Introduction                                                                        

Credit scoring and credit, in general, are pivotal instruments used by both private and public financial institutions. Credit offers the cash flow required for the development of all types of economic activity (Doumpos et al., 2019). The system of credit scoring allows analysis according to numerical values and decisions based on the applicant's overall score (Abdou & Pointon, 2011).

The enhanced credit scoring model considers a person's cognitive capacity and financial data, whereas the traditional credit scoring model is only founded on financial data (Kulkarni & Dhage, 2019). Lenders should expect to lower their risk in the short term by integrating network-based solutions, according to financial sector managers (Wei et al., 2016). To properly determine a person's financial situation, reliability and overall personality traits are also required (Kulkarni & Dhage, 2019). Credit and behavioral scoring have emerged as essential techniques for predicting financial risk in lending activities and assisting institutions in dealing with the risk of default in consumer loans (Thomas, 2000).

When consumers purposefully adjust their networks to enhance their score and deceitfully reap the rewards that result from it, the effectiveness of social scoring does not necessarily have to decline (Wei et al., 2016). There is no one ideal statistical methodology for designing credit scoring models and no technique exists for all data sets (Abdou & Pointon, 2011). Social scoring is more likely to be effective in online or offline communities and societies where customers have strong relationships (Wei et al., 2016).

Adding qualitative variables such as the depth and breadth of a person's social media relationships and alternative payment data modifies the algorithms used to calculate credit scores. As a result, the credit score will be determined by the users' prestige and social connections. Users will have access to their credit scores using the application, encouraging fiscal discipline in their lives (Lohokare et al., 2017).

Credit scoring has become a critical responsibility since the credit corporate has doubled over the last two decades. Because of its linked memory characteristic and generalization capabilities, the artificial neural network is becoming a preferred solution in credit scoring models (Lee et al., 2002). The main issue of credit scoring, which is a critical study topic in the banking sector, is anticipating bankrupts to acquire potentially profitable clients. The suggested technique divides consumers into five categories: best, good, satisfactory, bad, and worst (Marikkannu & Shanmugapriya, 2011).

In recent years, social media used for suggesting trades on the stock market has gained noticeable traction in Iran. Therefore, the introduction of new validation models to predict the validity of signals published by active telegram channels in the stock is one of the basic needs of the present age, which is the result of this compelling and practical research in this field. By examining various articles and methods used in them, it can be said that selecting the appropriate technique for the validation of active telegram channels offering stock signals is still an open research issue. The selection of the proper technique based on data and the use of advanced data mining techniques can lead to an efficient model for the sample. Selecting appropriate features based on the data collected will also be important in validating the review due to the lack of criteria to rely on the opinions expressed in the Iranian Telegram channels about the offers to buy and sell shares. So, the need to conduct this research seeks to implement new validation methods using the Telegram channels as one of the most popular offering platforms. Buying and selling stocks are considered in this analysis. What is worth mentioning is that we have implemented a novel approach for identifying reliable active stock traders through Iranian Telegram channels. By reviewing previous papers, we recognised that to identify attributes affecting the validation of stock traders that are active in the field of offering stock signals, we should utilize the concept of credit scoring to meet our scientific goal. So, we set our first question as below:

1. What are the factors affecting the validation of telegram channels active in the field of offering stock signals?

As social media platforms in Iran that recommend stock market trading have gained a lot of attention, we should do the related credit scoring method by considering financial and non-financial features existing in previous studies. Studies have also emphasized that a person's social media involvement can be used to assess their financial situation (Kulkarni & Dhage, 2019).

The financial situation in our case is the rate of return on stocks which was suggested by Iranian Telegram channels at one-month intervals. In other words, we had determined the ability of Iranian Telegram channels to propagate the right signal about buying or selling Iranian stocks. The importance of credit scoring Iranian Telegram channels propagating the right signals is that it can guide stockholders to gain economic benefits. This result has the potential to be categorized as a credit scoring concept. Due to the meaning of credit scoring, utilizing scoring in credit evaluation eliminates the need for personal judgment (Abdou & Pointon, 2011). We have also implemented the mentioned concept to identify the credit qualities of Iranian Telegram channels offering stock signals for reducing personal judgment.

