See discussions, stats, and author profiles for this publication at: [Link]

net/publication/381093192

CLOUD-NATIVE AI FOR RISK ASSESSMENT IN BANKING INSURANCE

Article · March 2024

CITATIONS READS

10 18

1 author:

Sridhar Madasamy
Indian Institute of Technology Kharagpur
12 PUBLICATIONS 123 CITATIONS

SEE PROFILE

All content following this page was uploaded by Sridhar Madasamy on 02 June 2024.

The user has requested enhancement of the downloaded file.

 © 2024 IJRTI | Volume 9, Issue 3 | ISSN: 2456-3315

CLOUD-NATIVE AI FOR RISK ASSESSMENT IN

BANKING INSURANCE
SRIDHAR MADASAMY M

Principal Solutions Architect

Aquilanz llc, USA.

Abstract- Risk assessment is pivotal in banking and insurance, guiding applicant classification. Underwriting
processes determine application decisions and policy pricing. As data volumes surge and analytics progress,
automation accelerates underwriting for swift application processing. This study endeavors to augment risk
assessment in banking and insurance through predictive analytics. Employing a real-world dataset with
anonymized attributes exceeding one hundred, the research conducts comprehensive analysis to propose
solutions for enhanced risk assessment in bank insurance operations. Initiating with dataset collection, the
framework ensures the acquisition of comprehensive data sources pertinent to risk evaluation. Subsequently,
data pre-processing procedures are implemented to cleanse the dataset of noise, outliers, and missing values,
essential for ensuring data integrity and consistency. Dimensionality reduction techniques are then applied to
streamline the dataset and enhance computational efficiency. Specifically, to find and extract important
characteristics helpful for risk assessment, Linear Discriminant Analysis (LDA) based feature extraction
techniques and Improved Artificial Bee Colony (IABC) based feature selection techniques are used. Leveraging
cloud infrastructure facilitates the scalability and flexibility required for processing large-scale datasets
efficiently. The multi-model selection phase encompasses the application of various algorithms, including
Recurrent Neural Networks (RNN), Convolutional Artificial Neural Networks (CANN), and Multiple Linear
Regression (MLR), for predictive modeling. Each algorithm brings unique strengths to the table, enabling
comprehensive risk assessment from different angles. The framework aims to optimize risk assessment
processes in banking and insurance domains, ultimately enhancing decision-making accuracy and operational
efficiency.

Keywords: Banking Insurance; Risk Assessment; IABC; LDA; RNN; CANN.

1. INTRODUCTION
In the fast-paced and dynamic landscape of banking and insurance, the ability to effectively access and manage risk is
important. With the advent of cloud-native artificial intelligence (AI) technologies, the traditional methods of risk
assessment are undergoing a profound transformation [1, 2]. This paradigm shift promises to revolutionize decision-
making processes, enhance predictive capabilities, and ultimately drive greater efficiency and profitability across the
industry [3, 4]. The fusion of cloud computing and AI has revolutionized risk assessment in banking and insurance.
Traditional methods relied on static models and historical data [5, 6]. However, leveraging cloud elasticity allows for
rapid deployment and scaling of AI-powered risk models. This agility optimizes resource use, reduces time-to-market,
and enhances competitiveness. Cloud-native AI excels in processing real-time data from diverse sources, a crucial
advantage in modern risk assessment [7, 8].
Traditional risk models often struggled to incorporate unstructured data such as social media feeds, news articles, and
sensor data, which can provide valuable insights into emerging risks and market trends [9]. However, AI platforms
excel at handling big data, allowing organizations to extract actionable intelligence from a wide range of sources and
make informed decisions with confidence [10]. Furthermore, AI solutions offer advanced machine learning
capabilities that can uncover hidden patterns and correlations within complex datasets. By analyzing historical data
and identifying underlying trends, these models can predict future market movements, identify potential risks, and
optimize portfolio performance [11, 12]. This predictive power enables banks and insurance companies to proactively
manage risk, optimize capital allocation, and maximize returns on investment. The potential of AI to automate and
streamline decision-making processes is another important advantage for risk assessment.
Traditional risk assessment models often required manual intervention and human oversight, which could introduce
errors and delays into the decision-making process [13]. Automating tasks like data collection and analysis allows
employees to concentrate on strategic projects, boosting efficiency and accuracy in risk assessment. In addition to
improving risk assessment capabilities, AI solutions can also enhance regulatory compliance and governance within
the banking and insurance sectors. With increased scrutiny from regulators and growing demand for transparency,
organizations must demonstrate robust risk management practices and compliance with industry regulations.

