In a stock portfolio trading system, there are two main tasks: stock selection and portfolio optimization. To consider market dynamics, such a system needs to have appropriate buy/sell trading signals that can be generated by applying a Convolutional Neural Network (CNN) as a powerful classifier. In a CNN architecture, pooling layers operate to decrease the input features' dimensionality while overcoming the overfitting issue. Although several approaches have been proposed for designing this layer based on different applications, especially in image processing, so far, no attempts have been made to design a pooling layer according to the characteristics of an algorithmic trading system. In this paper, an enhanced CNN has been proposed for stock selection in which a new pooling layer called ranked-based time-adjusted weighted pooling layer (RTAWP) has been suggested, in which activation values are ranked according to the recency of price time series. Our simplified RTAWP dramatically reduces the number of parameters while maintaining competing performance. To deal with the tail risk of the constructed portfolios, we consider Conditional Value at Risk (CVaR) as a coherent risk measure in our portfolio optimization model. To show the applicability of our proposed model, we use volume and open-high-low-close prices of sample stocks from the Tehran Stock Exchange during Jan 2020 and Sep 2023. The results indicate that the proposed model provides better computational and financial results for our sample stocks compared with average pooling and other benchmark models.