University of Isfahan Journal of Asset Management and Financing 2383-1189 14 3 2026 09 23 Designing an Early Warning System to Predict Price Bubbles in the Tehran Stock Exchange Using Deep Learning Designing an Early Warning System to Predict Price Bubbles in the Tehran Stock Exchange Using Deep Learning 31 58 29846 10.22108/amf.2025.143670.1943 FA Morteza Mazarei Ph.D. Candidate, Department of Financial Management, Kish International Campus, Tehran University, Kish, Iran 0009-0008-0601-7432 Ezatollah Abbasian Professor, Department of Financial Engineering, Faculty of Management, University of Tehran, Tehran, Iran Ebrahim Nasiroleslami Assistant Professor, Department of Statistics, Faculty of Basic Sciences, Bu-Ali Sina University, Hamadan, Iran Journal Article 2024 12 28 The primary objective of this study is to design an Early Warning System (EWS) based on a Long Short-Term Memory (LSTM) architecture for the timely forecasting of price bubbles in the Tehran Stock Exchange (TSE). A secondary objective is to compare the predictive performance of this model against a Logistic Regression (LR) benchmark, using evaluation metrics such as the AUC-ROC and confusion matrix. The system's performance was evaluated on five selected TSE indices. Utilizing monthly data from 2002 to 2023, bubble periods were identified via the Generalized Supremum Augmented Dickey-Fuller (GSADF) test and represented as a binary variable. This bubble variable was then modeled using price changes of key warning indicators. The proposed deep learning-based system achieved predictive accuracy ranging from 73% to 81%, with the highest performance on the Total Index (81%) and the lowest on the Basic Metals Index (73%). A comparative analysis demonstrates that the LSTM model outperformed the LR model across all selected indices. The evaluation metrics confirm the superior performance of the LSTM model. To the best of our knowledge, this study presents the first reported EWS designed for predicting price bubbles in this context using a deep learning approach. Keywords: Generalized Supremum Augmented Dickey-Fuller Test (GSADF), Early Warning System (EWS), Logistic Regression (LR), Long Short-Term Memory (LSTM) JEL Classification: G01, G32, C45, C53 Introduction The bursting of price bubbles can precipitate financial crises, rendering the study of bubbles and market collapses highly consequential for investment portfolio risk management (Kaliva & Koskinen, 2008). The early identification of such bubbles and the forecasting of their trajectories are therefore crucial for policymakers and market participants, as it enables preventive measures to mitigate or avert financial turmoil. Consequently, the development and deployment of financial EWS are essential to actively reduce economic vulnerabilities and counteract price bubbles (Claessens & Kose, 2013). As Phillips et al. (2015) note, an effective EWS must achieve a high degree of accurate detection to facilitate swift and effective policy implementation, while simultaneously maintaining a low false-positive rate to avoid unnecessary policy actions. Analyzing historical bubble episodes and the factors influencing their emergence is thus fundamental to informed decision-making and the control of market irregularities (Sadeghisharif et al., 2017). In this context, deep learning algorithms have significantly advanced machine-learning models by offering greater computational speed and predictive precision. The rapid evolution of this field has attracted considerable attention from economists addressing a range of problems, particularly in the domain of asset price forecasting (Khaliliaraghi et al., 2022). The primary objective of this study is to evaluate bubble periods in selected Tehran Stock Exchange (TSE) indices and, by incorporating price changes from a set of early warning indicators, to design a bubble prediction system using an LSTM model. This research further aims to compare the predictive accuracy and quality of the LSTM model against an LR benchmark. Accordingly, this research seeks to answer the following question: Does the employment of a deep learning approach improve the accuracy and quality of bubble forecasts across all selected indices in this study? Materials and Methods The target variables in this study were identified through a comparative analysis. Five TSE indices were selected: the Total Index, the 50 Most Active Companies Index, the Industry Index, the Financial Index, and the Basic Metals Index. Monthly data for these indices were collected for the period from 2002 to 2023. Bubble episodes were identified using the GSADF test, and the resulting bubble signals were used to construct a dummy bubble variable, which served as the dependent variable in the subsequent modeling. Based on a literature review, eighteen influential variables were drawn from macroeconomic indicators, market information, valuation multiples, and commodity prices to serve as independent warning indicators. The returns for these variables were calculated as logarithmic changes. Given the time-series nature of the data, these indicators were structured with a 5-period lag and normalized to serve as input features for the models. The contribution of these variables to predictive power, while indirect, is evidenced through the final model performance. Subsequently, bubble-forecasting models were developed and an EWS was designed using a Recurrent Neural Network with an LSTM architecture as the deep learning approach, alongside an LR model as a classical machine-learning benchmark. Finally, model