Strategic Management Dynamics and conventional SD

[gwr]

Hi fellow SD'ists,

having gone through some considerable effort to dwelve into the realm of Kim Warren's strategic management dynamics, I have reemerged from my studies with mixed feelings: On the one hand Kim has managed to (finally) convince me that SD can and should be useful to "ordinary" business people looking for guidance and further more that SD might finally deliver what Kaplan and Norton's Balanced Scorecard promised to do. How different from many other texts did Kim's tome reach out to me: At least someone had had the guts to play Prometheus and truly bring SD to practitioners.

... then came the unevitable "buts" - especially when trying to work things out in Vensim instead of using Kim's proprietary myStrategy - tool:

- how to easily use reference modes to enter a step function (cf. my other post)?
- how to nicely show behavior in the structure instead of the "crude" drafts Vensim offers at pressing the B-key?
(and Kim does it in a nice and convincing way!)
- how to calculate averages for stocks to have instant accounting figures work out for the year ended ...

... but wait a minute - calculating averages for stocks did not feel right - and wasn't right - at least from a classical SD point of view. A lot of Kim's presentation power does come from essentially returning to discrete event simulation setting DT = 1. Is this the future of our trade - to return to discrete event simulation because it is better and easier to sell to management?

Do we "scare people" using stocks and flows to report accounting figures? - I must admit I still haven't come up with a totally convincing answer myself. There may be times where using auxilaries to just go for accounting values in a straight forward fashion is the most appropriate way to go. But then again, in the end that will be how you "educate" the customer. The next time you show him a stock and flow to accumulate cash flow into a stock of cash he might wonder. On the other hand, if it is your first time with the client, should he really see that you "flush out" the stock of cash to give proper accounting figures for every year ended?

If you were to follow Strategic Management Dynamics you would have to calculate the average in the stock (in "truth" an integration for sure). Doing it Kim's way you would end up adding Stock(t-1) and Stock(t) and divide it by 2. But this is the beginning of a road to perdition as one step leads to the other: Having taken averages for the content in the stock you cannot simply return to integrating any derived flow: Let us say you have customers purchasing goods at a certain rate per year. Once you calculate averages in the stock you cannot interpolate between the values given for the purchase rate without spoiling the integration. So this is where we need to have the STEP FUNCTION come into play. With DT = 1 you soon end up struggling hard to find out whether any flows should be modelled :HOLD_BACKWARD: or :LOOK_FORWARD: using exogenous data once DT is set to 1.

All this makes me curious about your experiences with "selling" SD (legitimately) as THE "strategic management tool" of choice? Shall and must we follow Kim's way of going for it or should we rather stick to conventional SD with DT << 1 and without averages in stocks? Or is there an inbetween?

I for once have noted that Kim's way of presenting structure in an aggregated way is "the way to go". I do find it helpful to hide "technical" structure in the depth of tools like Vensim and use aggregated "icons" of SD to represent the model in front of the client.

What is your experience?

Kind regards,

Guido

[kim warren]

Thank you Guido for your generous comments. In truth, my efforts are purely presentational - I do not pretend to offer anything that was not already there in classical SD. The key focus, though is on the stocks and flows that so often get hidden in feedback presentations. Why? - because management of anything, [strategic management in my case] is all about the flows. Since performance at any time depends on the stocks, then if the stocks don't change [and other factors remain the same] then performance doesn't change either. How fast performance changes, then, can only be controlled at the flow-rates - in my cases, customers won and lost, staff hired and lost, products launched and dropped, reputation won and lost. That's why I never, ever show diagrams without explicit stocks and flows, and never without numbers of time-charts on the key items, especially the S&Fs.

I don't think the 'dT' question is such a big issue - I simply don't understand why, for example, an annual model with dT = 0.25 is better than a quarterly model with dT = 1, why a monthly model with dT = 0.25 is better than a weekly model. At least people see numbers they recognise - the customers or staff they have at end of a period, and the revenues and costs that relate to that period. The averaging question Guido mentions is only needed to calculate some things that depend on that average, e.g. sales = fn (average customers in the period), and salaries = fn (average staff). But these nuances are way, way less significant than the wide ranges of uncertainty and data limitations in the situations we model.

... but I may be wrong - I must confess to not being technically very expert in all this.

thanks again Guido. Kim

[gwr]

Kim,

thank you for the kind appraisal of my post. I would not have believed it but DT is - I am afraid - a big issue. I have enclosed a very simple model going just from year 0 to year 1 to make the differences obvious. We start out with 100 customers and they increase linearly to 110 by the end of the year (t=1). This amounts to an average of 105 customers during the year so multiplying this with sales revenue per customer per year of 10 amounts to a total revenue of 10'500 by the end of the year. Simple enough.