IJRTI2403070 International Journal for Research Trends and Innovation ([Link]) 485
 © 2024 IJRTI | Volume 9, Issue 3 | ISSN: 2456-3315

The banking and insurance sectors face significant challenges in accurately assessing and managing risks associated
with financial transactions and policies [1]. Traditional risk assessment methods often struggle to cope with the
increasing volume, velocity, and variety of data, leading to potential inaccuracies and inefficiencies [12]. Moreover,
evolving market dynamics and regulatory requirements further complicate risk management practices, necessitating a
more advanced approach. Therefore, there is a pressing need to develop innovative risk assessment frameworks that
harness the power of cloud-native AI technologies. By leveraging cloud computing resources and advanced AI
algorithms, these frameworks can integrate diverse data sources, model complex risk relationships, enhance
transparency and interpretability, and enable real-time risk monitoring and adaptive management strategies.
Ultimately, such frameworks aim to empower banking and insurance institutions with actionable insights, proactive
risk mitigation capabilities, and improved decision-making processes, thereby strengthening the resilience and
sustainability of the financial ecosystem.
The contributions of this paper are manifested below,
• Rigorous procedures are implemented to cleanse the dataset, removing noise, outliers, and missing values,
thereby ensuring data integrity and consistency crucial for reliable risk assessment.
• Using IABC for feature selection and LDA for feature extraction enhances computational efficiency by
refining the dataset while preserving crucial risk-related features.
• Various algorithms, including Recurrent Neural Networks (RNN), Convolutional Artificial Neural Networks
(CANN), and Multiple Linear Regression, are applied for predictive modeling, each offering unique strengths to
comprehensively assess risk from different perspectives.
The subsequent sections of this paper are structured as follows: Section II covers related works and the problem
statement. Section III outlines the proposed methodology. Section IV presents the results and discussions, while
Section V concludes the paper.

