I am using the statsmodels package in python to generate a set of ARIMA models for a series of log returns multiplied by 1000. I am iterating through possible models (p, d, q) starting from (1, 0, 0) and going all the way to (5, 2, 5) trying every combination. I get a ValueError with the message for many of the models:
"The computed initial AR coefficients are not stationary. You should induce stationarity, choose a different model order, or you can pass your own start_params."
From the line:
model.fit(trend = 'c', solver = "powell")
To me this means that perhaps my model does not represent the underlying series well. However, only about 20% of the models do not run into this error and the models that work are randomly spread out. For example (1,0,0) works but (1,0,1) and (1,0,2) do not, however (1,0,3) works again. To make matters more confusing when evaluating some of these models in R I do not run into this problem. Am I right to be concerned about this? What might the problem be? Would decreasing the tolerance and/or increasing the number of function evaluations help? Thanks for any tips.