Historical Data in Simulation Output

[Simone]

Hello everyone. I am not sure if it is the right place to ask this question or not, but I am not professional in running simulations in Vensim, and some how, I need to finish a project in Vensim DSS.
The question is I want to simulate some historical data, and the output is a system containing graphs of both historical data and simulation. The problem is that the simulation graph is so far from the historical data, and I have no idea how I can force the program to decrease the difference between the simulation and real data!
Should I calibrate the model, although I am not optimizing?
I really appreciate if you can help.

[Mohammad Jalali]

Should I calibrate the model, although I am not optimizing?
I really appreciate if you can help.

Hi Simone,

Can you post your model and dataset?

Best,
MJ

[Simone]

Can you post your model and dataset?

Hi MJ,
Thank you for your response, but unfortunately it is a very big and complicated model, and also because of some privacy issues I can not post it here.
Is there any chance that you can help me without taking a look to the model?

[Mohammad Jalali]

Can you post your model and dataset?

Hi MJ,
Thank you for your response, but unfortunately it is a very big and complicated model, and also because of some privacy issues I can not post it here.
It there any chance that you can help me without taking a look to the model?

I see. That's ok. Are you sure that your model is conceptually correct? If the model is not fully correct calibration may still work to generate behavior curves that fit with your dataset, but it's not recommended. When all loops and cause and effect relations along with equations make logical sense, and the model properly simulates, you can calibrate parameters (constants) in your model to get the the best fit with your historical dataset.

Best,
M.J.

[Simone]

I see. That's ok. Are you sure that your model is conceptually correct? If the model is not fully correct calibration may still work to generate behavior curves that fit with your dataset, but it's not recommended. When all loops and cause and effect relations along with equations make logical sense, and the model properly simulates, you can calibrate parameters (constants) in your model to get the the best fit with your historical dataset.
Best,
M.J.

I am not %100 sure that the model works correctly... Maybe I should check it again. I will post the results

Thanks

[Mohammad Jalali]

I am not %100 sure that the model works correctly... Maybe I should check it again. I will post the results

Thanks

Cool! Good luck!

M.J.

[tomfid]

Calibration is just optimization, where the objective is to minimize some measure of the mismatch between model and data.

It's always nice when you can hand calibrate the model, but automation is often better and faster.

You just need a payoff, specifying which model variables match which data series, and a control file, specifying which parameters to vary. If you don't have good intuition for the latter, you can start with an "all constants" parameter sensitivity run, to see which variables make a significant difference.

Tom

[Simone]

Thank you Tom, although I am not sure what you mean by Automation... Right now I am trying to optimize the model using "Policy", instead of "calibration". I hope it works at the end!

[Administrator]

You would need to use the calibration mode to fit to existing data.

Have a look at the example here
http://www.vensim.com/documentation/ind ... ?21295.htm

and just work your way through. That example fits a simulated graph to a historical one.

[LAUJJL]

Hi Simone.

The critical question in modeling is to master the model one is working with.

You write that there is a privacy question about your model and that it is too complicated.

The solution is to extract an ultra simplified model from the original model that summarizes your problematic. If you are not able to do that, then nobody will be able to help you and the best thing for you is first to understand the model you are working with until you are able to identify where there is a problem that you will be able to replicate with a simpler model.

The problem you are working has been either built by someone else out of your sight, or/and built too quickly without intermediate steps permitting to keep the model in control. If it is the case, the solution is to start again from scratch and built the model again step by step with understanding.
It may look like taking more time, but the results will make the difference.

Best regards.
JJ