ARIMA different results in Matlab and eViews

Dear all,

I am quite new in time series and I find some problems trying to estimate my model in Matlab. I have done the Box-Jenkins method and I got that the model identified is an ARMA(2,3). However, when I compared my results from Matlab with the ones obtained in EViews are a bit different the coefficients, but the information criteria gotten, in this case Akaike and BIC, diverges more. The series I am trying to estimate is a log difference of the IPI of a country, in case it helps. My intuition says it can be related to different way of operate the estimation of the model, but I am not sure about this. In fact, if this is the problem, is there any way to carry on the same criteria to get the same or similar results in both programs? Below is a screenshot of the regression in EViews and my results from Matlab. Thank you very much.

Matlab:

mdljenkins=arima(2,0,3)
estmdljenk=estimate(mdljenkins,difipi)
summarize(estmdljenk)

ARIMA(2,0,3) Model (Gaussian Distribution)

Effective Sample Size: 749

Number of Estimated Parameters: 7

LogLikelihood: 1692.1

AIC: -3370.2

BIC: -3337.87

Value StandardError TStatistic PValue

__________ _____________ __________ ________

Constant 0.0026424 0.0015368 1.7194 0.085543

AR{1} -0.7626 0.3694 -2.0644 0.038979

AR{2} -0.14975 0.081538 -1.8366 0.066275

MA{1} 0.56433 0.36739 1.536 0.12453

MA{2} -0.28602 0.12663 -2.2587 0.023903

MA{3} -0.2449 0.099087 -2.4716 0.013453

Variance 0.00063864 7.2488e-06 88.102 0

Eviews:

Answer 1

[Karanjot]

Hi Jorge,

The discrepancies in the coefficient estimates and information criteria (Akaike and BIC) between MATLAB and EViews may arise from differences in the initial parameter values, optimization algorithms, or convergence criteria used during the estimation process.

Please provide the data associated with the results you shared. Meanwhile, I recommend observing the following:

1. Ensure that initial parameter values are the same in both MATLAB and EViews
2. Examine the optimization settings in both MATLAB and EViews. Compare the optimization algorithms used, convergence criteria, and the maximum number of iterations allowed.
3. Verify that the input data and preprocessing steps (such as differencing or transformations) are identical in both MATLAB and EViews. Inconsistencies in data preprocessing can lead to differences in the estimated models.
4. I also observed that in the results shared, the ‘Effective Sample Size’ & ‘Included Observations’ in MATLAB and EViews respectively are different. Estimation Sample determines the observations used in the rule-of-thumb regressions to decide whether to log the dependent variable, the KPSS tests to determine the differencing order, and the estimation of individual ARMA models are all specific data points or cases that are utilized in these respective analyses

I recommend comparing and investigating the above. To learn more about this, please refer to documentation and examples below:

Create univariate autoregressive integrated moving average (ARIMA) model - MATLAB - MathWorks India

Create Autoregressive Integrated Moving Average Models - MATLAB & Simulink - MathWorks India

I hope this helps!

Regards,

Karanjot Singh