
Note: RHS network indicates, verbal factor by verbal factor, do banks i and j disclose the same risks with similar intensity?
Note: Verbal factors identified using LDA, and specifically from bank risks. Hence only systemically important risks are “eligible” to be part of network.
Hanley and Hoberg (2016)
Dynamic Emerging Systemic Risks

14. Our emerging risk model based on pairwise covariance
Run regression once per quarter. One observation is a bank-pair (i and j).
Dependent variable is return covariance of i and j
measured using daily returns.
Independent variable of interest is semantic theme of pair defined as the product Si,j = Si Sj
X are control variables including pairwise products: size, age, call report data, and industry.
Covariancei,j,t = α0 + γXi,j,t + εi,j,t ,
(1)
Covariancei,j,t = α0+β1Si,j,t,1+β2Si,j,t,2+β3Si,j,t,3+...+βT Si,j,t,18
+γXi,j,t + εi,j,t ,
(2)
Hanley and Hoberg (2016)
Dynamic Emerging Systemic Risks

15. Our emerging risk model II
Covariancei,j,t = α0+β1Si,j,t,1+β2Si,j,t,2+β3Si,j,t,3+...+βT Si,j,t,18
+γXi,j,t + εi,j,t ,
(3)
Goal: decompose the R2 of this model into parts related to (A) accounting variables, and (B) textual risk disclosures.
When R2 attributed to risk factors increases, a red flag IS RAISED. Dynamically interpret themes for channels driving increased R2.
Hanley and Hoberg (2016)
Dynamic Emerging Systemic Risks

16. Data Sources
We consider banks as identified by firms having SIC codes from 6000 to We exclude all other firms.
CRSP (stock returns), Compustat (accounting variables).
FDIC Failures and Assistance Transactions List. We also consider VIX data.
Call Reports for bank-specific accounting data.
metaHeuristica is used to extract risk factor discussions from bank 10-Ks from 1997 to 2014.
We require the firm to have a machine readable 10-K, with some non-empty discussion of risk factors, to be included.
Hanley and Hoberg (2016)
Dynamic Emerging Systemic Risks

17. Aggregate Systemic Risk Signal
Our Main Result: t -statistics of R2 due to textual factors 14 12 10 8 6 4 2 -2
t-statistic
Hanley and Hoberg (2016)
Dynamic Emerging Systemic Risks

18. Summary of 2008 Major Risks (t-stats)
Hanley and Hoberg (2016)
Dynamic Emerging Systemic Risks

19. Drill-down extended model
These semantic themes were chosen for more depth on marketable securities theme
Hanley and Hoberg (2016)
Dynamic Emerging Systemic Risks

20. Summary of 2015 Major Risks (t-stats)
Hanley and Hoberg (2016)
Dynamic Emerging Systemic Risks

21. Extend data to 4Q 2016 «NEW» Hanley and Hoberg (2016)
Dynamic Emerging Systemic Risks

22. Extend data to 4Q 2016 «NEW» II
Hanley and Hoberg (2016)
Dynamic Emerging Systemic Risks

23. Cross Section: Individual bank exposure to emerging risk
Define Emerging Risk Exposure: average quarterly predicted covariance bank i has with all other banks j using the main covariance model in Equation (3)
Does emerging risk exposure predict:
Bank’s quarterly stock return from 9/2008 to 3/2009 and 12/2015 to 2/2016
Dummy variable indicating whether the given bank failed in 3 years after Lehman bankruptcy
Hanley and Hoberg (2016)
Dynamic Emerging Systemic Risks

24. Predicting post-2008 crisis returns (9/2008-12/2012)
Row
Quarter
Emerging Risk Exposure #
Obs
Predictive Timing
(1)
2004 1Q
1.634 (1.13)
352
Predictive
(2)
2004 2Q
1.509 (1.38)
(3)
2004 3Q
3.892 (2.42)
368
(4)
2004 4Q
5.273 (1.99)
(5)
2005 1Q
5.898 (2.99)
391
(6)
2005 2Q
4.864 (2.60)
(7)
2005 3Q
2.571 (1.28)
421
(8)
2005 4Q
2.816 (1.05)
(9)
2006 1Q
(-1.21)
410
(10)
2006 2Q
(-1.40)
(11)
2006 3Q
(-1.27)
434
(12)
2006 4Q
(-1.27)
(13)
2007 1Q
(-2.14)
445
(14)
2007 2Q
(-1.93)
(15)
2007 3Q
(-2.83)
470
(16)
2007 4Q
(-2.50)
(17)
2008 1Q
(-3.27)
474
(18)
2008 2Q
(-3.46)
(19)
2008 3Q
(-4.59)
495
Non Predictive
(20)
2008 4Q
(-2.60)
497
(21)
2009 1Q
(-3.69)
525
(22)
2009 2Q
(-3.24)
(23)
2009 3Q
(-2.68)
536
(24)
2009 4Q
(-2.49)
529
Non-Predictive
Hanley and Hoberg (2016)
Dynamic Emerging Systemic Risks

