Content uploaded by Murat Donduran
Author content
All content in this area was uploaded by Murat Donduran on Dec 09, 2021
Content may be subject to copyright.
ICE-TEA2021 International Conference on Economics
April 09-11, 2021 Turkish Economic Association
Constructing Early Warning Indicators for the Banks Using
Machine Learning Models
1
Coskun Tarkocin
2
Yildiz Technical University, Turkey
ctarkocin@hotmail.com
Murat Donduran
Yildiz Technical University, Turkey
donduran@yildiz.edu.tr
Abstract:
This work contributes the bank liquidity management by applying supervised machine learning
models to provide banks with early warnings of liquidity stress using market base indicators.
Identifying increasing levels of stress as early as possible provides management with a crucial
window of time in which to assess and develop a potential response. In this study, market stress
was transformed into a classification problem. Publicly available data from 2007 to 2021 was
used to train the machine learning model; this period covers two severe stress periods, namely
the 2007-2008 Global Financial Crisis and the Covid-19 crisis. The St. Louis Fed Financial
Stress Index was used to define the level of stress in the market. Each day trading was assigned
a red-amber-green (RAG) status to identify the risk level for that day. Machine learning models
were then applied to predict the RAG status of each day. Due to a significantly limited number
of “red” status days, modelling became more challenging. An ensemble model with a random
under-sampling boosting algorithm (RUSBoost) was used to improve predictions of
imbalanced data. Initial results show the machine learning model used in this study can predict
80% of “red” risk days. The current version of the developed model has been back-tested using
data from the Covid-19 crisis and it was able to trigger “amber” and “red” status as crisis
unfolding. These findings show that the ensemble model with the RUSBoost algorithm predicts
“red” and “amber” days on average 36% more than other machine learning models and can
contribute to bank risk management.
Keywords: Early Warning Indicators, Financial Stress, Machine Learning, Ensemble model,
Liquidity Risk, Crisis Management, Covid-19 Crisis
JEL Codes: C51, C88, G21
1
Disclaimer: The views and opinions expressed in this paper are those of the authors and they do not necessarily
reflect the views of the HSBC Group or Yildiz Technical University.
2
https://orcid.org/0000-0002-3910-0602
2
1. Introduction
Bank risk management has become more complicated with the evolving regulatory framework
following the 2007-2008 global financial crisis. With increased regulation and advancements
in risk management practices, it is expected banks and the wider financial system will be more
resilient to shocks. However, any new crisis may unfold uniquely which will require
management to be on alert if the market stress level changes unexpectedly.
Banks have adapted to the current relatively complex environment with large costly
investments. Extensive transformation projects have been implemented for systems, reporting,
modelling and governance. The granularity of the risk data produced and reported has thus
increased significantly.
Producing and reporting a great level of information does not necessarily give all the answers,
however. A few questions remain open on an ongoing basis, e.g. what are the stress levels in
the market, what is the perception of the clients and counterparties for a specific bank’s risk
level, and in the very short term, is there any emerging stress? These questions all need an
answer to define an alert level for any development in the market which may jeopardise banks
survival in a stress environment. For this reason, banks regularly monitor the market and
internal bank indicators to have a view of developing events.
Indicators, providing early detection of developing stress, will be referred to as Early Warning
Indicators (EWIs). It is a standard practice to monitor several indicators with defined thresholds
to inform management. However, this study will bring a novel way of looking at these
indicators by using supervised machine learning models and transforming them into a
classification problem without creating another stress index.
There are three main motivations behind this study. Firstly, early warning indicators research
in the literature is often focussed on the policymaker perspective, but not the perspective of a
bank as an individual agent in the financial system. Secondly, research using machine learning
models to detect financial stress levels is in its early stages and still developing, therefore the
addition of the Ensemble classifier with a random under sampling algorithm, as applied by this
study, will provide a unique contribution to the field. Lastly, this study focusses on the
immediate nature of the warning based on daily data available publicly, which links real time
market data and internal EWIs within the institution. This last item is specifically important in
the variable selection, since different frequency and granularity of data is available for the
specific institution compared to the policymakers. In addition, the flexibility of the proposed
supervised machine learning model makes it possible to be used in different markets and by
different stakeholders such as investors, regulators, or central banks.
