Expert consultations are particularly appropriate when conducting qualitative research, if the subject of study is not the same in all its dimensions as in previous studies. This approach was chosen for data collection due to the fact that compliance reporting has not been discussed in more detail to date. Interviews typically capture the opinions of experts with superior knowledge of the subject of compliance reporting resulting from their role within the organization. Therefore, in contrast to in-person interviews, the researcher is more interested in the respondents’ roles as subject matter experts than in the interviewees themselves. knowledge bases, relevance structures, and socially held interpretations and constructions of reality are all implied in addition to the knowledge of a single individual. Theories are developed based on the information gathered from the interviews. The initial systematic theory discovery from the data is stated as the best way to be reasonably assured that the theory would fit and function. For this reason, the data were assessed using the grounded theory method. Its foundation is the methodical creation of hypothesis from data that is methodically gathered from social science research. Consequently, the grounded theory approach offers a methodical, exacting manual for developing theories. This approach of qualitative research aims to construct theories based on factual information. The goal of the theory is to identify the higher social structures that are responsible for the construction and reconstruction of pertinent conceptions as well as to illuminate the significance of human experiences. Grounded theory methodology is used to develop theories that are pertinent to the compliance reporting study issue ( Panitz et al., 2010 ). This study focuses on the insurance sector in Sudanese companies listed on the Khartoum Stock Exchange until 31 December 2022.

To conduct the expert interviews, a questionnaire was developed based on the research background to ensure that all topics relevant to the research area were addressed. As a subsequent measure, a pre-test was conducted to check the survey for suitability, and the questionnaire form was revised during a practical interview with an expert. The questionnaires were also revised to ensure ease of understanding and additional topics identified as important were added.

Since they are in charge of confirming compliance and are knowledgeable about compliance standards, 53 questionnaires were given to the study sample, which consists of internal audit managers and internal auditors in the companies under investigation. After completing the questionnaire, forty of them returned it.

To evaluate the accuracy of the prediction, the researcher used four statistical measures: mean square error (MSE) and root mean square error (RMSE), symmetric mean absolute percentage error (SMAPE), mean absolute percentage error (MAPE) ( Reda Abonazel, 2018 ). The formulas for MSE, RMSE, SMAPE, and MAPE are as follows ( Torsen et al., 2018 ).

$Meansquareerror\left(MSE\right)\text{=}\frac{1}{n}{{\displaystyle \underset{i=1}{\overset{n}{\sum }}\left(y_i-y_i^{\wedge }\right)}}^2$ (1)

$RootMeansquareerror\left(RMSE\right)=\sqrt{MSE}$ (2)

$SymmetricMeanAbsolutePercentageError\left(SMAPE\right)=\frac{1}{N}{\displaystyle \underset{i=1}{\overset{n}{\sum }}\frac{\left|y_{i-}{y}_i^{\wedge }\right|}{\frac{\left(\left|y_i\right|+\left|y_i^{\wedge }\right|\right)}{2}}}$ (3)

$AbsolutepercentageError\left(MAPE\right)=\frac{1}{n}{\displaystyle {\sum }_{i=1}^n\left|\frac{e_i}{y_i}\right|}$ (4)

Multi-Layer Perceptron (MLP) is a neural network architecture consisting of an input layer, an output layer, and a single or more hidden layers for many stacked neurons. In the perceptron model ( Hrizi et al., 2022 ). A neuron is required to include an activation process. A function imposes a threshold, such as the sigmoid function ( Abu Al-Haija et al., 2022 ). However, in the case of multi-layered sensory perception, Neurons have the flexibility to use any arbitrary activation function ( Miller et al., 2018 ). The diagram provided shows the structure in Figure 1 . The components of the multi-layer Perceptron network.

Decision tree is a classification and regression modeling tool and is an example of an algorithm that uses supervised learning ( Hamdi et al., 2023 ). Since regression is a technique for implementing predictive modeling, these trees can either classify data or make predictions about what will happen in the future ( Kamiński et al., 2018 ). Decision trees are diagrams that resemble flowcharts, and start at the root node with a particular data inquiry. In the field of machine learning, decision tree algorithms are utilized for both prediction and classification purposes. With the use of the decision tree and a predetermined collection of inputs, one is able to map the numerous outcomes that are the result of the decisions or consequences ( Povhan, 2020 ). The structure of DT follows a hierarchical tree pattern, wherein the central root node is surrounded on all sides by branches, internal nodes, and leaf nodes. Root nodes are the initial nodes situated at the inception of a decision tree. At this juncture within the tree, the population will commence to undergo segmentation into multiple groups based on the shared features they possess. The nodes that persist in the tree structure subsequent to the severance of the root nodes are commonly denoted as internal nodes. The term “Decision Node” is employed to designate these nodes. the leaf that Possesses Nodal Structures The plant’s leaf nodes or terminal nodes are defined as the nodes located on the leaves that lack the ability to undergo further division into additional divisions ( Povkhan, 2020 ). show the Decision Tree Classification Algorithm and presented in Figure 2 .

