本研究旨在探討深度學習模型應用於高頻股價預測任務之可行性，並評估 Transformer 與其衍生架構 Informer 在台灣股市資料下的表現差異。考量真實金融環境中資料分散且不均，本研究進一步設計模擬聯邦式學習架構，透過多檔股票模型參數的加權聚合與微調，以提升模型在資料稀疏條件下的穩定性與泛化能力。此外，針對 Informer 模型中的 ProbSparse 注意力機制，本研究亦探討稀疏性參數（sparsity factor）之調整對預測精度的影響。 實驗資料涵蓋台灣 11 檔大型上市公司之 15 分鐘 K 線資料，並搭配多項技術指標進行特徵建構。實驗結果顯示，在原始設定下 Informer 與 Transformer 各有優劣，惟引入聯邦式學習策略後，Informer 在多數股票上表現穩定提升。進一步提高注意力稀疏性亦有助於捕捉高頻波動訊號，增強模型預測能力。整體而言，本研究提出一套具彈性且可重現的高頻股價預測流程，未來可延伸應用於交易策略設計與跨市場金融預測任務。;This study explores the application of deep learning models for high-frequency stock price forecasting, with a focus on evaluating the performance of Transformer and its derivative, Informer, in the Taiwan stock market. To address challenges related to data sparsity and heterogeneity in real-world financial environments, a simulated federated learning framework is introduced. This framework aggregates and fine-tunes model parameters trained on multiple stocks, thereby enhancing prediction stability and generalization under limited data conditions. In addition, the study examines the effect of adjusting the sparsity factor in the Informer’s ProbSparse attention mechanism on predictive accuracy. The experimental dataset consists of 15-minute K-line data from 11 major Taiwanese stocks, combined with a range of technical indicators for feature construction. Results show that while Transformer and Informer exhibit mixed performance under default settings, the federated aggregation strategy consistently improves the accuracy and robustness of Informer. Moreover, increasing the sparsity factor enhances the model’s ability to capture short-term fluctuations, particularly in high-frequency trading environments. Overall, this study proposes a flexible and reproducible forecasting pipeline, offering potential for extension to trading strategy development and cross-market financial applications.
