Model computational Limits?

[stoddj]

I am using EViews 12. I have worked with EVeiws Systems and Model commands for several years now, but just ran into a problem I can’t diagnose. I built MOD_PNT2 from VAR_PNT2, then changing that to SYS_PNT2, which I then revised with exogenous taxes added onto the endogenous price terms. Running the model MOD_PNT2, I get the following error message.
Model: MOD_PNT2
Date: 03/25/24 Time: 13:04
Sample: 2020M01 2024M12
Solve Options:

Code: Select all

Dynamic-Stochastic Simulation
Solver: Broyden
Max iterations = 5000, Convergence = 1e-08
Requested repetitions = 10000, Allow up to 2 percent failures
Solution does not account for coefficient uncertainty in linked equations
Track endogenous: mean, 95% confidence interval

Calculating Innovation Covariance Matrix
Sample: 1973M01 2019M12
Matrix scaled to equation specified variances

Scenarios: Scenario 1 & Baseline
Solve begin 13:04:54
Repetitions 1-1: 1 failure(s) 13:04:54
Solve terminated - Unable to compute due to missing data in "D(LOG(CO2_A_0MB)) = @COEF(1) * ( LOG(CO2_A_0MB(-1)) - 0.53995156 ...
Simulation terminated before specified successful repetitions
0 successful repetitions, 1 failure(s)

When running the diagnostic check on the MODEL in deterministic mode, it tells me that the program gives NA for the variable CO2_A_0MB at date 2024M01.
2023M11 Solved (recursive block)
2023M12 Solved (recursive block)
2024M01 NA generated for CO2_A_0MB
2024M01 Not solved
Solve terminated - Unable to compute due to missing data in "D(LOG(CO2_A_0MB)) = @COEF(1) * ( LOG(CO2_A_0MB(-1)) - 0.53995156 ...

Because of the “missing data” message, I checked to see if any data might be missing for that or later dates, but it is not. The list of all variables can be checked in the model MOD_PNT2 and is collected in the workfile as Group02.

However, if I re-run the same model but excluding that date, with the estimation window now closed after 2023M12, the output is fine:
Model: MOD_PNT2
Date: 03/24/24 Time: 21:19
Sample: 2020M01 2023M12
Solve Options:
Dynamic-Stochastic Simulation
Solver: Broyden
Max iterations = 5000, Convergence = 1e-08
Requested repetitions = 10000, Allow up to 2 percent failures
Solution does not account for coefficient uncertainty in linked equations
Track endogenous: mean, 95% confidence interval

Calculating Innovation Covariance Matrix
Sample: 1973M01 2019M12
Matrix scaled to equation specified variances

Scenarios: Scenario 1 & Baseline
Solve begin 21:19:04
Repetitions 1-1600: successful 21:19:11
Repetitions 1601-3200: successful 21:19:17
Repetitions 3201-4800: successful 21:19:23
Repetitions 4801-6400: successful 21:19:29
Repetitions 6401-8000: successful 21:19:35
Repetitions 8001-9600: successful 21:19:42
Repetitions 9601-10000: successful 21:19:43
Solve complete 21:19:43
10000 successful repetitions, 0 failure(s)

I also find that I can run the model with the 12 additional output targets for 2024, but in a simplified form without the exogenous tax data in the underlying System equation. This pairing is shown with SYS_PNT222 and MOD_PNT222.

Thus, I am led to the preliminary conclusion that I have overstretched the computational bounds of the version of EViews I am using. So, 2 obvious questions:
(1) Is this conclusion correct?
(2) If so, what can I do about it?

I am happy to attach the EViews workfile referenced here, but do not see a way to do this currently.

Thank you,
Jim Stodder, Boston University
jim.stodder@gmail.com, jstodder@bu.edu

[stoddj]

I should add that my copy of EViews 12 is the Enterprise Edition, Oct. 18 2023 Build.

The data set is 13 columns of monthly data from 1973 to 2024 inclusive, or 13 x 52 x 12 = 8,112 data points. Therefore, it seems unlikely that the computational limits would have been exceeded.

The SYSTEM object underlying the MODEL is estimating a VAR of 10 endogenous and 3 exogenous variables in a Vector Error Correction Model with two identified Cointegrating equations.

[EViews Matt]

Hello,

Given that there are no missing values in the data and that simpler variations of your model solve successfully, my first thought is that the full model includes some type of numerical issue that only the data for that one observation happens to trigger. I would certainly like to examine your workfile. When posting a forum message, there should be an "Attachments" tab below the main message text field. Otherwise, you can send your workfile along to support@eviews.com.

[stoddj]

Thank you very much, Matt. Please find the file attached below. There's a lot of MODEL output (variable simulations in various forms), but the files I mentioned by name should all be there.

[stoddj]

Just a note, Matt. Following your initial thoughts on the problem, I tried simplifying the data entries for that one observation, 2024m01, by rounding them all down to integer values, down from the default 16 or so decimal places.

But I got the same error message from running model MOD_PNT2 in Deterministic mode:

2024M01 NA generated for CO2_A_0MB
2024M01 Not solved
Solve terminated - Unable to compute due to missing data in "D(LOG(CO2_A_0MB)) = @COEF(1)

So looks like I'm still stuck. // Jim Stodder

[EViews Matt]

Hello,

It turns out this is a run-of-the-mill missing data issue. In model MOD_PNT2, the equation for variable CO2_A_0MB references exogenous variable CPI_12(-1). Unfortunately, that series doesn't contain data (only NAs) after 2023M11, thus there's going to be a problem in 2024M01 as you've experienced.

[stoddj]

Thank you so much, Matt. I'm abashed -- it was actually a coding error on my part in the System equation. I should have been referencing the generated inflation variables which do span the time frame, rather than the actual CPI variable that does not. I'd love to know how you debugged it so I can do it myself if I get in a similar fix. Was there a diagnostic tool available to me?

You have done me a real service and I owe you. If there's an Eviews-sponsored charity, please let me know. // Jim Stodder

[EViews Matt]

My pleasure. While not exactly the debugging path I followed, this particular type of issue would be revealed via a simple visual scan of the exogenous data for NAs. The model "Proc -> Make Group/Table..." procedure is convenient for quickly generating a group of just the utilized exogenous variables, which can also help highlight an error in specification.

[stoddj]

Thanks again. I apologize for my silly oversight. Again, name a charity, if there is one that Eviews or you prefer. // Jim S.