2. LITERATURE REVIEW
In 2023, Settipalli and Gangadharan [14] presented a significant challenge, driving up medical costs by Health
insurance fraud. Current models struggle with accuracy due to complex, overlapping claims data. Our novel approach
employs unsupervised multivariate analysis, utilizing Weighted MultiTree (WMT) for categorical data and Density
Based Clustering (DBC) for continuous data. Empirical tests on CMS part B claims show improved detection
performance over existing models.
In 2022, Fursov et al. [15] proposed deep learning architectures process sequential patient visit data, enhancing fraud
detection. Empirical findings from health insurance data show superior performance with a ROC AUC of 0.873,
outperforming the best competitor's 0.815. This approach is robust to data corruption and applicable to similar
scenarios with semi-structured event sequences.
In 2021, Găbudeanu et al. [16] developed over the past decade, both industry initiatives and European Union
legislation, notably PSD2, have prioritized the enhancement of fraud detection capabilities. However, while
legislation outlines broad obligations, the market's varied solutions raise concerns regarding the aggregation of data
for fraud identification. This article delves into this discrepancy by analyzing existing fraud detection methodologies
in light of data protection laws and gauging respondents' attitudes toward privacy in fraud identification. It aims to
bridge the gap between fraud detection practices and compliance with data protection regulations.
In 2022, Hashemi et al. [17] used class weight-tuning hyperparameters to balance fraudulent and legitimate
transactions, employing Bayesian optimization to address data imbalance. CatBoost and XGBoost complement
LightGBM, enhancing performance through ensemble voting. Deep learning fine-tunes hyperparameters, notably
weight-tuning, resulting in improved metrics such as ROC-AUC, precision, recall, F1 score, and MCC, surpassing
existing methods.
Kolli and Tatavarthi [18] introduced the Harris water optimization-based deep recurrent neural network (HWO-based
deep RNN) in 2020 for fraud detection. The method comprises three phases: pre-processing, feature selection, and
fraud detection. Utilizing Box-Cox transformation in pre-processing and a wrapper model for feature selection, the
proposed model achieved superior performance metrics such as accuracy, sensitivity, and specificity, demonstrating
its efficacy in detecting fraudulent activities in bank transactions.
In 2021, Zhou et al. [19] used an intelligent and distributed Big Data approach for detecting Internet financial fraud,
leveraging the graph embedding algorithm Node2Vec. This method learns and represents topological features from
financial network graphs as low-dimensional vectors, facilitating intelligent classification and prediction using deep
neural networks.
In 2020, Subudhi and Panigrahi introduced a novel hybrid approach for detecting vehicle insurance fraud. Their
method combined supervised classifiers with Fuzzy C-Means clustering using Genetic Algorithm. The dataset was
split into train and test sets, with the train set undergoing clustering for undersampling. Various classifiers were then
applied to examine suspicious instances, demonstrating effectiveness on real-world data.

IJRTI2403070 International Journal for Research Trends and Innovation ([Link]) 486
 © 2024 IJRTI | Volume 9, Issue 3 | ISSN: 2456-3315

3. PROPOSED METHODOLOGY
Banking risk assessment integrates data collection, analysis, and risk evaluation techniques to identify and mitigate
potential risks. By leveraging both quantitative and qualitative methods, including statistical modeling and expert
judgment, the approach prioritizes risks based on severity and likelihood. Incorporating risk modeling and simulation
enables forecasting of potential outcomes, facilitating proactive risk management strategies. This comprehensive
process informs decision-making, enhancing the bank’s ability to safeguard financial stability and optimize
profitability.
3.1. Dataset Collection
The Prudential Life Insurance Assessment dataset [21] provides comprehensive information on life insurance
applicants, encompassing over a hundred variables. These variables cover various aspects such as product details,
applicant demographics (including age, height, weight, and BMI), employment history, insurance and family
background, medical history, and presence or absence of specific medical keywords. The final judgment related to
each application is represented by the dataset's target variable, "Response," an ordinal measure of risk with eight
levels. With a mix of categorical, continuous, and discrete variables, this dataset presents a rich source of data for
developing predictive models aimed at improving the efficiency and effectiveness of the life insurance application
process while maintaining privacy and accuracy standards. Out of the 59,381 instances and 128 attributes in the
dataset, those with more than 30% missing data, including Employment_Info_1, Employment_Info_4,
Employment_Info_6, and Medical_History_1, will be considered for analysis. Imputation techniques will be applied
to address missing values and maintain dataset integrity.
3.2. Data Pre-Processing
Data pre-processing involves cleaning the data by removing noisy data or outliers and addressing inconsistencies.
Specifically, missing values are treated to ensure consistency for analysis. Imputation methods are applied based on
the structure and mechanism of missing data, categorized as MCAR, MAR, or MNAR.