This study is organised as follows: Section 2 contains a review of existing literature; Section 3
provides definitions of and a framework for early warning indicators; Section 4 outlines the
data selected for this study and summarises the data transformation process; Section 5 outlines
the methodology of this study, while Section 6 discusses the results of this study. Section 7
summarises the conclusion of the empirical analysis and discusses its policy implications.
2. Literature Review
The literature around Early Warning Indicators is extensive, since understanding emerging
stress as early as possible — though not so early that it may increase the cost of policies
implemented to prevent or reduce the impact of stress — is vital. The main motivation behind
EWIs has been to detect stress signals early on and then provide governments, central banks
3
and regulators sufficient time to prevent or reduce the impact of the emerging crisis (BCBS,
2013).
Literature around EWIs in banking can be grouped under two categories. The first category
aims provide a warning for financial or banking crises. The second category aims to provide
early warning of individual bank failure, which may negatively impact the wider financial
system (Gaytán & Johnson, 2002). Models predicting distress in banks from a policymaker
perspective help to better understand underlying vulnerabilities and identify patterns preceding
financial stress (Betz, Oprica, Peltonen, & Sarlin, 2013). This study will bridge the two
categories of literature by focussing on a specific financial crisis from an individual bank
perspective, detecting market signals which then can be incorporated to the bank’s internal
indicators for decision making, rather than policymaking. Therefore selected literature around
this area will be discussed.
Demirguc-Kunt and Detragiache (1998) studied determinants of banking crisis using
multivariate logit models for a large panel of countries, and found that banking crises are more
likely in a weak macroeconomic environment. Navajas and Thegeya (2013) used a simple
multivariate logit model on financial soundness indicators and macroeconomic control
variables to investigate a correlation between the occurrence of banking crises and other
variables. While capital adequacy ratio and return on equity has a negative correlation, their
analysis found that lagged return on equity may be a leading indicator for a banking crisis
(Navajas & Aaron, 2013).
History has shown there has been an increased interest to understand the drivers of and
consequently propose policy recommendations following each major crisis period. Yoshitomi
& Shirai (2000) conducted a critical and comprehensive review of research following the Asian
crisis, and made policy recommendations to prevent another one. Bell and Pain (2000) reviewed
empirical studies performed since the financial crisis in Asia and assessed the usefulness of
leading indicator models. They concluded these models had significant weaknesses and
limitations for policymaking. Gaytán & Johnson (2002) reviewed the literature on an early
warning system for banking crises and classified literature by their methodological approaches.
Regulatory guidance before the global financial crisis does include Liquidity EWIs. Basel
Committee framework guidance for the managing liquidity BIS (1992) and BIS (2000) versions
do not refer to any EWIs to be monitored by individual banks. However, the BIS (2008)
“Principles for Sound Liquidity Risk Management and Supervision” version included a
recommendation to individual banks to have EWIs. BIS (2008) highlights the importance of
designing a set of indicators to identify increased risk or vulnerabilities.
After the 2007-2008 Global Financial Crisis (GFC), literature around early warning indicators
peaked and further developed. There was a similar increase in research interest in this area
following the financial crisis in Asia (Bräuning et al, 2019). For instance, Surjaningsih,
Yumanita, Deriantino (2014) developed EWIs for banking liquidity risk and could detect
liquidity imbalances one year before October 2008, with 67% predictive power reported.
Iachini and Nobili (2014) introduced a coincident indicator of systemic liquidity risk in the
Italian financial markets using standard portfolio theory. Individual raw indicators from the
equity and corporate market, Italian government bond market and money market were used.
The calculated indicator was then compared to the conducted survey of the most liquidity
stressful events for the Italian financial markets, which showed the systemic liquidity risk
indicator accurately identifies high systemic risk events (Iachini & Nobili, 2014).
Eross, Urquhart and Wolfe (2015) employed an autoregressive Markov regime-switching
model and showed US Libor-OIS spread can be used to predict when a liquidity crisis may
occur within the interbank market. Acharya et al (2017) developed a model-based measure of
systemic risk and with a variety of tests, showed market data from equity and CDS were able
4
to predict financial firms worst contributors to systemic risks. Aldasoro, Borio and Drehmann
(2018) assessed household and international debt as EWIs for banking distress, and found these
indicators provide useful information when modelling rare events such as crises.