As mentioned earlier, by doing a literature review, social and financial variables were extracted to identify factors affecting the validation of Iranian Telegram channels active in offering stock signals. Furthermore, to enrich our results, we realised the relation between Iranian market conditions and the rate of return on stocks recommended by Iranian Telegram channels offering stock signals. This analysis helps stockholders figure out appropriate features that can broaden their minds to find the creditworthiness of Iranian Telegram channels that propagate the right signals about buying or selling Iranian stocks. As a result, we set our first question as below:

2. Due to what market conditions can the credibility of telegram channels active in the field of stock offering be trusted?

This research is divided into five main chapters: In the first chapter, a brief description of the main subject of the study and its importance is given with a short reference to the method and steps of the study. In the second chapter, the literature is the subject, which includes an overview of the validation concepts, models of validation implementation, data mining, and its application in the issues related to validation. The third chapter focuses on the research methodology and provides a complete description of how the data have been collected and the variables have been defined. Also, in this chapter, the research structure based on Crisp methodology is explained step by step to explain all the steps. Then, the results of data processing and implementation of different models with the help of calculations, comparisons, and analysis of results based on the Crisp method are examined in detail. Finally, the research achievements and their innovative aspects are discussed, and the suggested tips for future research are stated.

Literature Review

The findings suggest that combining financial and non-financial elements leads to a more accurate prediction of future default occurrences than separately using each of these components (Grunert et al., 2005). The conventional technique of credit scoring relied on statistical analysis and the common judgment method. However, the novel techniques are centered on utilizing social media data, mobile data, and psychometrics (Hendricks & Budree, 2019). Classical credit reporting, which is primarily focused on financial data, may only provide such good assessment of SMEs that are restricted in financial data but flourishing in non-financial data, like big data from business, government, social media, and networks (Yadi et al., 2019). The significance of identifying financial and non-financial indicators to evaluate the credit risk associated with microfinance loan applications is inevitable (Abdullah et al., 2020).

Since some consumers' credit information is inadequate or unavailable, credit managers cannot assess their actual credit position. This difficulty, nevertheless, may be efficiently handled with the use of social data, particularly behavioral data, and a credit rating system (Yu et al., 2019). Social media has become more significant in spreading individual perspectives on a wide range of financial topics, constituting credit risk in investment decisions (Fei et al., 2015). With the collection of data on personal behavior and the advancement of machine learning techniques, social media data may now be used for personal credit rating (X. Yu et al., 2020). When financial or non-financial evaluations are absent or inaccurate, and credit analysts' subjective opinions affect the decision, social media data may be precious in evaluating organizations' credibility (Gül et al., 2018).

Credit scoring is becoming a reality due to the advancement of big data and machine learning models and algorithms (X. Yu et al., 2020). Artificial neural networks will allow the final credibility score to be calculated based on the different obtained data characteristics. Through SMS, a smartphone application will gather bank transaction data and data connected to internet transactions. Adding qualitative variables such as the depth and breadth of a person's social media networks and alternative payment data refines the algorithms used to calculate credit scores (Lohokare et al., 2017).

The proposed study effort intends to contribute to examining the behavioral features of online influencers is accomplished by calculating credit scores and examining the time development of these scores and time-series trends (Daou, 2019).

Related to the basis of the credit scoring model based on the decision tree algorithm, the borrowed credit score, the number of successes, prestige, the number of failures, repayment duration, and forum currency is the most important characteristics for forecasting default based on the credit scoring model and classification criteria (Zhang et al., 2016).

The Long Short-Term Memory model dynamically includes daily news from social media to account for market players' and public opinion views. Leveraging unstructured news crawled from social media to mitigate the influence of financial fraud on default probability prediction. Also, a neural network technique is presented for default probability prediction that integrates both structured financial parameters and unstructured social media data with proper time alignments (Zhao et al., 2019).

To build a credible personal credit evaluation system using social media data, to identify abnormal users, the logistic regression approach was utilized to score the individual credits of users before and after cleaning the data and discovered that the rank order of personal credit scoring has changed substantially (X. Yu et al., 2020).

