Re: [IM] training model based on historical data

[Pradeesh Kumar K V]

Hi Dave,

I am not sure but I think what you may be looking for is model calibration.

Once the structure of model is fixed, calibration takes real world data for the different stocks over time (time series) and calibrates the parameters in the model (for example, egg survival, age 1 survival etc.) so that the stocks replicate the real world time series.

I know calibration option is available in paid version of Vensim and may be in other similar softwares too. I should also add that I have done SD model calibration only during my course work. 

Hope this helps.

Regards,

Pradeesh

On Wed, Mar 10, 2021 at 1:14 AM Dave Kowalik <dmko...@gmail.com> wrote:

Thanks for the quick response Gene. Yes, I was imagining time -  so per my model, for instance, one could create a converter titled historical age 4 (1900 = 10000, 1950 = 5000, 2000 = 2500, 2021 = 1000....) then set the already present stock titled Age 4 equal to that converter and have the historical time trend inform the model as it projects into the future, as opposed to stopping at the last converter value in year 2021 which was 1000. I suppose this could be 1.) done outside of this model with a time series forecasting model and then those future values plugged back in to insightmaker, or 2.) I also suppose, one could assign a vector of values corresponding with various years and use that for the stock values, or 3.) One could create a random number distribution or function that is reflective of the overall trend for a given stock value through time (in the converter example I just provided this would be a decreasing linear, or curvilinear-like function.)

I was curious as to if there were specific examples/demos of this type of process being used as opposed to simply using a flat line of the last predicted converter value repeatedly into the future (which in our case would be 1000 in year 2021).

This may however be getting away from the intention of this style of projection modeling.

perhaps a more useful way to approach this would be to generate switches like you have used in your scalar/vector example (vectors example from the Phoenix Project). My next question then is - is it possible simply to add a switch for the overall model chain or do I have to define it for each stock, linkage, and/or variable? An example as it pertains to my model would be to add a multinomial switch with values 0 = 90% reduction in the total chain, values 1 = 0% change in total chain (current state), and 3 = 50% reduction in all values of chain). I went to play around with the scalar/vector rabbits example you provided but could not generate a grouped toggle switch like this. If you have an easy way of creating a switch for a big complex model full of random number generators where it may be difficult to pass vectors through each linkage in the chain I would be interested in seeing more.

Thanks so much for all you do, and greatly appreciate this tool

Dave

On Monday, March 8, 2021 at 9:34:13 PM UTC-8 Gene Bellinger wrote:

Dave,

I'm not sure about the replacement since the input to a converter is a value and the input to a stock is a flow, so if you try to replace the stock with a converter where would the appropriate X value come from, unless it's just Time.

On Tue, Mar 9, 2021 at 12:16 AM Dave Kowalik <dmko...@gmail.com> wrote:

Good morning, hope all is well. I was reviewing some insight maker historical functions, converter tutorials, and the community forum and had a couple of questions.

1.) in a stock - flow - transition model is it possible to train a model and supplement stocks, flows, or variables with historical/empirical data? essentially replace the stock with a converter for instance?

2.) if that is not possible, could someone please provide more examples/tutorials of clever converters being used to dictate stock and variable levels? imagining this is an ifelse statement, but would like to see how this could be incorporated maybe with a probability if one historical converter level is less likely than another.

I have provided my insight model as an example as I wish to use historical measurements of stocks and/or introduce predictor variables (converters) that have some historical measurements.

https://insightmaker.com/insight/220399/Chinook-Salmon-PVA-1

thanks in advance for the help.

Dave

--
You received this message because you are subscribed to the Google Groups "Insight Maker" group.
To unsubscribe from this group and stop receiving emails from it, send an email to insightmaker...@googlegroups.com.
To view this discussion on the web visit https://groups.google.com/d/msgid/insightmaker/CAB9gJXRmtuZ1dw_k0JkESmxcnRzskDmu0PWUo1FDY4OgusEWhA%40mail.gmail.com.

--
You received this message because you are subscribed to the Google Groups "Insight Maker" group.
To unsubscribe from this group and stop receiving emails from it, send an email to insightmaker...@googlegroups.com.
To view this discussion on the web visit https://groups.google.com/d/msgid/insightmaker/aba6f254-96c0-424e-a0e6-824aec4c176bn%40googlegroups.com.

[pradeeshkkv]

Oops. Sorry about this. Wrong email id.

Pradeesh