Lang, Peltonen and Sarlin (2018) proposed a general-purpose framework to build early warning
models and applied it to predict bank distress in Europe. This study also provided a flexible
modelling solution, combining the loss function approach to evaluate models with regularised
logistic regression and cross validation, which can be used to analyse emerging vulnerabilities
at micro and the macro level.
Padhan and Prabheesh (2019) documented the history of early warning models predicting the
financial crisis and also proposed a new agenda to augment existing models. Models were listed
under six different types and using machine learning, family artificial neural networks and
genetic algorithms were reported. In a more recent study employing machine learning model to
predict bank distress, Bräuning et al (2019) employed a decision tree model with the Quinlan
C5.0 algorithm to identify individual bank distress for small European banks. The study
reported low type I and type II errors, with the decision tree model outperforming the logit
model (Bräuning, Malikkidou, Scalone, & Scricco, 2019).
3. Early Warning Indicators (EWIs) Definitions and Framework
In this study, EWIs are defined as any data or information used to predict a potential stress
event. This can be numerical or categorical data. Market-based indicators reflect the health of
the economy, bank and firm, and can predict changes in financial conditions (Kliesen &
McCracken, 2020).
Examples of EWIs include, but are not limited to, the following:
Market EWIs: These are mainly external data sources such as equity prices, interest
rates, spreads, commodity prices, macroeconomic variables, credit ratings, futures,
foreign exchange rates, policy rates, stock indices, volatility indices, and secured
funding spreads.
Internal EWIs: These are bank-specific measures including capital position, deposit
outflows, maturity mismatch, stress test results, bank own credit rating and CDS
spread, credit losses measures, liquidity and funding metrics, market sentiment, share
price, funding spread compare to peers, concentration metrics, negative publicity, and
increasing currency mismatches.
A comprehensive list of Market and Internal EWIs can be seen in Venkat and Baird(2016)
chapter 6 (authored by Bruce Choy and Girish Adake) and the BIS (2008) suggested EWIs for
liquidity risk.
The proposed specific modelling and data experiment focus is Market EWIs, however
individual banks can employ the proposed ensemble classifier random undersampling
algorithm to generate red-amber-green status (RAG status) based on only internally available
daily data. This provides a comprehensive view of different stress types, detecting
vulnerabilities which may not be visible based on public information.
RAG Status is a common framework for risk measurement and monitoring. A Green indicator
means that the measure suggests normal market environment, while an Amber indicator advises
elevated stress levels which may need further investigation of underlying risk, and a Red
indicator highlights major stress levels in the market, demonstrating the need for immediate
management attention and a potential response.
5
Backtesting of the thresholds for Green, Amber and Red statuses is of critical importance to
make sure management receives warning light indicators when most relevant. Backtesting will
help to determine if recalibration is needed, for example where a measure exceeds the threshold
frequently it will report false alarms Venkat and Baird(2016) chapter 6 (authored by Bruce
Choy and Girish Adake) A recent example showing the importance of backtesting is the St
Louis Fed’s Financial Stress Index (STLFSI2), which was revised from STLFSI1 to STLFSI2,
since the original version of the index was not able to detect market movements around August
2011 and, most importantly, turmoil triggered by the Covid-19 pandemic around February-
March 2020 (Kliesen & McCracken, 2020). Models should aim to have a limited number of
false alarms, and following a major stress period, recalibration and review of the EWIs should
be performed.
4. Data
Publicly available data from 2007 to 2021 (3537 days) was used to train the machine learning
models in this study. Historical data for potential EWIs for the same period is sourced from
FRED Economic Data and Yahoo Finance for different markets such as Equity, FX, Interest
rate and Spreads.
Table 1: List of Raw Data and Transformation
Group
Raw Data
Transformed for Model
Definition
Volatility
VIX
Logarithmic return, 7days
standard deviation, high-low
range to low
CBOE Volatility Index (^VIX)
Commodity
Crude_F
Logarithmic return, 7days
standard deviation, high-low
range to low
Crude Oil Mar 21 (CL=F)
Futures
Nasdaq100_F
Logarithmic return, high-low
range to low
Nasdaq 100 Mar 21 (NQ=F)
Equity
Euronext100
Logarithmic return, 7days
standard deviation, high-low
range to low
EURONEXT 100 (^N100)
Spread
TED
Movement between two dates
TED Spread
Spread
CPFF
Actual spread as %, 7 days
standard deviation
3-Month Commercial Paper
Minus Federal Funds Rate
4.1. Data Transformation and Cleaning
The St. Louis Fed Financial Stress Index, which begins in late 1993, was used to define the
level of stress in the market and train the machine learning model. In this index, zero represents
normal market conditions. When the value drops below zero it means below-average financial
market stress, whilst above zero is interpreted as above-average market stress (Federal Reserve
Bank of St. Louis, 2021).