Through the utilization of Logistic Regression, while combining a variety of information sources such as consumer purchasing habits, social media activity, and geography details with conventional credit scoring methodologies yields fresh insights for appropriate and more accurate credit scoring (Hindistan et al., 2019).

To verify borrowers employing classification and regression trees, digital lenders usually gather vast quantities of data from their clients, such as communication patterns, social media activity data, and extensive mobile phone usage (Shema, 2019). By improving an ensemble learning model with a random forest, a novel computerized approach for extracting public emotions encoded in social postings is created to enhance standard financial indicators, like return-on-assets, for forecasting corporate credit ratings (Yuan et al., 2018). Regardless of whether a person has incredible, good, or bad credit, it is critical to review their credit situation. Money-related reliability, as well as a person's overall identity characteristic, are needed. This may be determined by a person's social media interaction with the means of the Naive Bayes classifier. Combining these two scores will provide the weighted advanced credit score, which will be more precise than the current method (Kulkarni & Dhage, 2019).

Research questions

1. What are the factors affecting the validation of telegram channels active in the field of offering stock signals?
2. Due to what market conditions can the credibility of telegram channels active in the field of stock offering be trusted?

Research Methodology

The Cross-industry standard process expands on earlier efforts to describe knowledge discovery techniques, commonly known as CRISP-DM, which are applied for data mining. The six critical steps in the process are provided below (Wirth & Hipp, 2000):

1. Business understanding
2. Data understanding
3. Data preparation
4. Modeling
5. Evaluation
6. Deployment
7. Business understanding: This is the first stage in which we define the business's goal. In this paper, the goal is to figure out how stockholders can rely on telegram channels' social behavior to gain one-month returns. Among the purposes pursued in this area is to identify financial factors affecting the identification of high-credit channels, criteria related to their social activity, and determine the impact of each of the variables affecting the credit of active Telegram channels to understand shareholders. Improving the current situation and providing solutions to properly evaluate the active telegram channels in stocks have been among the objectives of this research. The RFM model is used to understand the channel conditions in different groups. com is an intelligent stock social analysis system that helps the audience with everything they need from stock exchange information and social media to choose good stocks, trades and appropriate times. With the help of this system, we crawled Telegram channels data in a time of six months.
8. Data understanding: Data gathering is the first stage in this phase. We begin analyzing the data when it has been collected. We investigate data insights and detect quality issues, as well as discover intriguing data subsets. This type of data analysis is a crucial stage in the study of the data. The period that was selected is approximately from July 2020 to January 2021. Moreover, 270 telegram channels were selected, based on existing data, to identify their social behavior.
9. Data preparation: It includes all actions necessary to create the final dataset from raw data. This step is used to choose the table, and attributes, and to transform/separate junk data/from clean and useful data for modeling techniques. In this study, the available data are actual figures and statistics obtained from active telegram channels in stock offers by real users. In this study, analysis was performed on all data, and no sampling was performed.

Based on the three criteria in the RFM algorithm, a database of the performance of telegram channels in these three sections was created (Birant, 2011). Part of this database can be seen in the following tables:

1. The recency of the last broadcast signal (R): This factor represents the last time the signal was released, which means the time interval of the last broadcast signal of the current channels. The shorter the time interval, the larger the value.
2. Signal Frequency (F): This factor indicates repetition. It represents the number of signals emitted by the channels in the time interval.
3. The monetary value of broadcast signals (M): This factor expresses monetary value. In this study, the difference between the performance of the right and wrong signs in each of the broadcasted signals of the channels has been calculated. In this variable, if they function correctly, the buy and sell signals must have led to the growth and decline of the stock, respectively. Neutral signs have also been removed to reduce complexity.

Table 1. RFM dataset before Normalization

Table 2. RFM dataset before Normalization

In the next step, a dataset related to understanding the state and importance of both financial and social criteria, which are imported from the literature review, is created:

Financial criteria:

1. Stock return rate: The following formula is used to compare the returns of shares. This return is compared to time stability. In such a way, the considered periods are compared with each other seasonally, annually, etc. The efficiency achieved in fewer time intervals is the best (Kumar et al., 2021).

