This study utilizes stock price data from the Taiwan Economic Journal (TEJ) and transforms it into point matrices as input for a Convolutional Neural Network (CNN) model to predict stock price trends for the next 5, 20, and 60 days. The sample period covers from January 1, 1997, to December 31, 2023, spanning a total of 27 years. The CNN model outputs the probability of price increase, which is used to classify stocks into deciles. Four trading strategies are developed based on these probabilities to evaluate the model’s prediction accuracy and the performance of the trading strategies. The empirical results indicate that, among the trading strategies, the performance of predicting 60-day future returns surpasses that of predicting 20-day returns, which in turn outperforms the 5-day predictions. Additionally, long-short strategies generated higher average returns over the holding period compared to buy-only strategies. Among the four trading strategies, the equal-weighted long-short portfolio performed the best.