In this study, referring to this general definition and based on historical major stress events like
the global financial crisis and recent Covid-19 stress, we assumed 85% normal days (Green),
6
10% moderate stress days (Amber) and 5% severe stress days (Red). One limitation of the St
Louis Fed Financial Stress Index is it provides weekly rather than daily data, therefore the same
RAG status is assumed for the whole week. Although it is recognised some days may be
different level, this should have a limited impact on the main aim of this study.
With a few exceptions, historical data exists without issue. Below are two issues addressed in
data preparation, the impact of which is deemed to be limited:
- On 20 and 21 April 2021, crude oil futures prices fell below zero dollars which
resulted in a big outlier for the range for these two days than the previous day range
used.
- CPFF was not available for a number of days during the 2008 financial crisis and
Covid-19 crisis. For these days (78 days), the elevated spread level from the last
available date was used.
Raw data in Table 3 was transformed as below to use in the modelling. A seven day rolling
standard deviation was calculated to capture building volatility in the market represented as a
feature.
Table 2: Measures used in tranforming raw data for modelling
Measure
Formula
High-low range
TED Spread change
Logarithmic return
Seven days rolling standard deviation
High-low range was calculated to capture intra-day variation for the variables referred in Table
3. TED Spread, defined as the difference between the three-month US Treasury bill and the
three-month USD Libor, was used as a movement in the model. In order to capture daily
movement of the variables in the Table 3, logarithmic return was calculated as below.
Visual inspection shows the STLFSI2 captures two recession periods and major financial crises.
In the below figure (Figure 1) it can be seen that STLFSI2 significantly increases during the
2008-2009 global financial crisis and Covid-19 crisis around February-March 2020. This
supports the hypothesis that the model can be used as a benchmark to train machine learning
model for RAG status estimation. One limitation is that STLFSI2 is published weekly, therefore
day t features were used to predict the t+1 index level.
7
Figure 1: St. Louis Fed Financial Stress Index (STLFSI2)
Source: Federal Reserve Bank of St. Louis
5. Methodology
A subset of artificial intelligence, machine learning models are employed to conduct data
experiments in this study. The popularity of machine learning models in the banking and
financial sector has increased in recent years mainly due to the increasing availability of data,
computing power and improved software (BOE, 2019).
Machine learning models can be grouped into three main categories: supervised, unsupervised,
and reinforcement learning. Supervised learning models train data based on a given input and
output. By contrast, unsupervised learning models analyse data without a given output, and find
potential relationships through clustering data. In reinforcement learning models, the aim is to
maximise the defined reward for the specific action given (McKinsey&Company, 2021). In
addition, FSB (2017) identifies deep learning models under the machine learning umbrella and
defines these as algorithms that work with layers similar to how the human brain functions.
To the best of our knowledge, this study will be the first in the literature to use Ensemble
Classifiers with a random under-sampling algorithm to construct Early Warning Indicators.
EWIs produced by the model may trigger a management response which may have cost
implications. Due to this reason, explainability of the model and, where necessary. overlaying
the model with expert judgment is of utmost importance. In the following section selected
classifier and algorithm are briefly discussed. Machine learning models tend to have overfitting
problem if cross validation not performed, in order to prevent overfitting in this study K-fold
cross validation used with K=5.
5.1. Ensemble Classifiers
As discussed in the previous section, only Ensemble Classifiers will be used in this study since
the underlying data is materially imbalanced. Ensemble learning aims to combine predictions
from several base estimators to build a more robust single estimator. Two families of ensemble
methods are widely used. Firstly, there are Averaging methods, such as Bagging methods and
Forests of randomised trees. These group of models that independently make predictions and
then average those predictions. On the other hand, the second group – Boosting methods –
builds base estimators sequentially to achieve a combined estimator with reduced bias.
Examples of Boosting methods RUSBoost, AdaBoost, Gradient Tree Boosting. (scikit-learn,
2020).
