Abstract: Disk failures, the most common and major failures in storage systems, increase the risk of service interruption and data loss, and bring additional maintenance costs, which reduces system reliability. Disk failure prediction methods aim to forecast failures, initiating prompt data migration and disk replacement. Existing methods continuously optimize the models with different sampling methods and modeling algorithms. However, due to issues such as inaccurate sample labeling, insufficient data sampling, and improper sample segmentation, the predictive capabilities of existing models within the lookahead-window time are unstable and decline as the lookahead-window time increases. To address this, we propose LWCM (Lookahead-Window Constrained Model) to improve the predictability and stability of failure prediction models within the lookahead-window time. LWCM leverages dynamic sample relabeling methods based on lookahead-window time constraints and failure symptom durations to modify inaccurate sample labels. LWCM utilizes effective sample data by using the two-phase data sampling method including initial expectation sampling and subsequent segmented resampling. LWCM employs dynamic weighted optimization in backpropagation to enhance the predictability and stability of the disk failure prediction model. Experimental results show that LWCM has better failure prediction performance. The true positive and false positive rates surpass those of the offline-RF model by 38.7% and 92.4%, respectively. Furthermore, LWCM demonstrates its applicability across disk models while maintaining stability within the lookahead constraint window.