| چکیده انگلیسی مقاله |
Purpose: In recent years, most of the currency fluctuations in Iran have occurred after the spread of influential domestic and foreign political-economic news. Despite this significant effect, the impact of the news content is not considered, and its results are used in the form of numerical data in modeling. However, numerical indirect results are available after the impact of the news on the market and are practically useless in forecasting. Therefore, there needs to be a model that includes the news directly in forecasting the exchange rate. Another important phenomenon about the exchange rates and news is the framing effect. This effect is a linguistic-cognitive phenomenon resulting in people with two different attitudes, positive and negative, making different choices regarding the same item or reality. While this action is done explicitly or implicitly in policymakers' announcements, criticisms, and statements, there needs to be a systematic way to analyze and predict its effect on the exchange rate. This paper presents a comprehensive model based on machine learning to predict exchange rate moves by receiving all kinds of textual, numerical, and categorical data. Also, using this model, it is possible to analyze the effect of framing on exchange rate changes systematically.
Methodology: A processing method has been designed to predict the exchange rate from receiving data to producing the final model. In this method, various types of input data, including time series data, batch data, and political-economic news, are received and preprocessed from essential and relevant sources. In numerical data preprocessing, noises and outliers are removed, and values are normalized. Text data are also the texts normalized and converted into tokens. In the next step, custom embedding is provided based on Pars-BERT embedding and the collected news data. Using this embedding, the textual data are converted into numerical vectors, and all the data are given to the model. Here are three supervised models. The first and third models use textual and numerical data to predict the exchange rate, and the second model only uses numerical data. The core of all the models is a Bi-GRU, a deep neural network. Dropout and batch normalization and regularization techniques have been used in these models to avoid overfit and bias.
Findings and discussion: The tests performed to evaluate the models' efficiency are divided into two parts. The first part evaluates all the three developed models and their results. The second part is dedicated to assessing and analyzing the issue of news framing and the model's performance in this field. Based on the accuracy metric, the first to third models perform properly at the rate of 96.5, 87.2 and 97.07, respectively. The difference between the results of the second model and the first and third ones shows the effect of news on increasing the model's accuracy. In the second part, a dataset of the news affecting the exchange rate with different framings is prepared to evaluate the effect of framing. This set includes 74 main pieces of news and 435 secondary ones. In approximately 32% of the cases, where the news was announced to reduce the rate, the behavior of the market matched the purpose of the report. Meanwhile, in 68% of the other cases, the news that was expected to have a downward effect on the exchange rate did not match the market behavior. This can indicate their incorrect framing. On the other hand, the news that is expected to have a negative (increasing) effect on the exchange rate has worked about 78% of the time and has led to an increase in the rate. This observation also confirms the imbalance effect of positive and negative news on the currency market. Interestingly, the first and third models correctly identified 98.8% and 99.05% of the samples, including negative and positive news of internal and external origins, respectively. This indicates that the models resist different framings of a news story and have made correct predictions.
Conclusions and policy implications: In this article, using machine learning models, an operational method was presented to investigate the effect of framing on the exchange rate in Iran. The proposed architecture includes the five stages of preprocessing textual, numerical and categorical data, data vector representation, feature engineering, model training, and final evaluation and analysis. In the stage of textual data representation, customized embedding is used for economic data. In this regard, there are three models presented, including a model based on numerical data and two models based on numerical and textual data with a data fusion structure. All the three models are recurrent neural networks trained with deep learning techniques. The results of the experiments show that the models with combined data perform very well in forecasting the exchange rate. In addition to predicting the exchange rate, the produced models have been used to analyze the effect of the framing of the published news on the currency market. The research results show that negative news still has a far more significant impact on the currency market, and no positive framing can reduce its impact. Also, the mainly domestic report published to reduce the exchange rate has been successful in about 32% of the cases. Another significant achievement is the excellent performance of the presented models in framed samples. In other words, these models have correctly predicted market behavior by considering different aspects of a news story, news history, market history, and information, regardless of the type of the news framing. |