Figure 1: Overall Proposed Architecture

IJRTI2403070 International Journal for Research Trends and Innovation ([Link]) 487
 © 2024 IJRTI | Volume 9, Issue 3 | ISSN: 2456-3315

EDA explores feature distributions and relationships with the response variable. Visual analytics, including interactive
dashboards, aid in gaining insights into the structure and identifying suitable prediction models. Overall Proposed
Architecture is shown in the Figure 1.
In this study, data pre-processing involves missing data imputation, a crucial step aimed at handling incomplete or
missing values within the dataset. Missing data imputation techniques are employed to estimate and fill in the missing
values with plausible substitutes, ensuring that the dataset is complete and suitable for analysis. This process enables
the utilization of all available information, maintains data integrity, and prevents bias in subsequent analyses or model
building tasks. Various imputation methods may be employed based on the nature of the data and the characteristics
of the missing values, including regression imputation, mean or median imputation, or advanced techniques like
multiple imputations or k-nearest neighbors (KNN) imputation.
3.3. Dimensionality Reduction
Dimensionality reduction aims to streamline modeling by reducing the number of variables. It encompasses feature
selection, which involves choosing significant variables, and feature extraction, which transforms high-dimensional
data into fewer dimensions. This process accelerates model training and enhances accuracy by mitigating overfitting.
This study explores feature selection techniques like IABC and feature extraction methods such as principal
component analysis. Leveraging cloud resources ensures efficient dimensionality reduction, handling extensive data
volumes, and complex algorithms effectively.
3.3.1. IABC based Feature selection
ABC algorithm [26] is a metaheuristic optimization inspired by honey bees' foraging behavior. It consists of
employed, onlooker, and scout bees. Employed bees scout food sources and share data with onlooker bees, who
probabilistically select sources based on quality. Meanwhile, scout bees emerge from employed bees, exploring new
regions when abandoning their current sources, adding an exploration element to prevent local optima entrapment.
The ABC algorithm efficiently emulates social cooperation and intelligent foraging to iteratively refine solutions in
the search space for optimal or near-optimal outcomes, making it applicable to diverse optimization challenges.
In the ABC algorithm, the swarm consists of employed and onlooker bees, each comprising half of the total swarm
size, which equals the number of solutions. It starts with a randomly distributed population of 𝑠𝑠 solutions (food
sources), where 𝑠𝑠 is the swarm size. Each solution, 𝐴𝑖 = {𝑎𝑖,1 , 𝑎𝑖,2 , … , 𝑎𝑖,𝑑𝑠 } , with 𝑑𝑠 as the dimension size,
generates candidate solutions 𝐶𝑖 in its neighborhood to optimize as per Eq. (1).
𝑐𝑖,𝑗 = 𝑎𝑖,𝑗 + ∅𝑖,𝑗 ∙ (𝑎𝑖,𝑗 − 𝑎𝑘,𝑗 ) (1)
The algorithm randomly selects a candidate solution 𝐴𝑘 (where, 𝑖 ≠ 𝑘), and a dimension index 𝑗 from {1, 2, . . . , 𝑑𝑠},
then determines a random number ∅𝑖,𝑗 within [-1, 1]. Using these, a new candidate solution 𝐶𝑖 is generated. If
𝐶𝑖 fitness surpasses 𝐴𝑖 , 𝐴𝑖 is updated; otherwise, it remains unchanged. Employed bees share their food source info
through dances with onlooker bees. Onlookers evaluate this information and choose a food source probabilistically
based on nectar amount, akin to roulette wheel selection as Eq. (2) suggests.
𝑓𝑖𝑡
𝑟𝑖 = ∑𝑠𝑠 𝑖 ⊕ 𝐿𝑒𝑣𝑦(𝛽) (2)
𝑗=1 𝑓𝑖𝑡𝑗
In the ABC algorithm, incorporating Lévy flight enhances resource search efficiency by enabling a random walk
during the employed bee phase. This novel mechanism generates new solutions around the best food source,
improving search direction and algorithm performance. The selection probability of the 𝑖𝑡ℎ food source is determined
by its fitness value, denoted as 𝑓𝑖𝑡𝑖 . A higher fitness value indicates a higher probability of selecting the 𝑖𝑡ℎ food
source. If a position cannot be enhanced within a predefined number of cycles (referred to as the limit), the food
source is abandoned. In such cases, denoted by 𝐴𝑖 , the scout bee identifies a new food source to replace it, as per Eq.
(3).
𝐴𝑖,𝑗 = 𝑙𝑜𝑏𝑗 + 𝑟𝑎(0,1) ∙ (𝑢𝑝𝑏𝑗 − 𝑙𝑜𝑏𝑗 ) (3)
Within the range [0, 1], a normal distribution is used to create the random number 𝑟𝑎(0,1). The lower and upper
bounds of the 𝑗𝑡ℎ dimension are denoted by the symbols 𝑙𝑜𝑏 and 𝑢𝑝𝑏, respectively.
3.3.2. LDA based Feature extraction
LDA, utilized in statistics, pattern recognition, and machine learning, determines a linear combination of features to
express one dependent variable. Unlike PCA and factor analysis, LDA explicitly models differences between data
classes, rather than ignoring or focusing solely on similarities. LDA identifies vectors in the data space that best
discriminate between classes, creating a linear combination of independent features to maximize mean differences
between classes as per Eq. (4) and Eq. (5).
𝑛𝑗 𝑗 𝑗 𝐵
𝑠𝑤 = ∑𝑐𝑗=1 ∑𝑖=1(𝑥𝑖 − 𝜇𝑗 )(𝑥𝑖 − 𝜇𝑗 ) (4)
𝑗
𝑥𝑖denotes the 𝑖𝑡ℎ sample of class, 𝜇𝑗 is the mean of class 𝑗, 𝑐 represents the number of classes, 𝑛𝑗 signifies the
number of samples in class 𝑗, 𝜇 denotes the mean of all classes.
𝑇
𝑠𝑤 = ∑𝑐𝑗=1(𝜇𝑗 − 𝜇)(𝜇𝑗 − 𝜇) (5)

