基于Leddam-iTransformer-FCB的日频股价预测方法
Daily Frequency Stock Price Prediction Method Based on Leddam-iTransformer-FCB

日频股价预测同时受到市场噪声、弱非平稳性以及价格变量与流动性相关变量耦合关系的共同影响，因而仍是量化金融中的一项具有挑战性的研究问题。本文围绕归档的Leddam-iTransformer-FCB实现，考察其在一步日频股价预测场景下的模型机理与实验表现。该模型首先对输入样本执行样本级归一化，随后利用可学习分解模块提取各变量中的平滑主成分，再通过倒置Transformer将变量轨迹视为token以建模跨变量依赖关系，并在潜在空间中引入Fourier Convolution Block进行全局频谱混合，最终完成预测输出。依据归档代码、checkpoint与测试预测序列，可以重建出该模型工作于一个包含5个变量、共2085个日频观测、采用60%/20%/20%划分、回溯窗口长度为12个交易日、预测步长为1日的股价预测设置。实验结果表明，完整模型在测试集上的性能为R² = 0.9653、MSE = 1.5547、RMSE = 1.2469、MAE = 0.9294、MAPE = 1.2655%。结果表明，可学习平滑、跨变量注意力与潜在频谱混合在短期股价预测任务中具有明显的互补作用。
Daily stock price forecasting remains a challenging research topic in quantitative finance due to the combined effects of market noise, weak non-stationarity, and coupling relationships between price variables and liquidity-related variables. This study investigates the model mechanism and experimental performance of the archived Leddam-iTransformer-FCB framework in one-step daily price prediction scenarios. The model first performs sample-level normalization on input data, then employs a learnable decomposition module to extract smooth principal components from variables. An inverted Transformer treats variable trajectories as tokens to model cross-variable dependencies, while Fourier Convolution Blocks in latent spaces facilitate global spectral mixing for prediction output. Using archived code, checkpoints, and test sequences, the model operates under the following parameters: 5 variables, 2085 daily observations, 60%/20%/20% data split, 12-day backtracking window, and 1-day prediction horizon. Experimental results demonstrate the model achieves R² = 0.9653, MSE = 1.5547, RMSE = 1.2469, MAE = 0.9294, and MAPE = 1.2655% on the test set. The results indicate that learnable smoothing, cross-variable attention, and latent spectrum mixing exhibit significant complementary effects in short-term stock price prediction tasks.