2019 IEEE International Conference on Data Mining (ICDM)
An Integrated Multimodal Attention-Based
Approach for Bank Stress Test Prediction
1st Farid Razzak 2nd Fei Yi 3rd Yang Yang 4th Hui Xiong
Rutgers University Northwestern Polytechnical University Nanjing University Rutgers University
farid.razzak@rutgers.edu yifeinwpu@gmail.com yangy@lamda.nju.edu.cn hui.xiong@business.rutgers.edu
Abstract—Since the financial crisis in late 2008-2009, several TABLE I: Data Summary for Xi , Yi (1990-2017)
global regulatory authorities have mandated stress-testing ex- Type Variable Mean Std. Med. Max
Comm. & indtrl 5.2 13.0 1.2 205.0
ercises to evaluate the potential capital shortfalls & systemic Constr. & ld dev.
Consr.(excl.CredC)
10.5
13.1
2.1
114.6
.43
.22
275.5
1710.4
impacts that large banks may face during adverse economic con- Xloancat (In Mn USD)
CredCrd
HELOCs
3.9
1.7
13.4
6.6
.007
.24
162.4
121.7
ditions. Thus, having the ability to analyze economic conditions & Multifam-RealEstate
Non-farm-nonres-CRE
.57
5.1
2.0
2.4
.14
1.3
6.9
798.6
Res-RealEstate(excl.HELOCs) 2.2 16.1 1.4 372.4
banking performance profiles together to determine relationships Comm. & indtrl
Constr. & ld dev.
1.1
2.2
1.9
4.7
.44
.34
11.0
25.9
among their respective features may provide insights for stress- Multifam-RealEstate
Consr.(excl.CredC)
.93
3.6
2.0
5.8
.11
1.2
12.2
35.2
YncoR (0-100%)
testing tasks. CredC
HELOCs
8.3
1.7
1.0
1.7
4.1
.15
54.8
12.0
Non-farm-nonres-CRE 1.9 4.3 .15 23.1
In this paper, we propose an Integrated Multimodal Bank Res-RealEstate(excl.HELOCs) .80 1.9 .13 12.5
Comb-LoanLoss 2.4 2.1 1.7 14.7
Stress Test Prediction (IMBSTP) model framework consisting of Net-interest income
Non-interest income
4.6
2.8
6.6
5.1
2.3
0.9
35.9
30.7
a two-stages; (1) economic conditions estimator to approximate Yppnr (0-100%)
Trading income
Compensation expense
0.2
2.8
0.7
4.7
.001
1.16
4.9
26.5
joint representation among the exogenous factors using genera- Fixed assets expense
Non-interest expense
.66
5.3
1.0
9.6
.29
1.61
6.3
53.7
T1 Common Equity 16.9 24.0 13.0 66.0
tive models, (2) bank capital & loss forecaster to project stress- YC apRatio (%)
T1 Risk 17.0 27.4 14.0 70
Ttl-Risk 18.2 28.1 16.0 72.0
test measures based on dimensional & temporal features selected T1 Leverage 5.8 20.9 9.0 37.0
from the exogenous economic conditions & banking performance
profiles using a dual-attention recurrent neural network.
Extensive experimentation is performed on historical economic to project bank capital & loan loss ratios in estimated eco-
conditions & consolidated financial statements of U.S. bank
holdings companies to show the effectiveness of our approach nomic conditions. Particularly, we incorporate additional ex-
when compared to state-of-the-art baseline methods. ogenous factors beyond regulator standards to depict economic
conditions & learn non-linear latent relationships among rele-
Keywords-Deep Learning, Multimodal Conditional Generative
Models, Recurrent Neural Networks, Bank Stress-Test vant economic & financial market indicators that can be used
to understand banking performance. The contributions of this
I. I NTRODUCTION paper can be summarized as follows:
The recent financial crisis & it’s ensuing economic reces- • An integrated model framework that robustly consolidates
sion in the United States have brought major advances to multimodal economic conditions estimation & banking
capital adequacy requirements, such as testing the solvency capital & loss prediction based on dimensional & tempo-
of banks through hypothetically adverse economic scenarios ral importance for stress-test related tasks.
by examining expected performance in said dire conditions. • Additions to the state of the art literature by exploring the
Effectively, regulators want to ensure large banks are providing use of deep learning techniques to understand bank stress-
loans with considerations for economic & systemic risks that test properties [1], [3], [4] using generative & non linear
could impact bank solvency. However, this regulatory strategy auto-regressive exogenous neural network modeling.
is counter to the profit generation goals of the bank, which can
create a unique circumstance of conflicts that warrant regula- II. P ROBLEM S TATEMENT
tory oversight. Effective stress-testing can provide insights &
A. Preliminaries
potential mitigation strategies to prevent catastrophic losses by
financial institutions during severe economic conditions [1]. Definition 1 (Economic Conditions) In a top-down approach
Currently, the Federal Reserve’s Comprehensive Capital to stress-testing, exogenous economic factors, in our case
Analysis & Review (CCAR) is the most recognized regulatory ECOmod = [Zmacro , Zmicro , MSP , MF CI ], are believed to
exercise in the U.S [2]. An adverse outcome from the exercise have influence on the trajectory of a bank’s performance.
may serve as an industry-wide signal for counter-party & Currently, the historical macro-economic variables that de-
systemic risk as well as have repercussions for the bank that pict the U.S. & parts of the global financial economy consists
may include restrictions on a firm’s capital distribution plan of domestic & international variables, ZmacroI [2], [4].