8
The RUSBoost algorithm was first introduced by Seiffert et al (2008) to reduce class imbalance
problems in the data set. RUSBoost uses random data sampling with boosting, which as a result
improves the classification performance of the training data. Financial stress or distress bank
classification problems have imbalanced data, wherein one class has far fewer members than
others. For this reason, the RUSBoost algorithm is used for the modelling in this study. For a
comprehensive overview of the RUSBoost algorithm, please refer to Seiffert et al (2010).
Matlab 2019a version is used for the implementation of the models which has eleven ensemble
learning algorithms. Out of these eleven methods, Random Undersampling Boosting
(RUSBoost method) is a better fit for imbalanced data and can be used for binary and multiclass
classification. Matlab’s random under-sampling algorithm takes the same number of
observations from each class of data, which is the same as the number of the minority class.
Following sampling, adaptive boosting is applied to construct the ensemble. For full details see
Matlab documentation under Ensemble Algorithms (MATLAB, Ensemble Algorithms, 2019).
The boosting procedure for RUSBoost applies adaptive boosting for multiclass classification in
calibrating weights and constructing ensembles. Adaptive boosting for multiclass classification
in MATLAB uses weighted pseudo-loss for N observation and K classes. Calculated pseudo-
loss is used as a measure of classification accuracy (MATLAB, Ensemble Algorithms, 2019).
- Each step represented by t, k represents class, N number of observations, K classes
- is a vector of predictor values for observation n,
- represents the true class value taking one of the
- represents the prediction of the learner for each step t
- is the confidence of learner prediction at step t, class k range from zeo to one,
-
represents observation weights of class k in step t
5.2. Performance Evaluation Metrics and Definitions
Confusion Matrix
The Confusion Matrix is used to evaluate the performance of the models in the data
experiments. Using data presented on the Confusion Matrix, several performance measures are
calculated.
The below table (Table 4) summarises information presented on the Confusion Matrix. As the
table shows, G+A+R = N, the total number of observations.
9
Table 3: Confusion Matrix
RAG Status Predicted Class
RAG
Status
Actual
Class
Green
Amber
Red
Tota
l
Green
True Green (TG)
False Amber_1
(FA1)
False Red_1
(FR1)
G
Ambe
r
False Green_1
(FG1)
True Amber (TA)
False Red_2
(FR2)
A
Red
False Green_2
(FG2)
False Amber_2
(FA2)
True Red (TR)
R
Total
G*
A*
R*
N
Formulas for the measures above are outlined below:
Accuracy (Acc) =
. This metric measures the percentage of true class
predictions in all observations.
Error (Err) = . This metric provides the misclassification percentage.
In this study, we will extend two measures that can be used specifically for predicting the RAG
status of EWIs:
Warning Score (WS) =
. The rationale behind this score is that an
institution would at least like to be warned of elevated stress levels. Once a Warning
Score is received in combination with another internal measure, final management
status can be decided. This measure will ignore misclassification between Red and
Amber days,
True Red Accuracy (TRA) =
. This is to control accuracy of red status days, since
red days are most impacted by the imbalanced data, for instance if overall AUC is
very high but TRA below 50%, model may not be fit for EWIs usage.
ROC Curves
Another method used in analysing model results performance is Receiver Operating
Characteristics (ROC curves). ROC curves represent true positive rates versus false positive
rates to show trade-off between correctly identified and false prediction. The area under the
curve (AUC) is used to compare model performance for the classification models (Seiffert et
al, 2010).
6. Empirical Results
In the MATLAB program (2019a version), under six model types a total of 24 classification
models run to predict the RAG status of each day. Six performance evaluation metrics,
presented in the table below (Table 6), compare model performances for the EWIs’ RAG status
prediction. For each model type, the results for the model with highest accuracy is presented in
bold for comparison purposes.
When using Accuracy (or inversely Error) as a measure, ‘Ensemble – Boosted Trees’ shows
the highest predictive power, however when focus is moved to the prediction of only Red days
or both Red/Amber days, ‘Ensemble – RUSBoost’ clearly shows better performance. The WS
for the latter returned 95% accuracy compared to just 74% for the average of the other best
performing models. This trend is supported by the TRA measure as well.
10
It must be noted that k-fold cross-validation was performed to prevent overfitting. Without cross
validation, in-sample accuracy would be very high but at the cost of sacrificing the ability to
make predictions beyond the sample.