IJRTI2403070 International Journal for Research Trends and Innovation ([Link]) 488
 © 2024 IJRTI | Volume 9, Issue 3 | ISSN: 2456-3315

Integrating cloud capabilities into the dimensionality reduction step enhances scalability, efficiency, and flexibility in
managing large-scale data processing tasks. By leveraging cloud infrastructure and services, organizations can
accelerate dimensionality reduction computations and optimize algorithm performance, ultimately improving the
quality of downstream analytics and decision-making processes.
3.4. Multi-Model Selection
This section will detail various algorithms applied to the dataset for predictive modeling, including existing models
RNN, MLR, and proposed CANN.
3.4.1. RNN
An RNN is a neural network designed for sequential data. Unlike feedforward neural networks, RNNs incorporate
connections that create directed cycles, enabling them to retain hidden state information and process sequential input.
This makes them particularly effective for tasks involving time series data or natural language processing, where the
order of input elements matters. RNNs utilize recurrent units, such as the Gated Recurrent Unit (GRU) and Long
Short-Term Memory (LSTM), to manage long-term dependencies and update hidden states over time. They employ
techniques like Backpropagation through Time (BPTT) to update parameters based on the sequential structure of the
data. Additionally, RNNs are deployed in security frameworks to analyze incoming data continuously, detecting
anomalies or deviations from expected patterns, which may indicate security threats or abnormalities in the system.
The hidden state ℎ𝑠𝑡 at time 𝑡 in an RNN is computed using Eq. (6).
ℎ𝑠𝑡 = 𝜎(𝑤ℎ𝑠𝑖 𝑖𝑡 + 𝑤ℎ𝑠ℎ𝑠 ℎ𝑠𝑡−1 + 𝑏𝑖ℎ𝑠 ) (6)
𝑖𝑡 represents the input at time, 𝑤ℎ𝑠𝑖 is the weight matrix for the input, 𝑤ℎ𝑠ℎ𝑠 is the weight matrix for the hidden state,
𝑏𝑖ℎ𝑠 is the bias term, and 𝜎 denotes the activation function, typically a non-linear function such as the sigmoid or
hyperbolic tangent (𝑡𝑎𝑛ℎ). The output 𝑜𝑡 at time 𝑡 is computed based on the hidden state.
𝑜𝑡 = 𝑠𝑜𝑓𝑡𝑚𝑎𝑥(𝑤𝑜ℎ𝑠 ℎ𝑠𝑡 + 𝑏𝑖𝑜 ) (7)
Where, 𝑤𝑜ℎ𝑠 is the weight matrix connecting the hidden state to the output, 𝑏𝑖𝑜 is the bias term, 𝑠𝑜𝑓𝑡𝑚𝑎𝑥 is the
activation function, commonly used for multi-class classification problems. The RNN processes sequences by
iterating through time steps, updating the hidden state at each step based on the current input and the previous hidden
state.
3.4.2. CANN
CNNs, a type of deep learning model, are adept at learning and extracting significant patterns from raw data like
images. They comprise convolutional, pooling, activation, fully connected layers, and an output layer. Convolutional
layers extract features using filters, while pooling layers downsample features. Activation functions introduce non-
linearities, and fully connected layers make predictions. During training, CNNs learn to optimize their parameters
through techniques like backpropagation and gradient descent, minimizing a chosen loss function. This iterative
process allows CNNs to adapt and improve their ability to recognize and classify objects in images. The architecture
of an ANN typically consists of three types of layers: input layer, hidden layers, and output layer. Each layer contains
neurons or nodes, and connections between these neurons carry information. The CANN architecture is depicted in
Figure 2.
• Input Layer
The input layer forwards raw data or features to hidden layers for processing, with one neuron per feature. No
computation occurs within this layer.

