Forecasting cryptocurrency price and trend using high frequency data remains a challenging task due to the inherent noise and complexity of market microstructures. While deep learning models have achieved notable success, those based primarily on standard open, high, low, close, volume (OHLCV) or limit order book (LOB) data often suffer from overfitting and limited generalization to out-of-sample scenarios. This paper explores whether incorporating cyclical temporal features—such as minute-of-hour, hour-of-day, and day-of-week—can improve prediction performance. We apply this approach to high-frequency Ethereum data and evaluate its effectiveness using three deep learning architectures: LSTM, CNN, and a hybrid model. Across both minute-ahead price prediction and short-term trend prediction tasks, our results show that integrating these periodic time features leads to consistent and significant improvements in forecasting accuracy. The source code is available and maintained in the GitHub repository ( https://github.com/LHSang6403/HFT-Time-Cyclic-Improvement ).