ارائه يك معماري تركيبي پيش‌بيني قيمت بيت‌كوين بر اساس اخبار مرتبط و داده‌هاي قيمت

A Hybrid Architecture for Bitcoin Price Prediction Based on Related News an‎d Price Data

مهندسي نرم افزار

The cryptocurrency market, particularly Bitcoin, is considered one of the most complex financial markets due to its high volatility an‎d susceptibility to economic, political, an‎d social factors. In recent years, numerous studies have investigated the application of statistical models, machine learning, an‎d deep learning techniques for Bitcoin price prediction. Some of these studies have also incorporated news data an‎d sentiment analysis alongside price data. However, a considerable proportion of existing research has faced limitations, including insufficient modeling of short-term an‎d long-term temporal dependencies simultaneously, static an‎d non-adaptive integration of textual news an‎d numerical data, an‎d limited eva‎luation of model performance across different forecasting horizons. To address these limitations, this study proposes an‎d implements an integrated framework for Bitcoin price prediction. The proposed framework jointly utilizes numerical market data, including prices an‎d technical indicators, an‎d Bitcoin-related news data. Following preprocessing, text summarization, an‎d sentiment analysis, the news data are transformed into numerical feature vectors, enabling the quantitative incorporation of news content an‎d market sentiment into the prediction process. The proposed architecture combines a Long Short-Term Memory (LSTM) network with a Temporal Fusion Transformer (TFT) branch to capture short-term an‎d long-term temporal dependencies in price data simultaneously. Subsequently, features extracted from the numerical an‎d news branches are dynamically integrated through a self-attention-based fusion layer, allowing the relative contributions of news an‎d price information to be adaptively determined at each time step. To improve learning stability an‎d mitigate overfitting, advanced training an‎d validation strategies are employed. The proposed model is trained an‎d eva‎luated across three forecasting horizons: short-term, medium-term, an‎d long-term. Experimental results demonstrate that the model achieves a coefficient of determination ((R^2)) of 0.9907, a mean absolute error (MAE) of 0.007942, an‎d a root mean squared error (RMSE) of 0.010673 for the one-hour forecasting horizon. As the forecasting horizon increases to four an‎d twelve hours, the (R^2) values decrease to 0.9856 an‎d 0.9750, respectively, while the RMSE values increase to 0.013307 an‎d 0.017521. These results indicate that the model maintains robust predictive performance across different forecasting horizons an‎d highlight the potential contribution of adaptive news an‎d price information fusion to improving prediction accuracy

6

171906