Designing neural networks for stock market trading is a challenging task, as it’s not as straightforward as solving a classification or regression problem. The stochastic behavior in the stock market makes it extremely difficult to predict price movements accurately. To tackle this, a novel framework is proposed in this paper to effectively train the model by designing a customized objective function that seeks to maximize cumulative rewards during the training process. As a result, the model becomes adept at identifying patterns that lead to positive price fluctuations while simultaneously avoiding market risks. To align with the objective function, the reward for each input is linked to the future price change. An adaptive preprocessing method is used to center the rewards for each batch at zero and eliminate noise, preventing the model from converging to solutions that trigger buying signals for all inputs due to the inherent biases in the rewards. Within this framework, continuous and discrete features are extracted from normalized price and volume inputs and combined as the input of the convolutional neural network to compute the probability of initiating buying actions. Experiments conducted on various datasets show that agents trained using this framework can generate excess returns across different markets.