放弃学习周期形状，用低秩分解实现高效精准的周期时间序列预测。
Don't Learn the Shape: Forecasting Periodic Time Series by Rank-1 Decomposition
- 将周期序列视为低秩矩阵，仅通过奇异值分解提取主成分，避免复杂建模。
- 在97个测试场景中，其相对误差（relMASE）达0.838，优于多数复杂模型。
- 适合资源受限场景，无需调参、无预训练，推理快且参数极简。
预测周期性时间序列需要多少参数？将小时级电力数据重构成24行（每日一列）的矩阵后，其秩接近1：每日形态由每日水平调制（GIFT-Eval上中位秩-1能量为0.82）。是否应学习该形态？我们测试了8种替代方案（如傅里叶、EWMA、James-Stein、低秩SVD），在97个GIFT-Eval配置下，经霍姆校正，均未显著优于固定基线；其中两个更差。由此提出FLAIR方法：(a) 有效：在整体GIFT-Eval上表现媲美PatchTST（relMASE 0.838 vs 0.849）；(b) 紧凑：小时级仅需28个标量，周级57个；(c) 快速：单核MacBook Pro仅需22分钟；(d) 闭式解、免调参：每周期候选值执行一次SVD，GCV平均岭回归，无需GPU、预训练或任务微调。在高秩-1、多周期场景中，额外灵活性反而带来估计噪声。
原文摘要 · Abstract (English)
How few parameters do we really need to forecast a periodic time series? An hourly electricity series, reshaped as a 24-row matrix with one column per day, is approximately rank-1: a daily shape modulated by a daily level (median centered rank-1 energy 0.82 on GIFT-Eval). Should we learn the shape? Smoothing, shrinkage, and low-rank fits all seem like obvious upgrades over the simple average of the last K=2 cycles. On all 97 GIFT-Eval configurations, we tested 8 such alternatives (e.g., Fourier, EWMA, James-Stein, rank-r SVD): none significantly beats the frozen baseline under Holm correction; two are significantly worse. The resulting method, FLAIR, is (a) Effective: matches PatchTST on aggregate GIFT-Eval (relMASE 0.838 vs 0.849); (b) Compact: 28 scalars for hourly, 57 for weekly; (c) Fast: 22 minutes on one CPU core of a MacBook Pro; (d) Closed-form & Hands-Off: one SVD per period candidate, GCV-averaged Ridge, no GPU, no pre-training, no per-task tuning. In the high-rank-1, many-cycle regime, extra flexibility is estimation noise.
Thank you to arXiv for use of its open access interoperability. PaperDance 不是 arXiv 官方产品；中文卡片由大模型生成，请以原文为准。