Evaluating GRU Algorithm and Double Moving Average for Predicting USDT Prices: A Case Study 2017-2024

Abstract

The cryptocurrency market is highly volatile, requiring advanced analytical methods for accurate price forecasting. This study evaluates the effectiveness of Gated Recurrent Units (GRU) and Double Moving Average (DMA) in predicting USDT (Tether Coin) prices using historical data from 2017 to 2024, sourced from Investing.com. Implemented in Jupyter Notebook, the research explores the strengths of each method in analyzing market fluctuations and price trends. GRU, a deep learning-based recurrent neural network, processes sequential data using a gating mechanism, making it effective for capturing short-term price dynamics. DMA, in contrast, is a statistical method that filters market noise to identify long-term trends, making it more reliable for stable market conditions. Performance evaluation shows DMA achieving lower errors (MAE: 5.494, MAPE: 0.0339%) than GRU (MAE: 5.984, MAPE: 0.0369%), suggesting higher accuracy for trend-based predictions. However, GRU’s lower RMSE (8.531 vs. 8.715 for DMA) indicates better adaptability to sudden price fluctuations, making it more responsive to volatile markets. A hybrid approach combining GRU and DMA reveals their complementary strengths—DMA’s minimal bias (-0.0013% MPE) supports stable trend analysis, while GRU’s slight positive bias (0.0286% MPE) captures short-term fluctuations. Additionally, a comparison with Long Short-Term Memory (LSTM) demonstrates its superior predictive accuracy, outperforming both GRU (MAE: 5.98, RMSE: 8.53) and DMA (MAE: 5.49, RMSE: 8.72) with the lowest MAE (4.31), MAPE (0.027%), and RMSE (5.64), alongside minimal bias (MPE: 0.007%). This study highlights the need for integrating multiple forecasting techniques in cryptocurrency price prediction. While DMA is well-suited for stable trends and GRU excels in volatile conditions, LSTM outperforms both, reinforcing the effectiveness of deep learning for financial time-series forecasting.