محمدرضا داودابادی

In the current era, financial markets play a crucial role in the global economy and are recognized as its beating heart. Financial markets are critically important in several aspects, including optimal resource allocation, facilitating investment flows in industries, determining the true value of assets, contributing to economic development, and risk management, acting as the driving engine of the economy. Additionally, the existence of efficient financial markets reflects the dynamism and health of a society's economy, and changes in these markets can significantly impact the lives of many individuals. For this reason, predicting indicators as markers of financial market conditions is of great importance and can aid in informed decision-making, proper resource allocation, risk reduction for investors, and achieving optimal returns. The goal of forecasting financial markets is to predict prices, returns, and market trends so that this information can be used for financial decision-making and investment.

To date, various approaches, such as statistical and economic methods, have been introduced and developed for forecasting financial markets, each with its own strengths and weaknesses. However, with the emergence of machine learning and deep learning algorithms, increased computational power, and access to large volumes of data, time series forecasting methods for financial markets using new machine learning and deep learning-based approaches have gained attention in the scientific community due to their higher accuracy and speed in developing predictive models, leading to improved forecasting accuracy.

Among these methods, transformer-based models have gained attention in recent years due to their ability to learn long-term dependencies and enhance model interpretability through attention mechanisms, particularly in fields such as natural language processing, computer vision, and time series analysis.

In this research, the Temporal Fusion Transformer (TFT) model, a transformer-based model, has been used to forecast the S&am   ۵۰۰ index and the Tehran Stock Exchange's overall index to evaluate its effectiveness in predicting these indices. Additionally, to assess the performance of the TFT model, its results have been compared with well-known models such as LSTM, simple moving average, and support vector regression. The LSTM and TFT models had lower errors compared to the simple moving average and support vector regression. The error of the TFT model in predicting the overall index of the Iranian stock market was lower than that of LSTM, but in predicting the S&am   ۵۰۰ index, the LSTM model achieved lower error in both evaluation metrics.