Return =

Due to the limited time frame of the present study, this return has been considered only for thirty days. In this study, the denominator of the deduction is the final price of the share on the first day in thirty days, and the numerator also includes the final price thirty days later. Finally, the average of these returns for the channels is calculated (to reduce complexity, records for which there was no end price were separated from the others).

2. Performance score: The following formula has been used to calculate the performance score for the rate of positive and negative emotions published on Twitter, which includes the amount of difference between negative and positive emotions of each user divided by the total difference between negative and positive emotions of users (Leitch & Sherif, 2017).

Score =

In this study, according to the previous variable formula, the signals emitted by each channel, the difference between the performance of the right and wrong signals in each channel divided by all the signals with a positive and negative performance of the channels has been calculated. In this variable, if they function correctly, the buy and sell signals must have led to the growth and decline of the stock, respectively. In this case, TRUE and otherwise FALSE are considered. Neutral signals have also been removed to reduce complexity.

Social criteria:

1. Normal activity rates: The behavior of some users on social networks is not normal in such a way that they publish a lot of information in a limited period and then remain silent for a long time. Scores greater than 0.05 and less than 10 are considered good performance in this criterion. The denominator of the deduction includes the number of user activities (in this study, the number of published signals), and the denominator of the deduction consists of the difference between the first and last days of signal propagation in the interval (X. Yu et al., 2020).

 =

In the present study, this criterion is calculated weekly and in total.

2. Monthly average of published signals: To calculate the past behavior in relation to the loan recharge, the frequency of this recharge every month has been calculated as validation variables (Shema, 2019).

This study includes the total number of signals published in the period in question and the average weekly and monthly number of channel signals to offer shares.

3. Monthly average of active days to spread the signal: To calculate the past behavior concerning loan characteristics, the amount of active days after activating the account on a monthly basis has been calculated as validation variables(Shema, 2019).

This study includes the total number of active days to broadcast the signal in the period and the average weekly and monthly number of active days of the channels to post.

Table 3. Validation variables from review studies

4,5. Modeling and Evaluation: For modeling, we use data mining techniques in this step. Then, for evaluation, one of the primary goals is to determine whether any critical business issues have been overlooked. A choice on how to use the data mining results should be made near the end of this stage.

Our example study utilized a basic K-means method to define multiple clusters based on the RFM criteria. We used the R tool for the clustering process, a free software environment for analyzing data and implementing statical computing and graph representation (Hornik, 2012). The K-means clustering method attempts to arrange related objects into groups. K represents the number of groupings. K-means clustering is a well-known and very effective unsupervised machine learning technique. It is used to address a wide range of complex unsupervised machine learning issues (Malik, 2011). The elbow technique is the most used strategy for addressing this issue when using the k-means clustering algorithm. We compute WCSS, the within-cluster sum of a square, for each value of K. WCSS is the sum of the squared distances between each point in a cluster and its centroid. When we display the WCSS with the K value, we get an Elbow. The WCSS value decreases as the number of clusters rises. (Syakur et al., 2018).

By implementing the elbow method, four clusters are the best choice for this dataset (as it is shown in the following figure).

Figure 1. The Elbow Method

The center of each cluster and other information is detailed in table 4 :

Table 4. Cluster Centers

The view of the scattering of clusters is illustrated in figure 2:

Figure 2. Clusters Scatter Plot

Also, we used the Python tool, which has proven to be a complete programming solution to advance data science goals. Python's ease of learning and flexibility make it one of the fastest-growing languages. Python's evolving libraries also make it a convenient option for data analysis (Nagpal & Gabrani, 2019). We have used a Tree-based Pipeline Optimization Tool (TPOT) to create machine learning pipelines with competitive classification accuracy and identified innovative pipeline operators (Olson et al., 2016). The results of running this algorithm on the validation variables are as follows: ExtraTreesClassifier is proposed as the algorithm that performs best. ExtraTreesClassifier is an ensemble learning approach, also known as Extremely Randomized Trees (Bhati & Rai, 2020). It mainly consists of heavily randomizing both attribute and cut-point selection while splitting a tree node. In the most extreme example, it generates completely randomized trees with models independent of the learning sample's output values (Geurts et al., 2006). These algorithms were implemented in the Python tool. Then, cross-validation was implemented, which is a resampling procedure used to evaluate machine learning models. For each cluster, the accuracy of implementing this algorithm is: cluster1=0.705, cluster2=0.717, cluster3=0.764, cluster4=0.714.

