The paper introduces MLF (Multi-period Learning Framework), a novel deep learning framework for financial time series forecasting that simultaneously processes multiple input windows of different lengths (short, medium, long). Unlike existing single-period models, MLF captures both short-term fluctuations and long-term trends by integrating multi-period information. The framework addresses two key challenges: integration of multi-period inputs and computational efficiency. To improve integration, MLF proposes three modules: Multi-period self-Adaptive Patching (MAP) ensures equal patch counts across periods to avoid bias; Inter-period Redundancy Filtering (IRF) removes redundant information between periods to enhance self-attention; and Learnable Weighted-average Integration (LWI) adaptively combines forecasts from different periods. For efficiency, a Patch Squeeze module reduces intra-period redundancy by compressing patches via a lightweight encoder-decoder with reconstruction loss. Experiments on a proprietary fund sales dataset from Alipay and several public benchmarks demonstrate that MLF achieves superior accuracy and efficiency compared to state-of-the-art methods. The framework has been deployed in Alipay's fund inventory management system, leading to improved forecast accuracy and operational efficiency.