[3] , rectification of compliance risk program, & renewal of Historical micro-economic variables represent specific as-
loss projections meeting regulatory thresholds [2]. pects of the financial economy. To this end, US treasuries,
In this paper, we propose the Integrated Multimodal Bank inflation rates, major commodity indices, stock indices returns,
Stress Test Prediction (IMBSTP) model framework, designed government bond rates, interest rate swaps, currency swap
2374-8486/19/$31.00 ©2019 IEEE 1282
DOI 10.1109/ICDM.2019.00161
, rates, & major commodities prices are collected quarterly
between 1976-2017 to depict the micro-economy, Zmicro .
Historical sector-based indices, MSPF IN , are collected over
the same time period & frequency as previous variables to
depict the financial & real-estate sectors, since initial signals of
the most recent financial crisis were first noticed at mentioned
sectors. The S&P 500 Financial Sector Index consists of three
tickers that illuminate the conditions specific to the financial
industry, while the S&P Real Estate Indices,MSPRE , consists Fig. 1: The overall framework of proposed IMBSTP model.
of tickers that tracks the real estate industry.
Several industry indices that depict the overall financial
conditions of U.S banking & financial economy are also (XEqP O ), payment of taxes (τ = 35%), & regulatory capital
collected. These indices & sub-indices, MF CI , are developed deductions (XRegDt ). The calculations can be summarized as
by the regional Federal Reserve Boards as well as corporate XEqCapt = XEqCapt−1 + (1 − τ ) ∗ N
etrevlosst − XEqP Oi,t−1 ,
research entities to capture the directional conditions in money where N etrevlosst = ( j Yppnrit − j YncoRit ).
markets, debt markets, equity markets & traditional shadow The remaining capital in proportion to the bank’s previous
banking systems [5]. risk weighted assets, XRW At−1 ), is considered to be the
Definition 2: (Banking Performance Profile) Top-down stress- capital ratio (YT 1CR ), a measure that best depicts the bank’s
testing relies on publicly released bank financial statements to overall capital adequacy. Regulators are most interested in
assess bank loan portfolios, loan loss rates & net revenues, to the Tier-1 common ratio, which only considers bank equity
XEqCap −XRegDtt−1
determine capital adequacy. elements as part of the capital, YT 1CR = XRW A .
t−1
For the purposes of our study, the regulatory codes &
calculations discussed in [3] are used to acquire corresponding
features. B. Problem Formulation
Loan portfolios breakdown seven major loan categories, The prediction tasks can be separated into two stages:
Xloancati,j , of bank holding companies acquired from their 1) Economic Conditions Estimation: Given a set of eco-
respective consolidated financial statements through ”Bank nomic conditions, ECOmod , the task is to estimate fu-
Regulatory” dataset [6], a summary of which can be seen ture conditions by learning the true joint probability dis-
in Table I. The loan categories represent a snapshot view of tribution among all the variables within each modality,
the bank’s lending practice to different segments of borrowers Pr(ECO ˆ mod ). This representation of the overall economic
whom could be impacted by economic conditions & therefore ˆ mod ), allows for both the prediction of
conditions, Prθ (ECO
affect the bank’s risk exposure. the most likely upcoming economic conditions & sampling
Determining insight from Xloancati,j , where i is an individ- of economic conditions from different probability densities to
ual bank & j is a loan category, by examining the temporal acquire more dire but plausible conditions.
evolution from t − 1 to t can provide details into the bank’s In our approach, normalization & principal components
growth & loss rates in conjunction with economic conditions, dimension reduction, ECO ˆ mod = [ŵ1 , ŵ2 , ...ŵj ] ∀wj ∈
ECOmod,t . ECOmod , of each modality in ECOmod is performed prior
The net losses in loan categories are depicted by the net- to estimation for the purposes of scaling numerical values &
charge-off amounts, XN COi,j which considers both losses obtaining representative variables.
& recoveries from each respective loan category to then 2) Bank Capital & Loss Ratio Prediction: Given the
XN CO set of banking performance profiles, [YncoR , Yppnr , YT 1CR ],
determine the loan loss rates, YncoRi,j,t = 100 ∗ Xloancat i,j .
i,jt−1
banking loan portfolios, Xloancat , economic conditions,
Banks are able to generate revenue from interest earning ECOˆ mod & economic estimations from the previous stage,
loans, trading income & other revenue-generating services, Prθ (ECOˆ mod ), the task of predicting future bank capital,
however they also have to consider expenses such as compen- loan loss & net revenue is a function of each target variable’s
sation, fixed assets, & other non interest earning operations, past with consideration for economic conditions estimations,
XCmpppnri,j . To determine the net revenue proportional to the ECOˆ mod .
t
bank’s consolidated assets, XCnsldAstsi,t−1 , we can derive the
Forecasting banking capital & performance is typically
pre-provisional net revenue ratio of the bank, Yppnri,jt [3]. Un-
derived from projected loan loss & revenue [1], [3], as:
derstanding the bank’s ability to generate revenue during dy-
XCmpppnr
namic economic conditions, Yppnri,jt = 100 ∗ XCnsldAsts i,j P r(YncoRt , Yppnrt , YT 1CRt |YncoRt−1 , Yppnrt−1 , Xloancatt−1 ,
i,t−1
, can be crucial to offset losses for more appropriate forecasts. YT 1CRt−1 , ECOˆ mod )
t
The bank’s monetary reserves, retained net earnings, or (1)
equity capital, XEqCap , is considered to be the funds it has
However, for the purposes of this paper, YT 1CRt is directly
available after expenses & losses are deducted, however these
projected using loan losses, net revenue, previous loan portfo-
funds may be further reduced due to capital distributions
lio, & estimated economic conditions.
1283
