在 Python 中使用季节性分解时我做错了什么？

Question

我有一个每月间隔的小时间序列。我想绘制它，然后分解为季节性、趋势、残差。我首先将 csv 导入 pandas，然后只绘制工作正常的时间序列。我遵循This 教程，我的代码是这样的：

%matplotlib inline
import matplotlib.pyplot as plt
import matplotlib.dates as mdates
import pandas as pd

ali3 = pd.read_csv('C:\\Users\\ALI\\Desktop\\CSV\\index\\ZIAM\\ME\\ME_DATA_7_MONTH_AVG_PROFIT\\data.csv',
 names=['Date', 'Month','AverageProfit'],
 index_col=['Date'],
 parse_dates=True)

\* Delete month column which is a string */
del ali3['Month']


ali3
plt.plot(ali3)

Data Frame

在这个阶段，我尝试像这样进行季节性分解：

import statsmodels.api as sm 
res = sm.tsa.seasonal_decompose(ali3.AverageProfit)  
fig = res.plot() 

导致以下错误：

---------------------------------------------------------------------------
ValueError                                Traceback (most recent call last)
<ipython-input-41-afeab639d13b> in <module>()
      1 import statsmodels.api as sm
----> 2 res = sm.tsa.seasonal_decompose(ali3.AverageProfit)
      3 fig = res.plot()

C:\Users\D063375\AppData\Local\Continuum\Anaconda2\lib\site-packages\statsmodels\tsa\seasonal.py in seasonal_decompose(x, model, filt, freq)
     86             filt = np.repeat(1./freq, freq)
     87 
---> 88     trend = convolution_filter(x, filt)
     89 
     90     # nan pad for conformability - convolve doesn't do it

C:\Users\D063375\AppData\Local\Continuum\Anaconda2\lib\site-packages\statsmodels\tsa\filters\filtertools.py in convolution_filter(x, filt, nsides)
    287 
    288     if filt.ndim == 1 or min(filt.shape) == 1:
--> 289         result = signal.convolve(x, filt, mode='valid')
    290     elif filt.ndim == 2:
    291         nlags = filt.shape[0]

C:\Users\D063375\AppData\Local\Continuum\Anaconda2\lib\site-packages\scipy\signal\signaltools.py in convolve(in1, in2, mode)
    468         return correlate(volume, kernel[slice_obj].conj(), mode)
    469     else:
--> 470         return correlate(volume, kernel[slice_obj], mode)
    471 
    472 

C:\Users\D063375\AppData\Local\Continuum\Anaconda2\lib\site-packages\scipy\signal\signaltools.py in correlate(in1, in2, mode)
    158 
    159     if mode == 'valid':
--> 160         _check_valid_mode_shapes(in1.shape, in2.shape)
    161         # numpy is significantly faster for 1d
    162         if in1.ndim == 1 and in2.ndim == 1:

C:\Users\D063375\AppData\Local\Continuum\Anaconda2\lib\site-packages\scipy\signal\signaltools.py in _check_valid_mode_shapes(shape1, shape2)
     70         if not d1 >= d2:
     71             raise ValueError(
---> 72                 "in1 should have at least as many items as in2 in "
     73                 "every dimension for 'valid' mode.")
     74 

ValueError: in1 should have at least as many items as in2 in every dimension for 'valid' mode.

谁能阐明我做错了什么以及如何解决？非常感谢。

编辑：这就是数据框的样子

Date            AverageProfit

2015-06-01          29.990231
2015-07-01          26.080038
2015-08-01          25.640862
2015-09-01          25.346447
2015-10-01          27.386001
2015-11-01          26.357709
2015-12-01          25.260644

Answer 1

您有 7 个数据点，这通常是执行平稳性分析的一个非常小的数字。

您没有足够的积分来使用季节性分解。要查看这一点，您可以连接您的数据以创建一个扩展的时间序列（只需在接下来的几个月重复您的数据）。让extendedData 成为这个扩展数据框，data 成为您的原始数据。

data.plot()

extendedData.plot()

res = sm.tsa.seasonal_decompose(extendedData.interpolate())
res.plot()

季节性估计的频率 (freq) 是根据数据自动估计的，并且可以手动指定。

---

您可以尝试第一个差异：生成一个新的时间序列，从前一个数据值中减去每个数据值。在您的情况下，它看起来像这样：

接下来可以应用平稳性测试，如 here 所述