In the realm of financial transactions, the allocation of household assets often lacks proper guidance, resulting in suboptimal utilization and limited income for residents. This study aims to address this issue by introducing rational asset allocation strategies, improv-ing efficiency, and increasing household income. Specifically, the study focuses on en-hancing Deep Reinforcement Learning (DRL) algorithms, particularly the Deep Q-Net-work algorithm. The current network is optimized by employing a dual network structure, leading to improved performance. Moreover, the proposed model incorporates Grubbs’ improved algorithm to effectively denoise stock data samples, enabling iterative training of diverse stock agents. The results demonstrate the following findings: (1) The application of the Grubbs’ improved algorithm to the cumulative returns of different stocks significantly surpasses the results without noise reduction; (2) The Sharpe ratio of the stock agents using the Grubbs improvement algorithm is noticeably higher than that of the un-improved DRL algorithm. The Sharpe ratio of stock agents under the unimproved algo-rithm consistently remains below 1.0; (3) The maximum drawdown of the stock agents using the Grubbs’ improved algorithm is significantly lower than that of the unimproved DRL algorithm. The unimproved algorithm exhibits relatively higher maximum draw-down rates. These findings indicate that Grubbs’ improved algorithm possesses notable advantages in noise reduction and enhancing the cumulative returns of assets.