Table 4: Comparison of Models
MATLAB Model Name
Acc
Err
WS
TRA
AUC
Ensemble – RUSBoost
88%
12%
95%
83%
0.98
Average of below models
91%
9%
74%
70%
0.97
Tree – Medium Tree
92%
8%
73%
77%
0.97
Quadratic Discriminant
89%
11%
74%
60%
0.96
Naïve Bayes
90%
10%
85%
73%
0.94
Support Vector Machine –
Quadratic
93%
7%
79%
71%
0.96
k-Nearest Neighbor (weighted
KNN)
92%
8%
58%
62%
0.98
Ensemble – Boosted Trees
94%
6%
78%
78%
0.99
A model which perfectly predicts normal days performs best based on Acc and AUC measures.
However, with imbalanced data predicting normal days is not the aim. For distress prediction,
WS and TRA are the proposed measures for determining model selection. ‘Ensemble –
RUSBoost’ shows the best performance for this task. Therefore, the results for this model will
be examined more closely.
The Confusion Matrix (Figure 1) for the best performing model (Ensemble – RUSBoost) shows
a true positive rate of 83% for Red status; 84% for Amber status, and 89% for the normal days
with Green status. Improved Red and Amber day prediction comes at the cost of missing some
of the normal days; since it is preferable to miss Green days rather than higher-risk Red and
Amber days, this is acceptable. Significantly, the ‘Ensemble – RUSBoost’ model did not
misclassify any Red status days as Green. This means a warning will only be produced when
stress levels are elevated. Achieving 95% WS and 83% TRA supports the hypothesis that the
trained model can be automated to predict daily RAG statuses.
11
Figure 2: Confusion Matrix- Ensemble Model RUSBoost
The Receiver Operating Characteristic (ROC) curve corroborates the observation that
‘Ensemble – RUSBoost’ is the best performing model. A high number in the Area Under Curve
(AUC) suggests a classifier is performing well. Perfect results, where no misclassification
occurs, appear in the top left corner of the plot; random results appear along a diagonal line at
45 degrees across the plot, and poor results appear below this line, towards the bottom right
corner. Table 7 shows the ROC curve for ‘Ensemble Model – RUSBoost’, where Red days are
true positives. It also shows that if a classifier’s TPR is increased it will compromise false
positive rates.
Figure 3: ROC Curve for ‘Ensemble – RUSBoost’
12
Based on the above empirical results, it can be concluded that the predictive performance of the
Ensemble Model with RUSBoost algorithm is satisfactory to be employed for EWI RAG status
prediction. Future work may involve employing more features and different model
combinations, as well as performing hyperparameter tuning to further optimise model results
before deployment.
7. Conclusions
This paper employed supervised machine learning models to predict EWI RAG statuses for the
market across 3537 days between January 2007 and January 2021. Data used for training the
model covered two major stress periods: the 2007-2008 global financial crisis, and the Covid-
19 market crisis in 2020. The St. Louis Fed Financial Stress Index (STLFSU2) was used to
create three class classification problems and to train the machine learning models.
This study showed the Ensemble method with random undersampling algorithm (Ensemble –
RUSBoost) outperforms 23 other models in the MATLAB program based on Warning Score
and true red accuracy measures. This method and measuring approach for EWIs distinguishes
this study from existing literature.
The model and framework proposed in this study can be applied in a bank setting. This will
enable financial institutions to combine their internal metrics with market stress measures,
thereby producing stronger risk warning signals specific to their business model or balance
sheet structure. Utilising machine learning model for this purpose is dynamic in nature and can
be automated for integration into management information systems. With further calibration,
the link to the real-time data may be possible.
Although in this study addresses the problem of risk measurement from an individual bank
perspective, future work can be done by regulators using their available data superiority to
incorporate dynamic EWI RAG status prediction.
1.8 References
Acharya, V. V., Pedersen, L. H., Philippon, T., & Richardson, M. (2017). Measuring Systemic
Risk. The Review fo Financial Studies, Volume 30, Issue 1, 2-47.
Aldasoro, I., Borio, C., & Drehmann, M. (2018). Early warning indicators fo banking crises;
expanding the family. BIS Quarterly Review, 29-45.
BCBS. (1992). A Framework for Measuring And Managing Liquidity.
BCBS. (2000). Sound Practices for Managing Liquidity in Banking Organisations.