performance was evaluated based on the model type and stock index, using forecasting accuracy, the Area Under the Receiver Operating Characteristic Curve (AUC-ROC), and the confusion matrix. This evaluation framework allowed for a comprehensive assessment and comparison of the models' predictive capabilities. Findings As illustrated in Table 1, the highest forecasting accuracy for the LR model was achieved by the Financial Index at 79%, while the lowest was observed for the Basic Metals Index at 65%. The AUC evaluation for the LR model indicates its strongest performance on the Financial Index (92%) and its weakest on the Basic Metals Index (79%). Furthermore, the confusion matrix assessment for the LR model reveals notably low true positive rates for the test data, ranging from 0% to 17%. According to the results in Table 2, the LSTM model attained its highest forecasting accuracy on the Total Index (81%) and its lowest on the Basic Metals Index (73%). A comparative analysis of predictive accuracy demonstrates that the LSTM model outperformed the LR benchmark across all indices. The AUC evaluation for the LSTM model also shows its best performance on the Financial Index (92%) and its lowest on the Basic Metals Index (81%). The confusion matrix for the LSTM model indicates substantially higher true positive rates, ranging from 38% to 71%, with the highest rate for the Total Index (71%) and the lowest for the 50 Most Active Companies Index (38%). Table (1) Evaluation of the accuracy of the LR model for training and testing datasets by index Testing Training Indices 0.69 0.83 Total 0.69 0.85 Industry 0.79 0.88 Financial 0.67 0.94 50 Most Active Companies 0.65 0.85 Basic Metals Table (2) Evaluation of the accuracy of the LSTM model for training and testing datasets by Index Testing Training Indices 0.81 0.85 Total 0.77 0.88 Industry 0.80 0.84 Financial 0.75 0.94 50 Most Active Companies 0.73 0.87 Basic Metals Discussion and conclusion The results demonstrate that the proposed EWS achieves predictive accuracy ranging from 73% to 81% across all selected indices, with the highest performance on the Total Index (81%) and the lowest on the Basic Metals Index (73%). A comparative analysis of model performance confirms that the LSTM model consistently outperforms the LR benchmark across all indices. This superiority is particularly evident in the case of the Total Index, where the LSTM model's accuracy of 81% represents a 12-percentage-point improvement over the LR model's 69%. Similarly, the accuracy improved by 8 percentage points for the Industrial, 50 Most Active Companies, and Basic Metals indices, while a marginal improvement of 1 percentage point was observed for the Financial Index. In terms of predictive quality, as evaluated by the AUC-ROC and confusion matrices, the findings present a nuanced picture. The AUC-ROC metrics indicate that the performance of the LSTM and LR models is somewhat comparable. In contrast, the evaluation based on confusion matrices reveals a substantially superior performance for the LSTM model, particularly in its ability to correctly identify true positives. The primary objective of this study is to design an Early Warning System (EWS) based on a Long Short-Term Memory (LSTM) architecture for the timely forecasting of price bubbles in the Tehran Stock Exchange (TSE). A secondary objective is to compare the predictive performance of this model against a Logistic Regression (LR) benchmark, using evaluation metrics such as the AUC-ROC and confusion matrix. The system's performance was evaluated on five selected TSE indices. Utilizing monthly data from 2002 to 2023, bubble periods were identified via the Generalized Supremum Augmented Dickey-Fuller (GSADF) test and represented as a binary variable. This bubble variable was then modeled using price changes of key warning indicators. The proposed deep learning-based system achieved predictive accuracy ranging from 73% to 81%, with the highest performance on the Total Index (81%) and the lowest on the Basic Metals Index (73%). A comparative analysis demonstrates that the LSTM model outperformed the LR model across all selected indices. The evaluation metrics confirm the superior performance of the LSTM model. To the best of our knowledge, this study presents the first reported EWS designed for predicting price bubbles in this context using a deep learning approach. Keywords: Generalized Supremum Augmented Dickey-Fuller Test (GSADF), Early Warning System (EWS), Logistic Regression (LR), Long Short-Term Memory (LSTM) JEL Classification: G01, G32, C45, C53 Introduction The bursting of price bubbles can precipitate financial crises, rendering the study of bubbles and market collapses highly consequential for investment portfolio risk management (Kaliva & Koskinen, 2008). The early identification of such bubbles and the forecasting of their trajectories are therefore crucial for policymakers and market participants, as it enables preventive measures to mitigate or avert financial turmoil. Consequently, the development and deployment of financial EWS are essential to actively reduce economic vulnerabilities and counteract price bubbles (Claessens & Kose, 2013). As Phillips et al. (2015) note, an effective EWS must achieve a high degree of accurate detection to facilitate swift and effective policy implementation, while simultaneously maintaining a low false-positive rate to avoid unnecessary policy actions. Analyzing historical bubble episodes and the factors influencing their emergence is thus fundamental to