Figure 2: CANN Architecture

IJRTI2403070 International Journal for Research Trends and Innovation ([Link]) 489
 © 2024 IJRTI | Volume 9, Issue 3 | ISSN: 2456-3315

• Hidden Layers
Hidden layers carry out the primary computation in neural networks. Neurons within a hidden layer receive inputs
from the preceding layer, compute a weighted sum, and apply an activation function to generate an output. The
configuration of hidden layers is tailored to problem complexity.
𝑛𝑖−1
𝑎𝑖𝑗 = 𝑓(∑𝑘=1 𝑤𝑖𝑘 𝑎𝑖𝑘 + 𝑏𝑖𝑗 ) (8)
The output of each neuron in a hidden layer is calculated using a weighted sum of inputs and a bias term.
• Output Layer
The output layer generates the neural network's final output, with the number of neurons determined by the problem
type.
3.4.3. MLR
MLR is a statistical method employed to forecast the outcome of a variable by considering the values of two or more
independent variables. It extends simple linear regression by accommodating multiple predictors. The model assumes
a linear relationship between the dependent variable and each independent variable. By estimating the regression
coefficients, it quantifies the impact of each independent variable on the dependent variable, allowing for the
prediction of outcomes based on known values of the predictors. The formula for multiple linear regression is given as
per Eq. (9).
𝑙𝑖 = 𝛽0 + 𝛽1 𝑥𝑖1 + 𝛽2 𝑥𝑖2 + ⋯ + 𝛽𝑜 𝑥𝑖𝑜 + 𝜖𝑖 (9)
𝑙𝑖 represents predicted value. 𝛽0 denotes y-intercept, 𝛽1 , 𝛽2 , … , 𝛽𝑜 represents regression coefficients. 𝑥𝑖1 , 𝑥𝑖2 , … , 𝑥𝑖𝑜 are
the values of the independent variables for the 𝑖-th observation. 𝜖𝑖 is the model’s random error term.