25. Predicting current period returns (12/2015-2/2016)
Row
Quarter
Emerging Risk Exposure #
Obs
Predictive Timing
(1)
2010 1Q
(-3.44)
335
Predictive
(2)
2010 2Q
(-2.54)
(3)
2010 3Q
(-1.14)
342
(4)
2010 4Q
(-0.51)
(5)
2011 1Q
(-1.89)
353
(6)
2011 2Q
(-1.98)
352
(7)
2011 3Q
(-1.04)
358
(8)
2011 4Q
(-0.15)
(9)
2012 1Q
0.005 (0.04)
349
(10)
2012 2Q
(-0.57)
(11)
2012 3Q
0.120 (0.30)
360
(12)
2012 4Q
0.044 (0.14)
(13)
2013 1Q
(-1.55)
351
(14)
2013 2Q
(-1.55)
(15)
2013 3Q
0.155 (0.47)
368
(16)
2013 4Q
(-0.94)
(17)
2014 1Q
(-1.08)
356
(18)
2014 2Q
(-1.04)
(19)
2014 3Q
(-1.97)
367
(20)
2014 4Q
(-2.61)
(21)
2015 1Q
(-1.16)
(22)
2015 2Q
(-3.38)
(23)
2015 3Q
(-2.69)
378
(24)
2015 4Q
(-2.82)
377
Non Predictive
Hanley and Hoberg (2016)
Dynamic Emerging Systemic Risks

26. Predicting bank failures
Row
Quarter
Emerging Risk Exposure
Log Assets
Loans Assets
Loss/ Assets
Cap- ital
(1)
2004 1Q
(-0.13)
(-0.77)
0.149 (4.63)
(-2.77)
(-0.79)
(2)
2004 2Q
0.003 (1.09)
(-0.71)
0.149 (4.57)
(-1.75)
(-0.77)
(3)
2004 3Q
(-0.21)
(-0.81)
0.149 (4.66)
(-2.12)
(-0.78)
(4)
2004 4Q
(-1.58)
(-0.88)
0.150 (4.73)
(-2.24)
(-0.78)
(5)
2005 1Q
0.005 (1.01)
(-0.20)
0.160 (5.10)
(-0.30)
(-2.61)
(6)
2005 2Q
0.003 (0.78)
(-0.34)
0.159 (5.31)
(-0.35)
(-2.69)
(7)
2005 3Q
0.007 (2.41)
(-0.13)
0.159 (5.23)
(-0.05)
(-2.81)
(8)
2005 4Q
0.007 (2.32)
(-0.12)
0.159 (5.22)
0.444 (0.07)
(-2.83)
(9)
2006 1Q
0.000 (0.04)
(-1.11)
0.220 (7.17)
(-1.08)
(-2.86)
(10)
2006 2Q
0.012 (1.71)
(-0.61)
0.221 (7.13)
4.409 (1.31)
(-3.31)
(11)
2006 3Q
0.014 (2.50)
(-0.56)
0.222 (7.25)
2.109 (0.69)
(-3.45)
(12)
2006 4Q
0.019 (5.03)
(-0.17)
0.221 (7.14)
2.094 (0.52)
(-3.73)
(13)
2007 1Q
0.020 (4.92)
(-0.25)
0.224 (6.02)
(-0.01)
(-2.65)
(14)
2007 2Q
0.025 (6.90)
(-0.06)
0.223 (6.45)
7.694 (0.55)
(-2.63)
(15)
2007 3Q
0.023 (6.18)
(-0.14)
0.223 (6.21)
(-0.01)
(-2.53)
(16)
2007 4Q
0.028 (6.39)
0.219 (6.25)
3.324 (0.25)
(-2.30)
(17)
2008 1Q
0.020 (5.13)
(-0.54)
0.222 (2.47)
(-2.15)
(-0.42)
(18)
2008 2Q
0.020 (5.32)
(-0.47)
0.224 (2.48)
(-1.92)
(-0.33)
(19)
2008 3Q
0.023 (4.30)
(-0.27)
0.224 (2.55)
(-2.28)
(-0.71)
(20)
2008 4Q
0.018 (3.82)
(-0.87)
0.224 (2.52)
(-1.76)
(-0.47)
(21)
2009 1Q
0.025 (6.82)
(-1.51)
0.136 (3.77)
(-16.64)
(-2.95)
(22)
2009 2Q
0.009 (8.76)
(-2.03)
0.141 (3.83)
(-18.47)
(-3.20)
(23)
2009 3Q
0.019 (6.42)
(-2.21)
0.141 (3.41)
(-18.97)
(-3.05)
(24)
2009 4Q
0.001 (0.22)
(-2.11)
0.142 (3.93)
(-13.68)
(-3.25)
Hanley and Hoberg (2016)
Dynamic Emerging Systemic Risks

27. Conclusions
We propose a dynamic model of emerging systemic risks based on computational linguistic analysis of
financial firm disclosures and return covariances.
Benefits of model:
Provides little or no signal in “normal times”.
Provides aggregate measure of trading on systemic risks. When systemic risk is building, produces interpretable information about specific channels.
Model is dynamic and reveals risks researcher might be unaware of. Yet SVA also allows researcher to drill down.
Suggests interpretable early warning system is possible.
Also suggests SEC’s risk factor disclosure program is valuable (not a priori clear from existing work).
Hanley and Hoberg (2016)
Dynamic Emerging Systemic Risks