正在打开模型库
正在打开模型库
把时间序列差分到平稳，再用自回归和滑动平均捕捉残留结构。
先让序列“站稳”，再认为今天的值可由昨天的值和过去的误差线性组合出来。
有连续时间点、要预测未来几期，并且你能写出平稳性和残差检验时用。
Python 代码
import numpy as np
series = np.array([21, 23, 24, 28, 30, 31, 34, 36], dtype=float)
diff = np.diff(series)
# 极简 AR(1)：diff_t = c + phi * diff_{t-1}
x, y = diff[:-1], diff[1:]
phi = np.cov(x, y, bias=True)[0, 1] / np.var(x)
c = y.mean() - phi * x.mean()
next_diff = c + phi * diff[-1]
print("下一期预测:", series[-1] + next_diff)