informed decision-making and the control of market irregularities (Sadeghisharif et al., 2017). In this context, deep learning algorithms have significantly advanced machine-learning models by offering greater computational speed and predictive precision. The rapid evolution of this field has attracted considerable attention from economists addressing a range of problems, particularly in the domain of asset price forecasting (Khaliliaraghi et al., 2022). The primary objective of this study is to evaluate bubble periods in selected Tehran Stock Exchange (TSE) indices and, by incorporating price changes from a set of early warning indicators, to design a bubble prediction system using an LSTM model. This research further aims to compare the predictive accuracy and quality of the LSTM model against an LR benchmark. Accordingly, this research seeks to answer the following question: Does the employment of a deep learning approach improve the accuracy and quality of bubble forecasts across all selected indices in this study? Materials and Methods The target variables in this study were identified through a comparative analysis. Five TSE indices were selected: the Total Index, the 50 Most Active Companies Index, the Industry Index, the Financial Index, and the Basic Metals Index. Monthly data for these indices were collected for the period from 2002 to 2023. Bubble episodes were identified using the GSADF test, and the resulting bubble signals were used to construct a dummy bubble variable, which served as the dependent variable in the subsequent modeling. Based on a literature review, eighteen influential variables were drawn from macroeconomic indicators, market information, valuation multiples, and commodity prices to serve as independent warning indicators. The returns for these variables were calculated as logarithmic changes. Given the time-series nature of the data, these indicators were structured with a 5-period lag and normalized to serve as input features for the models. The contribution of these variables to predictive power, while indirect, is evidenced through the final model performance. Subsequently, bubble-forecasting models were developed and an EWS was designed using a Recurrent Neural Network with an LSTM architecture as the deep learning approach, alongside an LR model as a classical machine-learning benchmark. Finally, model performance was evaluated based on the model type and stock index, using forecasting accuracy, the Area Under the Receiver Operating Characteristic Curve (AUC-ROC), and the confusion matrix. This evaluation framework allowed for a comprehensive assessment and comparison of the models' predictive capabilities. Findings As illustrated in Table 1, the highest forecasting accuracy for the LR model was achieved by the Financial Index at 79%, while the lowest was observed for the Basic Metals Index at 65%. The AUC evaluation for the LR model indicates its strongest performance on the Financial Index (92%) and its weakest on the Basic Metals Index (79%). Furthermore, the confusion matrix assessment for the LR model reveals notably low true positive rates for the test data, ranging from 0% to 17%. According to the results in Table 2, the LSTM model attained its highest forecasting accuracy on the Total Index (81%) and its lowest on the Basic Metals Index (73%). A comparative analysis of predictive accuracy demonstrates that the LSTM model outperformed the LR benchmark across all indices. The AUC evaluation for the LSTM model also shows its best performance on the Financial Index (92%) and its lowest on the Basic Metals Index (81%). The confusion matrix for the LSTM model indicates substantially higher true positive rates, ranging from 38% to 71%, with the highest rate for the Total Index (71%) and the lowest for the 50 Most Active Companies Index (38%). Table (1) Evaluation of the accuracy of the LR model for training and testing datasets by index Testing Training Indices 0.69 0.83 Total 0.69 0.85 Industry 0.79 0.88 Financial 0.67 0.94 50 Most Active Companies 0.65 0.85 Basic Metals Table (2) Evaluation of the accuracy of the LSTM model for training and testing datasets by Index Testing Training Indices 0.81 0.85 Total 0.77 0.88 Industry 0.80 0.84 Financial 0.75 0.94 50 Most Active Companies 0.73 0.87 Basic Metals Discussion and conclusion The results demonstrate that the proposed EWS achieves predictive accuracy ranging from 73% to 81% across all selected indices, with the highest performance on the Total Index (81%) and the lowest on the Basic Metals Index (73%). A comparative analysis of model performance confirms that the LSTM model consistently outperforms the LR benchmark across all indices. This superiority is particularly evident in the case of the Total Index, where the LSTM model's accuracy of 81% represents a 12-percentage-point improvement over the LR model's 69%. Similarly, the accuracy improved by 8 percentage points for the Industrial, 50 Most Active Companies, and Basic Metals indices, while a marginal improvement of 1 percentage point was observed for the Financial Index. In terms of predictive quality, as evaluated by the AUC-ROC and confusion matrices, the findings present a nuanced picture. The AUC-ROC metrics indicate that the performance of the LSTM and LR models is somewhat comparable. In contrast, the evaluation based on confusion matrices reveals a substantially superior performance for the LSTM model, particularly in its ability to correctly identify true positives.

https://amf.ui.ac.ir/article_29846_df89e86f7b1a65c9b6820cb92fb9d95a.pdf