4. RESULT AND DISCUSSION

The experiment employed the Waikato Environment for Knowledge Analysis (WEKA) for feature selection and
extraction. BestFirst search with CfsSubsetEval was used for IABC-based feature selection, yielding 33 variables.
LDA, using Ranker search on Components, ranked all 117 attributes and combined them to create new features. A
threshold of 0.5 standard deviation was used to retain 20 attributes for prediction models using MLR, CNN, and RNN
classifiers.
Table 1: Comparison of algorithms between IABC and LDA
Proposed IABC LDA
Algorithm
MAE RMSE MAE RMSE
MLR [22] 1.5972 2.0409 1.6149 2.0145
RNN [22] 1.6212 2.7125 2.0305 2.8442
Proposed CANN 1.4512 2.0122 1.6973 2.1607

Each model's performance was evaluated based on error measures, as shown in Table 1. The proposed CANN
classifier, based on IABC, showed superior performance with the lowest MAE of 1.4512 and RMSE of 2.0122. In
contrast, multiple linear regression, in LDA-based models, achieved the lowest MAE and RMSE values of 1.6149 and
2.0145, respectively. Generally, IABC resulted in lower error rates compared to LDA. Multiple linear regression and
RNN performed better with IABC, while CANN excelled with LDA, highlighting the importance of selecting the
appropriate feature selection or extraction method based on dataset characteristics and machine learning algorithms.
Figure 3 illustrates a graphical comparison between the proposed IABC and LDA algorithms, showcasing their
performance in feature selection and extraction.

IJRTI2403070 International Journal for Research Trends and Innovation ([Link]) 490
 © 2024 IJRTI | Volume 9, Issue 3 | ISSN: 2456-3315

2.5

Values 2

1.5 MLR
RNN
1
ANN
0.5

0
MAE RMSE MAE RMSE
IABC LDA
Figure 3: Comparison of IABC and LDA algorithms graphically

5. CONCLUSION
This work developed predictive analytics to improve risk assessment in banking and insurance. Utilizing an actual
dataset with more than a hundred anonymized features, the study performs a comprehensive analysis and makes
recommendations for enhancing risk assessment in bank insurance operations. The system guarantees the acquisition
of comprehensive data sources pertinent to risk evaluation, starting with dataset collection. In order to preserve data
integrity and consistency, the dataset is cleaned of noise, outliers, and missing values using subsequent data pre-
processing procedures. After that, dimensionality reduction techniques are used to improve computational
performance and streamline the dataset. To be more precise, characteristics that are important for risk assessment are
found and extracted using techniques based on LDA and proposed IABC feature selection. Large-scale dataset
processing requires flexibility and scalability, which is made possible by utilizing cloud infrastructure. Using different
predictive modeling techniques, such as existing models RNN, MLR, and proposed CANN, which each offer distinct
viewpoints on risk assessment, is part of the multi-model selection phase. The framework’s ultimate goal is to
improve risk assessment procedures in the banking and insurance sectors, which will improve the precision of
decisions and operational effectiveness.

CONFLICTS OF INTEREST
The authors declare no conflicts of interest.
DATA AVAILABILITY STATEMENT
Not Applicable

REFERENCES:
1. Mukhopadhyay, A., Chatterjee, S., Bagchi, K. K., Kirs, P. J., & Shukla, G. K. (2019). Cyber risk assessment
and mitigation (CRAM) framework using logit and probit models for cyber insurance. Information Systems
Frontiers, 21, 997-1018.
2. Ashraf, B. N., Zheng, C., Jiang, C., & Qian, N. (2020). Capital regulation, deposit insurance and bank risk:
International evidence from normal and crisis periods. Research in International Business and Finance, 52,
101188.
3. Gaganis, C., Hasan, I., Papadimitri, P., & Tasiou, M. (2019). National culture and risk-taking: Evidence from
the insurance industry. Journal of Business Research, 97, 104-116.
4. Dhieb, N., Ghazzai, H., Besbes, H., & Massoud, Y. (2020). A secure ai-driven architecture for automated
insurance systems: Fraud detection and risk measurement. IEEE Access, 8, 58546-58558.
5. Aslam, F., Hunjra, A. I., Ftiti, Z., Louhichi, W., & Shams, T. (2022). Insurance fraud detection: Evidence
from artificial intelligence and machine learning. Research in International Business and Finance, 62, 101744.
6. Grima, S., Kizilkaya, M., Sood, K., & ErdemDelice, M. (2021). The perceived effectiveness of blockchain for
digital operational risk resilience in the European Union insurance market sector. Journal of Risk and
Financial Management, 14(8), 363.
7. Severino, M. K., & Peng, Y. (2021). Machine learning algorithms for fraud prediction in property insurance:
Empirical evidence using real-world microdata. Machine Learning with Applications, 5, 100074.