It should be noted that the performance score variable is considered the target variable because it means the channel's overall performance, in terms of rising stock prices following the signal, compared to other channels. As a result, the importance of other criteria is considered concerning this variable.

Figure 3. TPOT result

Feature importance and summary of variables are shown in the following figures:

Cluster 1:

Figure 4. Factors of importance in cluster 1

Figure 5. Factors statistics in cluster 1

In figures 6,9,12 and 15, line charts named 1 and 2 are the positive and negative performance of signals, respectively.

Figure 6. signal trends in cluster 1

Table 5. Channels Analysis

4. Deployment: During the deployment process, we have created a report that includes the results. Through clustering, the telegram channels, hidden patterns in their behavior can be traced to identify the validity of their signal and improve shareholder decisions about relying on these signals to buy or sell shares.

Discussion and Conclusion

Now, using the data and results presented in this research, channels active in the field of stock signal propagation have been categorized, and for each category, an appropriate validation strategy should be adopted so that firstly, shareholders choose reputable channels to rely on for their analysis and secondly, channels are encouraged to improve the quality of their transmitter signals. With the help of this research, it is possible to validate the accuracy of the financial status of channels active in the field of stock signal propagation according to their different activities in the Telegram social network.

Now we should answer the research questions:

1. What are the factors affecting the validation of telegram channels active in the field of offering stock signals?

According to the literature review and the results of the algorithms implemented in this research are: Social variables include the rate of normal activity around the weekly and total stock signal propagation (abnormal behavior of some users on social networks is such that they publish a large amount of information in a limited period, and then do not send signals for a relatively long time) / weekly, monthly average, and the total number of signals and the weekly, monthly average and the total number of activity days around the signal propagation. Financial variables also include the rate of return on stocks at one-month intervals and the performance score (the difference between the right and wrong signals per channel over the total signals with positive and negative channel performance). Among the mentioned criteria, to identify the channels whose stock financial return status is less consistent with the type of their published signals, we can refer to the total number of active days when the stock signal is released (in total desired return) and calculate the normal performance rate. These channels paid more attention to this (as weekly reviews). The fact that channels whose number of activities peaks each week on certain days and then do not work cannot be trusted in the one-month horizon to the returns of stocks that have recommended buying or selling. Also, identifying channels with more credibility for one-month returns does not necessarily mean that they have the highest daily activities or the highest proportion of signals.

2. Due to what market conditions can the credibility of telegram channels active in the field of stock offering be trusted?

Examining the clusters; the primary and most reliable part when viewing the published signals around stock trading is that the channels usually perform more correctly around the forecast for stock buying or selling when the overall index grows positively. It would not be wise to rely on telegram channel signals when the growth of the overall index is negative, and the stock market is not in a favorable position. Also, the high number of activity days and the high amount of channel signals do not necessarily mean that they can give high credits to the returns of their proposed stocks in the one-month horizons.

In any research, some limitations reduce the accuracy of doing so. There are limitations in this research as well. The most important of these are:

1. Limited access to data in a short period, six months
2. Only 270 channels that could be shared between variables were extracted. Their number can be increased in future research.

Research suggestions: Other researchers are suggested to take action in the following cases:

1. If it is possible to access the data for a more extended period, in addition to using the RFM model, the LRFM model can also be used. This model, along with the previous three variables, has the variable L, which indicates the duration of the channels' activity in propagating the stock signal in the telegram.
2. The weight of RFM variables in this study is the same. But depending on the industry of research and the opinion of experts, these weights can have different degrees of importance.
3. A combination of other clustering algorithms can be used in future research.
4. Using neutral signals (neither buy nor sell) in credit scoring
5. Validation variables have been extracted according to the literature review. In future research, the opinion of stock industry experts can also be used in this regard.

Declaration of Conflicting Interests

The authors declared no potential conflicts of interest concerning the research, authorship and, or publication of this article.

Funding

The authors received no financial support for the research, authorship and, or publication of this article.