SARIMAX python np.linalg.linalg.LinAlgError: LU分解錯誤

Question

我對時間序列分析有疑問。 我有一個包含 5 個特征的數據集。 以下是我的輸入數據集的子集：

date,price,year,day,totaltx
1/1/2016 0:00,434.46,2016,1,126762
1/2/2016 0:00,433.59,2016,2,147449
1/3/2016 0:00,430.36,2016,3,148661
1/4/2016 0:00,433.49,2016,4,185279
1/5/2016 0:00,432.25,2016,5,178723
1/6/2016 0:00,429.46,2016,6,184207

我的內生數據是價格列，外生數據是 totaltx 價格。

這是我正在運行並收到錯誤的代碼：

import statsmodels.api as sm
import pandas as pd
import numpy as np
from numpy.linalg import LinAlgError

def arima(filteredData, coinOutput, window, horizon, trainLength):
    start_index = 0
    end_index = 0
    inputNumber = filteredData.shape[0]
    predictions = np.array([], dtype=np.float32)
    prices = np.array([], dtype=np.float32)
    # sliding on time series data with 1 day step
    while ((end_index) < inputNumber - 1):
        end_index = start_index + trainLength
        trainFeatures = filteredData[start_index:end_index]["totaltx"]
        trainOutput = coinOutput[start_index:end_index]["price"]

        arima = sm.tsa.statespace.SARIMAX(endog=trainOutput.values, exog=trainFeatures.values, order=(window, 0, 0))
        arima_fit = arima.fit(disp=0)
        testdata=filteredData[end_index:end_index+1]["totaltx"]
        total_sample = end_index-start_index
        predicted = arima_fit.predict(start=total_sample, end=total_sample, exog=np.array(testdata.values).reshape(-1,1))
        price = coinOutput[end_index:end_index + 1]["price"].values

        predictions = np.append(predictions, predicted)
        prices = np.append(prices, price)

        start_index = start_index + 1
    return predictions, prices

def processCoins(bitcoinPrice, window, horizon):
    output = bitcoinPrice[horizon:][["date", "day", "year", "price"]]
    return output

trainLength=100;
for window in [3,5]:
    for horizon in [1,2,5,7,10]:
        bitcoinPrice = pd.read_csv("..\\prices.csv", sep=",")
        coinOutput = processCoins(bitcoinPrice, window, horizon)
        predictions, prices = arima(bitcoinPrice, coinOutput, window, horizon, trainLength)

在此代碼中，我使用滾動 window 回歸技術。 我正在為start_index:end_index訓練 arima 並使用end_index:end_index+1預測測試數據

這是從我的代碼中拋出的錯誤：

Traceback (most recent call last):
  File "C:/PycharmProjects/coinLogPrediction/src/arima.py", line 115, in <module>
    predictions, prices = arima(filteredBitcoinPrice, coinOutput, window, horizon, trainLength, outputFile)
  File "C:/PycharmProjects/coinLogPrediction/src/arima.py", line 64, in arima
    arima_fit = arima.fit(disp=0)
  File "C:\AppData\Local\Continuum\Anaconda3\lib\site-packages\statsmodels\tsa\statespace\mlemodel.py", line 469, in fit
    skip_hessian=True, **kwargs)
  File "C:\AppData\Local\Continuum\Anaconda3\lib\site-packages\statsmodels\base\model.py", line 466, in fit
    full_output=full_output)
  File "C:\AppData\Local\Continuum\Anaconda3\lib\site-packages\statsmodels\base\optimizer.py", line 191, in _fit
    hess=hessian)
  File "C:\AppData\Local\Continuum\Anaconda3\lib\site-packages\statsmodels\base\optimizer.py", line 410, in _fit_lbfgs
    **extra_kwargs)
  File "C:\AppData\Local\Continuum\Anaconda3\lib\site-packages\scipy\optimize\lbfgsb.py", line 193, in fmin_l_bfgs_b
    **opts)
  File "C:\AppData\Local\Continuum\Anaconda3\lib\site-packages\scipy\optimize\lbfgsb.py", line 328, in _minimize_lbfgsb
    f, g = func_and_grad(x)
  File "C:\AppData\Local\Continuum\Anaconda3\lib\site-packages\scipy\optimize\lbfgsb.py", line 273, in func_and_grad
    f = fun(x, *args)
  File "C:\AppData\Local\Continuum\Anaconda3\lib\site-packages\scipy\optimize\optimize.py", line 292, in function_wrapper
    return function(*(wrapper_args + args))
  File "C:\AppData\Local\Continuum\Anaconda3\lib\site-packages\statsmodels\base\model.py", line 440, in f
    return -self.loglike(params, *args) / nobs
  File "C:\AppData\Local\Continuum\Anaconda3\lib\site-packages\statsmodels\tsa\statespace\mlemodel.py", line 646, in loglike
    loglike = self.ssm.loglike(complex_step=complex_step, **kwargs)
  File "C:\AppData\Local\Continuum\Anaconda3\lib\site-packages\statsmodels\tsa\statespace\kalman_filter.py", line 825, in loglike
    kfilter = self._filter(**kwargs)
  File "C:\AppData\Local\Continuum\Anaconda3\lib\site-packages\statsmodels\tsa\statespace\kalman_filter.py", line 747, in _filter
    self._initialize_state(prefix=prefix, complex_step=complex_step)
  File "C:\AppData\Local\Continuum\Anaconda3\lib\site-packages\statsmodels\tsa\statespace\representation.py", line 723, in _initialize_state
    self._statespaces[prefix].initialize_stationary(complex_step)
  File "_representation.pyx", line 1351, in statsmodels.tsa.statespace._representation.dStatespace.initialize_stationary
  File "_tools.pyx", line 1151, in statsmodels.tsa.statespace._tools._dsolve_discrete_lyapunov
numpy.linalg.linalg.LinAlgError: LU decomposition error.

Answer 1

看起來可能是一個錯誤。 同時，您可以通過使用不同的初始化來解決此問題，如下所示：

arima = sm.tsa.statespace.SARIMAX(
    endog=trainOutput.values, exog=trainFeatures.values, order=(window, 0, 0),
    initialization='approximate_diffuse')

如果有機會，請在https://github.com/statsmodels/statsmodels/issues/new提交錯誤報告！

Answer 2

我有同樣的錯誤，

錯誤代碼：mod = sm.tsa.SARIMAX(y, order=(0 1,0), seasonal_order=(1,0,0,12)) res = mod.fit()

這給了我錯誤：LinAlgError: Schur decomposition solver error

在對代碼進行一些修改后，我能夠解決此錯誤：mod = sm.tsa.SARIMAX(y, order=(0 1,0), seasonal_order=(1,0,0,12),enforce_stationarity=False) res = mod.fit()

這解決了我的問題