我有一个多元时间序列，因此我正在尝试进行向量自回归移动平均（VARMA）预测。我相信VARMA最适合该系列，因为该系列确实具有多个变量，所有变量都是内生的。
根据多个来源（包括statsmodel文档），VARMAX类可用于完成VARMA计算。实际上，我可以使用VARMA成功拟合模型，并提供以下摘要。
from statsmodels.tsa.statespace.varmax import VARMAX
# fit model
varma = VARMAX(df_pca, order=(1, 1))
varma_fit = varma.fit(maxiter=1000, disp=False)
print (varma_fit.summary())
Statespace Model Results
==============================================================================
Dep. Variable: [0, 1, 2] No. Observations: 13199
Model: VARMA(1,1) Log Likelihood 59909.901
+ intercept AIC -119765.801
Date: Fri, 02 Nov 2018 BIC -119563.628
Time: 23:50:27 HQIC -119698.300
Sample: 01-02-1962
- 10-29-2018
Covariance Type: opg
=====================================================================================================
Ljung-Box (Q): 163.74, 127.73, 137.46 Jarque-Bera (JB): 149761.37, 41120716.89, 29871.12
Prob(Q): 0.00, 0.00, 0.00 Prob(JB): 0.00, 0.00, 0.00
Heteroskedasticity (H): 1.41, 2.31, 0.80 Skew: -0.12, 6.25, 0.13
Prob(H) (two-sided): 0.00, 0.00, 0.00 Kurtosis: 19.50, 276.16, 10.37
Results for equation 0
==============================================================================
coef std err z P>|z| [0.025 0.975]
------------------------------------------------------------------------------
const -0.0005 0.002 -0.251 0.802 -0.005 0.004
L1.0 0.9998 0.000 4756.659 0.000 0.999 1.000
L1.1 0.0022 0.002 1.105 0.269 -0.002 0.006
L1.2 0.0081 0.008 1.039 0.299 -0.007 0.023
L1.e(0) 0.1021 0.005 21.093 0.000 0.093 0.112
L1.e(1) 0.1295 0.016 8.233 0.000 0.099 0.160
L1.e(2) -0.0327 0.031 -1.063 0.288 -0.093 0.028
Results for equation 1
==============================================================================
coef std err z P>|z| [0.025 0.975]
------------------------------------------------------------------------------
const -1.522e-05 0.001 -0.025 0.980 -0.001 0.001
L1.0 -5.163e-05 5.98e-05 -0.863 0.388 -0.000 6.56e-05
L1.1 0.9981 0.001 1584.755 0.000 0.997 0.999
L1.2 -0.0063 0.002 -2.797 0.005 -0.011 -0.002
L1.e(0) 0.0044 0.002 2.512 0.012 0.001 0.008
L1.e(1) 0.0390 0.006 6.702 0.000 0.028 0.050
L1.e(2) -0.0389 0.012 -3.375 0.001 -0.061 -0.016
Results for equation 2
==============================================================================
coef std err z P>|z| [0.025 0.975]
------------------------------------------------------------------------------
const -3.385e-05 0.000 -0.150 0.881 -0.000 0.000
L1.0 5.207e-06 2.41e-05 0.216 0.829 -4.21e-05 5.25e-05
L1.1 -0.0002 0.000 -0.775 0.438 -0.001 0.000
L1.2 0.9916 0.001 1100.941 0.000 0.990 0.993
L1.e(0) -0.0075 0.001 -9.812 0.000 -0.009 -0.006
L1.e(1) -0.0045 0.003 -1.554 0.120 -0.010 0.001
L1.e(2) -0.1088 0.005 -20.812 0.000 -0.119 -0.099
Error covariance matrix
================================================================================
coef std err z P>|z| [0.025 0.975]
--------------------------------------------------------------------------------
sqrt.var.0 0.1405 0.000 409.870 0.000 0.140 0.141
sqrt.cov.0.1 0.0051 0.000 28.134 0.000 0.005 0.006
sqrt.var.1 0.0463 8.13e-05 569.743 0.000 0.046 0.046
sqrt.cov.0.2 -0.0016 0.000 -15.317 0.000 -0.002 -0.001
sqrt.cov.1.2 0.0073 0.000 61.894 0.000 0.007 0.007
sqrt.var.2 0.0232 7.59e-05 306.162 0.000 0.023 0.023
================================================================================
但是，当我尝试向前预测该模型时（与VAR类一样），如下所示：
yhat = varma_fit.forecast(steps=10)
我收到以下错误/回溯信息：
/opt/conda/lib/python3.6/site-packages/statsmodels/tsa/base/datetools.py in _date_from_idx(d1, idx, freq)
84 offset. For now, this needs to be taken care of before you get here.
85 """
---> 86 return _maybe_convert_period(d1) + int(idx) * _freq_to_pandas[freq]
87
88
TypeError: unsupported operand type(s) for *: 'int' and 'NoneType'
文档说，在各个地方VARMA都有问题，但我还没有看到为什么从技术上讲这不起作用的任何原因。任何输入将不胜感激。