بررسی تأثیر سیگنال‌‌های معاملاتی منتشرشده در کانال‌‌های تلگرامی بر ارزش معاملات بازار سرمایه با به‌کارگیری الگوریتم‌‌های یادگیری ماشین

Objective
The expansion of social networks and messaging platforms in recent years has fundamentally transformed the structure of information circulation in financial markets. Due to its high sensitivity to information, expectations, and collective investor behavior, the capital market is among the most affected domains by this transformation. In this context, Telegram, as one of the most widely used platforms among Iranian capital market participants, plays a significant role in the dissemination of analyses, news, rumors, and trading signals. Many retail investors, alongside fundamental and technical analyses, rely on the content published on Telegram channels when making decisions about buying or selling stocks. Therefore, an empirical and scientific examination of the relationship between trading signals published in these channels and actual market behavior can contribute to a deeper understanding of how social media influence market dynamics. The main objective of this study is to investigate the relationship between the number of trading signals published in stock-market-related Telegram channels and the daily trading value of the Iranian capital market. The key innovation of this research lies in its focus on “trading signals” rather than general news, its use of Persian-language data from Telegram, and its examination of “daily trading value” as an indicator of market dynamism, whereas most previous studies have primarily focused on price, return, or the overall market index.
Research Methodology
This study is applied in terms of purpose and descriptive-analytical and correlational in terms of method. It was conducted using data mining, text mining, and machine learning approaches. The research data consists of two main categories: first, textual data extracted from Telegram channels active in the field of the stock market; and second, numerical data related to the daily trading value of the stock market. The textual dataset includes messages published in stock-market-related Telegram channels during the year 1401 in the Iranian calendar. The selected sample consisted of channels that regularly published analyses and buy/sell recommendations for stocks. In the numerical section, the study used the daily value of retail stock and pre-emptive rights transactions in the Tehran Stock Exchange and Iran Fara Bourse, excluding block trades, over 237 trading days in 1401. The textual data were extracted using crawlers and automated data collection tools. In total, more than 3.5 million messages were included in the analysis process. Among these data, approximately 5,000 messages were manually labeled by an individual familiar with capital market terminology. The messages were classified into three categories: “buy signal,” “sell signal,” and “neutral/news related.” The main criterion for labeling was the presence of an explicit or implicit recommendation to buy or sell a stock in the text of the message.
In the preprocessing stage, tokenization, noise removal, deletion of unnecessary characters, letter normalization, stop-word removal, and stemming were performed on the texts so that raw social media data could be transformed into standardized and processable data. Subsequently, numerical representation of the texts was carried out using methods such as Bag-of-Words, TF-IDF, and transformer-based models, including BERT. In the modeling stage, supervised machine learning algorithms, including Naive Bayes, Support Vector Machine, and a BERT-based model, were used to classify the messages. The labeled data were divided into training, validation, and test sets. After model evaluation, the trained model was applied to classify all extracted messages. Then, the daily number of buy signals, sell signals, and neutral messages was calculated and matched with daily trading value data. Pearson’s correlation coefficient was used to examine the relationship between the variables. In addition, the performance of the classification model was evaluated using the confusion matrix and the indices of accuracy, precision, and recall.
Findings
The modeling results showed that the designed algorithm was able to classify Telegram messages into the three categories of buy, sell, and neutral with acceptable accuracy. In the final evaluation, with a focus on the two main classes of “buy” and “sell,” the model achieved an overall accuracy of 96.6%. Moreover, the model’s precision and recall were reported to be approximately 89.15% and 90.82%, respectively, indicating a balanced performance in identifying trading signals. In particular, the model performed very well in detecting sell signals, with precision of about 98% and recall of about 97%. This finding suggests that the linguistic pattern of selling signals in Telegram messages is relatively clearer and that the machine learning model was able to identify these patterns successfully. In the statistical analysis section, Pearson’s correlation coefficient indicated a positive and relatively strong relationship between the number of sell signals and the daily trading value of the market, with a correlation coefficient of 0.80. The number of buy signals also showed a positive correlation of 0.72 with daily trading value. Neutral or news-related messages had a weaker correlation with trading value, equal to 0.58. The difference between the number of buy and sell signals showed a correlation of 0.61 with trading value. These results indicate that trading signals, particularly sell signals, have a stronger association with increased levels of trading activity in the market compared with neutral or news-related messages. In other words, whenever signaling activity in Telegram channels intensified, the daily trading value of the market also tended to increase.
Discussion & Conclusion
The findings of this study indicate that unstructured data published on social networks can, if properly processed, be transformed into quantitative and analytical indicators. The study demonstrated that Telegram trading signals can be regarded as reflections of the sentiments and expectations of capital market participants and that they have a statistically significant relationship with the level of daily trading dynamism. From a theoretical perspective, the results are consistent with the principles of behavioral finance and with approaches that emphasize the role of informal information, collective expectations, and imitative behaviors in investors’ decision-making processes. Due to their high speed of dissemination and broad user accessibility, social networks can play an influential role in shaping collective perceptions of market conditions and, through this mechanism, affect trading intensity. Nevertheless, it should be emphasized that the findings of this study indicate correlation rather than a definitive causal relationship. Therefore, it cannot be concluded with certainty that Telegram signals alone caused an increase in trading value. Rather, it can be stated that increased signaling activity on Telegram was accompanied by an increase in trading value. From an applied perspective, the results of this study may be useful for capital market supervisory institutions, the Securities and Exchange Organization, analysts, and investors. The development of intelligent social media monitoring tools can help identify abnormal behaviors, track waves of trading signals, reduce the possibility of market manipulation, and enhance informational transparency. Furthermore, investors can make more informed decisions by understanding both the role and the limitations of trading signals published on social networks.