BCBS. (2008). Principles for Sound Liquidity Risk Management and Supervision.
BCBS. (2013). Evaluating early warning indicators of banking crisis: Satisfying policy
requirements. BIS Working Papers No 421.
13
Betz, F., Oprica, S., Peltonen, T. A., & Sarlin, P. (2013, October 11). Predicting Distress in
European Banks. Retrieved from ECB Working Paper no. 1597:
https://ssrn.com/abstract=2338998
BOE. (2019, October). Machine Learning in UK Financial Services. Retrieved from
https://www.bankofengland.co.uk/-/media/boe/files/report/2019/machine-learning-in-
uk-financial-services.pdf
Bräuning, M., Malikkidou, D., Scalone, S., & Scricco, G. (2019). A new approach to Early
Warning Systems for small European banks. ECB Working Papers Series No 2348.
Demirguc-Kunt, A., & Detragiache, E. (1998). The Determinants of Banking Crises in
Developing and Developed Countries. IMF Staff papers, Volume 45, Number 1.
Eross, A., Urquhart, A., & Wolfe, S. (2015, September 29). An Early Warning Indicator for
Liquidity Shortages in the Interbank Market. Retrieved from Eross, Andrea and
Urquhart, Andrew and Wolfe, Simon, An Early Warning Indicator for Liquidity
Shortages in the Interbank https://ssrn.com/abstract=2658797
Federal Reserve Bank of St. Louis. (2021, February). Retrieved from
https://fred.stlouisfed.org/series/STLFSI2
FSB. (2017). Artificial intelligence and machine learning in financial services. Financial
Stability Board.
Gaytán, A., & Johnson, C. A. (2002, October). A Review of the Literature on Early Warning
Systems for Banking Crises. Working Papers Central Bank of Chile 183. Retrieved
from https://ideas.repec.org/p/chb/bcchwp/183.html
Iachini, E., & Nobili, S. (2014, April 23). An Indicator of Systemic Liquidity Risk in the Italian
Financial Markets. Retrieved from Bank of Italy Occasional Paper No. 217:
https://ssrn.com/abstract=2489885
Kliesen, K., & McCracken, M. (2020, March 26). The St. Louis Fed’s Financial Stress Index,
Version 2.0. Retrieved from https://fredblog.stlouisfed.org/2020/03/the-st-louis-feds-
financial-stress-index-version-2-0/
Lang, J. H., A., P. T., & Sarlin, P. (2018). A framework for early-warning modeling with an
application to banks. ECB Working Paper Series.
MATLAB. (2019). Ensemble Algorithms. Retrieved from
https://uk.mathworks.com/help/stats/ensemble-algorithms.html
14
MATLAB. (2021, 03 09). Machine Learning in Matlab. Retrieved from
https://uk.mathworks.com/help/stats/machine-learning-in-matlab.html
McKinsey&Company. (2021, 03 09). An executive's guide to AI. Retrieved from
https://www.mckinsey.com/business-functions/mckinsey-analytics/our-insights/an-
executives-guide-to-ai
Navajas, M. C., & Aaron, T. (2013). Financial Soundness Indicators and Banking Crisis. IMF
Working paper.
Nosratabadi, S., Mosavi, A., Duan, P., Ghamisi, P., Filip, F., Band, S. S., . . . Gandomi, A. H.
(2020). Data Science in Economics: Comprehensive Review of Advanced Machine
Learning and Deep Learning Methods. Mathematics 2020, 8(10), 1799.
scikit-learn. (2020). 1.11. Ensemble methods. Retrieved from Scikit Learn Documentation:
https://scikit-learn.org/stable/modules/ensemble.html
Seiffert, C., Khoshgoftaar, T. M., Hulse, J. V., & Napolitano, A. (2008). RUSBoost: A Hybrid
Approach to Alleviating Class Imbalance. 19th International Conference on Pattern
Recognition, (pp. 1-4).
Seiffert, C., Khoshgoftaar, T. M., Van, H. J., & Napolitano, A. (2010). RUSBoost: A Hybrid
Approach to Alleviating Class Imbalance. IEEE Transactions on Systems, Man, and
Cybernetics- Part A: Systems and Humans Vol.40 No.1, 185-197.
Venkat, S., & Baird, S. (2016). Liquidity Risk Management A Practitioner's Perspective. Wiley
Finance Series.