IJRTI2403070 International Journal for Research Trends and Innovation ([Link]) 491
 © 2024 IJRTI | Volume 9, Issue 3 | ISSN: 2456-3315

8. Carmona, P., Climent, F., & Momparler, A. (2019). Predicting failure in the US banking sector: An extreme
gradient boosting approach. International Review of Economics & Finance, 61, 304-323.
9. Hanafy, M. O. H. A. M. E. D., & Ming, R. (2021). Using machine learning models to compare various
resampling methods in predicting insurance fraud. J. Theor. Appl. Inf. Technol, 99(12), 2819-2833.
10. Vosseler, A. (2022). Unsupervised insurance fraud prediction based on anomaly detector ensembles. Risks,
10(7), 132.
11. Subudhi, S., & Panigrahi, S. (2020). Use of optimized Fuzzy C-Means clustering and supervised classifiers
for automobile insurance fraud detection. Journal of King Saud University-Computer and Information
Sciences, 32(5), 568-575.
12. Chaudhry, S. M., Ahmed, R., Huynh, T. L. D., & Benjasak, C. (2022). Tail risk and systemic risk of finance
and technology (FinTech) firms. Technological Forecasting and Social Change, 174, 121191.
13. Dalal, S., Seth, B., Radulescu, M., Secara, C., & Tolea, C. (2022). Predicting fraud in financial payment
services through optimized hyper-parameter-tuned XGBoost model. Mathematics, 10(24), 4679.
14. Settipalli, L., & Gangadharan, G. R. (2023). WMTDBC: An unsupervised multivariate analysis model for
fraud detection in health insurance claims. Expert Systems with Applications, 215, 119259.
15. Fursov, I., Kovtun, E., Rivera-Castro, R., Zaytsev, A., Khasyanov, R., Spindler, M., & Burnaev, E. (2022).
Sequence embeddings help detect insurance fraud. IEEE Access, 10, 32060-32074.
16. Găbudeanu, L., Brici, I., Mare, C., Mihai, I. C., & Șcheau, M. C. (2021). Privacy intrusiveness in financial-
banking fraud detection. Risks, 9(6), 104.
17. Hashemi, S. K., Mirtaheri, S. L., & Greco, S. (2022). Fraud Detection in Banking Data by Machine Learning
Techniques. IEEE Access, 11, 3034-3043.
18. Kolli, C. S., & Tatavarthi, U. D. (2020). Fraud detection in bank transaction with wrapper model and Harris
water optimization-based deep recurrent neural network. Kybernetes, 50(6), 1731-1750.
19. Zhou, H., Sun, G., Fu, S., Wang, L., Hu, J., & Gao, Y. (2021). Internet financial fraud detection based on a
distributed big data approach with node2vec. IEEE Access, 9, 43378-43386.
20. Subudhi, S., & Panigrahi, S. (2020). Use of optimized Fuzzy C-Means clustering and supervised classifiers
for automobile insurance fraud detection. Journal of King Saud University-Computer and Information
Sciences, 32(5), 568-575.
21. The Kaggle Website. [Online]. [Link]
22. Xu, W., Wang, Q. and Chen, R., 2018. Spatio-temporal prediction of crop disease severity for agricultural
emergency management based on recurrent neural networks. GeoInformatica, 22, pp.363-381.

IJRTI2403070 International Journal for Research Trends and Innovation ([Link]) 492
View publication stats