时间序列通常包含这些组成部分:线性趋势(Trend),季节变化(Seasonality),循环变化(Cycle),不规则变化(Irregularity)
多步预测问题
图片来自:《Temporal Fusion Transformers for Interpretable Multi-horizon Time Series Forecasting》
时间序列预测的基本思路
- 自回归模型
- 循环模型,隐状态模型
- 曲线拟合模型
图片来自《Neural Decomposition of Time-Series Data for Effective Generalization》
从深度学习的角度:CNN, RNN, Attention
图片来自:《Time Series Forecasting With Deep Learning: A Survey》
多步预测的五种策略
可分为单步预测(one-step-ahead)和多步预测(muti-step-ahead)
多步预测的五种策略《Comparison of Strategies for Multi-step-ahead Prediction of Time Series using Neural Network》:
- recursive (or iterated) strategy
- direct strategy
- combination of both the recursive and direct strategies, called DirREC
- the Multi-Input Multi-Output (MIMO) strategy
- DirMO strategy
常用指标
y ^ \hat{y} y^ 为预测值, y y y 为实际值, N N N 为预测数:
-
MSE
M S E = 1 N ∑ i = 1 N ( y ^ i − y i ) 2 MSE = \frac{1}{N}\sum_{i=1}^{N}(\hat{y}_i-y_i)^2 MSE=N1i=1∑N(y^i−yi)2 -
RMSE
R M S E = 1 N ∑ i = 1 N ( y ^ i − y i ) 2 RMSE = \sqrt{\frac{1}{N}\sum_{i=1}^{N}(\hat{y}_i-y_i)^2} RMSE=N1i=1∑N(y^i−yi)2 -
NMSE
N M S E = 1 N ⋅ σ y 2 ∑ i = 1 N ( y ^ i − y i ) 2 NMSE = \frac{1}{N\cdot \sigma_y^2}\sum_{i=1}^{N}(\hat{y}_i-y_i)^2 NMSE=N⋅σy21i=1∑N(y^i−yi)2 -
MAE
M A E = 1 N ∑ i = 1 N ∣ y ^ i − y i ∣ MAE = \frac{1}{N}\sum_{i=1}^{N}|\hat{y}_i-y_i| MAE=N1i=1∑N∣y^i−yi∣ -
MAPE
M A P E = 1 N ∑ i = 1 N ∣ y ^ i − y i y i ∣ MAPE = \frac{1}{N}\sum_{i=1}^{N}\left|\frac{\hat{y}_i-y_i}{y_i}\right| MAPE=N1i=1∑N∣∣∣∣yiy^i−yi∣∣∣∣ -
sMAPE
symmetric mean absolute percentage error
s M A P E = 1 N ∑ i = 1 N ∣ y ^ i − y i ( y ^ i + y i ) / 2 ∣ sMAPE = \frac{1}{N}\sum_{i=1}^{N}\left|\frac{\hat{y}_i-y_i}{(\hat{y}_i+y_i)/2}\right| sMAPE=N1i=1∑N∣∣∣∣(y^i+yi)/2y^i−yi∣∣∣∣ -
MASE
mean absolute scaled error
M A S E = 1 h ∑ i = 1 h ∣ y ^ i − y i ∣ 1 l − p ∑ i = p + 1 − l 0 ∣ y ^ i − y i − p ∣ MASE = \frac{\frac{1}{h}\sum_{i=1}^h|\hat{y}_i - y_i|}{\frac{1}{l-p}\sum_{i=p+1-l}^0|\hat{y}_i - y_{i-p}|} MASE=l−p1∑i=p+1−l0∣y^i−yi−p∣h1∑i=1h∣y^i−yi∣其中 l l l 是训练集样本数, h h h 是预测长度, p p p 是季节长度