Sarbanes-Oxley Act and other laws impose strict compliance standards, and businesses are forced to invest a lot of money to comply with them ( Ghose & Koliadis, 2007 ). Timely implementation of Compliance shields businesses from fines and other legal repercussions from the government by assisting in the prevention and detection of legal infractions of laws or regulations (2022). Rules governing corporate governance are yet another important component of compliance. Violations of corporate governance regulations will raise operational risks within the firm, thereby impacting stakeholders’ interests ( Zerban & Madani, 2018 ). There is an increasing need to appoint a compliance officer and conduct compliance audits ( Thottoli, 2022 ). The quantity, intricacy, and significance of the requirements for compliance are continuously increasing, prompting corporations to do so. Where various efforts are made to ensure compliance with legal regulations, regulatory standards or voluntarily imposed obligations ( Panitz et al., 2010 ).

One of the most recent developments in management is the Balanced Scorecard (BSC). Actually, it is a strategic instrument created by Kaplan and Norton and detailed in their book The Balanced Scorecard, published in 1996. Within the realm of business, the balanced scorecard has aroused intense curiosity at the international level ( Nørreklit, 2003 ). The goal of the balanced scorecard is to find a solution of the background of finance metrics for accounting systems. This is done Through the incorporation of financial and non-financial strategic measurement variables into a cause-and-effect connection that presupposes the following: growth and learning inside the organization metrics → internal company operations process metrics → customer perspective metrics → finance metrics. Assuming a cause-and-effect relationship between the proposed measurement domains is crucial because non-financial measurements domains Apply feed-forward control to the performance measuring system ( De Haas & Kleingeld, 1999 ). Measurement is as fundamental for managers as it has been for scientists ( Kaplan, 2009 ). One study also concluded that the balanced scorecard approach may require some substantial changes in culture within the organization. A company’s top management must comprehend, support, and be committed to the balanced scorecard. Balanced scorecards are going to evolve. The company will discover new metrics to use., and new objectives in several fields, to improve the effectiveness and balance of the balanced scorecard in achieving goals ( Chavan, 2009 ).

In order to increase performance in the future, machine learning is a model that may allude to learning from past experiences, which in this case are historical data. This field focuses only on autonomous learning techniques. Learning is the process of automatically adjusting or enhancing an algorithm without the need for outside assistance from a human based on prior “experiences” ( Das & Behera, 2017 ). It is a subfield of artificial intelligence (AI), mainly concerned with computer algorithms that can automatically discover and learn from data. When ML is properly integrated with data analytics, it can ultimately maximize the use of big data AI and current computer technology. Innovations have led to the rapid creation of powerful deep learning (DL) algorithms that enable them to improve prediction and reduce involvement from humans ( Surya, 2016 ). (ML) replaces its conventional statistical counterpart in forecasting future occurrences, which yields incredibly helpful findings and is applied in a variety of domains including engineering, the medical sciences, and finance ( Sarkar et al., 2019 ).

Machine learning is of great interest in finding optimal solutions in various fields ( El Misilmani & Naous, 2019 ). Machine learning is typically defined as a discipline that “concerns the question of how to create computer programs that get better on their own with use ( Kuznetsov, 2004 ). Many methods based on statistics (Bayesian networks, instance-based approaches) and artificial intelligence (logic-based, perceptual-based techniques) have been developed ( Kotsiantis et al., 2006 ). As evidenced by a number of current studies, machine learning approaches have garnered significant attention from the research community due to their superior classification accuracy when compared to alternative data categorization methods. Reaching extraordinary forecasting accuracy is essential because it can result in the right kind of protection ( Dwivedi, 2018 ).

The data employed in this investigation, according to the MLP, included a total of 180 records ( Table 1 ), of which 129 (71.70%) were used as training sets and 51 (28.30%) as test sets.

Table 2 shows the units in the hidden layers: Sigmoid function is used for the dependent variable of the output layer (focus), and sum of squares is used as the error function. The activation function of the hidden layer is hyperbolic tangent.

Figure 3 shows the Predictor pseudo probability” is a predictive model’s projected likelihood or score of an outcome based on input predictors. It helps grasp and interpret model predictions in predictive modeling and machine learning applications, where strongly agree varied between 0.4 and 1.

Figure 4 shows the Impact of Input Changes on Neural Network Output Sensitivity suggesting a direct correlation between the input and output.

Figure 5 illustrates how the gain of the network affects its output, with a steeper line indicating a higher gain. High gain amplifies the input signal to a greater extent, which can have both advantageous and possibly destabilising effects.

Classification and Regression Tree (CRT) is used as a growth methodology. The maximum depth of the tree is 3, the minimum number of cases in the parent node is 100, the minimum number of cases in the child node is 50, and the independent variables include 3 nodes. The number of end nodes is 2 while the depth result is 1. And shown in Table 3 .

It is clear from Table 4 that the degree of response of respondents in the study sample regarding the statements of the dependent variable reached 3 for strongly agreeing at a rate of 50% and 2 for agreeing at a rate of 33%, while the number of neutral ones was 1 at a rate of approximately 17%.

Figure 6 shows that the number of neutrals in node 1 is 3, those who agree are 35, and those who strongly agree are 31, and the total percentage for the first node is 38.3%. As for node 2, the number of neutrals is 4, those who agree are 4, and those who strongly agree are 102, and the total percentage for the second node is 61.7%.

One of the most straightforward ways to interpret forecast values is to compare them across different forecasting models or techniques. In general, lower prediction values indicate higher accuracy, while higher prediction values indicate less accurate predictions. Looking at Table 5 , it was found that the prediction accuracy measures in the decision tree model are low compared to the multi-layer receiver model, which indicates the prediction accuracy of the decision tree model.