Deep Learning to Enhance the Accuracy and Lag-time of Time Series Data Forecasting

Abstract

Time series forecasting is crucial in the present, applied in areas such as weather forecasting, finance, and beyond. Gold price forecasting is highly complex due to significant influences from economic factors. Although current studies have explored gold price forecasting using deep learning models, issues of prediction lag persist, affecting the accuracy of directional forecasts. To address this problem, this research focuses on developing a gold price time series forecasting model with improved accuracy and reduced prediction lag. This is achieved by combining the strengths of the LSTM (Long Short-Term Memory) model, which excels at learning long-term dependencies in data, and the CNN (Convolutional Neural Network) model, which efficiently extracts spatial features from data. Additionally, a data grouping method based on price trend characteristics is employed, allowing the model to learn specific patterns of each trend type appropriately. This reduces lag caused by delayed responses to actual data changes and enhances the accuracy of gold price forecasts in both numerical value and directional movement. Consequently, the model demonstrates strong potential for effective application in real-world financial trading decisions. This study uses gold price data from January 2014 to July 2025, totaling 2,943 days. Price trends are classified into three groups uptrend, sideways and downtrend. Four models are trained, which are LSTM, CNN, LSTM_G3, and CNN_G3. The most suitable model for each group is then selected to create the LSTM+CNN_G3 model. The results show that the LSTM+CNN_G3 model achieves the lowest overall MAE (Mean Absolute Error) at 17.68 US dollars and the highest overall MPM (Movement Prediction Metric) at 54.08%, outperforming the baseline LSTM and CNN models that do not use data grouping. This demonstrates that the grouping method effectively reduces prediction lag. In a gold trading simulation, the LSTM+CNN_G3 model yields the highest profit of 1,870 US dollars, representing a 101.87% return on capital. These findings indicate that grouping data by price trends combined with selecting appropriate architectures can improve accuracy and reduce lag, enabling the model to be effectively applied in real-world financial markets.