With Euler integration and DT=1 you will get an error of 4.76% for just one year with a trivial example. With DT = 0.5 one still gets 2.38% error. By now I must disagree with you, Kim: With all the "uncertainty and unavailability of data" out there the last thing somebody in management will want is some "weird iconography" that adds error to uncertainty especially if it is obvious in a simple example. Will this be inclined to create trust? Probably not - let us hope they will believe that SD will handle the "difficult" problems while "failing" the easy ones?

So no wonder you have gone a different road. I would still recommend not to take it:

1. Setting the integration method to Runge-Kutta in the simple example will give perfect results with classical SD starting from DT=1. So one more reason to keep experimenting with the integration method.

2. While taking the average will give the proper result for the variable "sales revenue at year ended" even with Euler and DT = 1 this variable will loose its meaning if DT is set to a different value than 1 which for many reasons still is good practise (there may be delays). So in a way you are fixing yourself to a time step without having to (cf. 1.).

3. Using stock and flows for the revenue is perfectly reasonable from an accounting point of view - it is even quite easily understandable for people used to accounting as stocks and flows are common ground - something to be used to create mutual understanding. Also one can accumulate accounting figures to have balance sheet stocks a feature that is lost in the direct method or needs tricky programming (?) losing the advantage of the easy presentation. Why not educate the clients from the beginning: SD is stocks and flows.

4. Comparing a model with yearly data and DT=0.25 to a quarterly model with DT=1.0 is an unjust comparison. If you are taking averages than at the very least you would have to compare a model at whatever time period and DT=1 with another of the same period set to DT=0.5. The last model at least allows the rates to change twice as often - it might matter with feedback.

5. Finally building a model with DT=1 in mind blurrs the fact that things happen in continuous time in SD. Even if we only have annual data it is quite a difference whether I have population data for school children at mid year and sales data at the end of the year. I might change prices in the middle of the year and record customers at the end of the year... Building the model with continuous time in mind should be more robust - and in the end you will have to play with DT.

To me it has become obvious that SD is far from easily "sold" and "used" on the client side. It seems essential to have a technically robust model fit to the purpose and to show the model in an aggregate fashion to the clients. In this regard the way you do this in Strategic Management Dynamics seems the way to go IMHO. I can't quite see a shortcut inbetween.

Kind regards,

Guido

---
Guido Wolf Reichert
Management Consultant
Selent, Germany

[Administrator]

Just a note (as I don't recall seeing it mentioned above).

Straight from the Vensim documentation
"TIME STEP should be smaller than 1/3 of the shortest time constant in the model (not applicable with automatic step size adjustment in Runge-Kutta integration)."

"TIME STEP should be smaller than the shortest period for which a significant change in model behaviour is at all likely."

I've never had a problem with time step as long as you follow these guidelines.

Tony.


PS. I tried downloading and opening the model, it failed. Can you upload the unpublished model?

[gwr]

Tony,

I have exchanged the .vpm-file - but I wonder why it failed?! (maybe a path-problem) - and I have enclose the model. In my opinion Kim does have a point as flow-integration will give bad results for DT=1 and the averages are more intuitive. I also have to add that the point with the meaning of the variable has to be modified as the average is usually calculated adding the value of the net flow of customers going through a pipeline-delay (1 year) so that it will always give the year end value for the year before.

I have to admit that I have to take a closer look at the Runge-Kutta as it does not work with flushing a stock like cash in the given model at the end of each year.

Guido

[gwr]

Hi,

I noted that there is silence on this post which maybe is due to my touching an old issue with a beard as long as...?
Nevertheless as to Kim's opinion on the difference between an annual model with DT=0.25 and a quarterly model with DT=1 I have just now come upon Andrew Ford's remarks in "Modeling the environment" (2nd ed.) where he points out that DT has nothing to do with any real world time period and that DT=1 leads to discrete thinking which is not what we use SD for - nevertheless I still wonder whether setting up models with DT=1 makes everthing so "nice and easy" that the advatages might in many cases and for high level modeling purposes indeed surpass the disadvantages?

Anyhow, I still would like to know how we might "join" conventional SD (varying DT and usually setting DT below 1) and strategic management dynamics which focusses on presenting recognizable numbers and not having to explain integration of pulses (and other difficult flow-patterns) to clients used to spreadsheet presentations.

So the question remains: How best to report AVERAGE FLOWS DURING THE PERIOD and ANNUAL VALUES OF INTEGRATED FLOWS in tools like Vensim?

I have build two models with very explicit structures which might then be put into a macro to make presentation and use easier. To me there seem to be two straight forward ways of reporting the average flow during a period:
a) integrate the flow into stock 1 and integrate the delayed (fixed delay of on period) flow into stock 2 and take the difference
b) integrate the flow into a stock that is flushed out (-stock/time step) each period and use that stock to calculate the average flow as well

I found it interesting that there are some differences in the methods. The most important was that the flushing method turned out to be incompatible with RK4-integration which produces correct results early on in terms of DT.

So is the difference method with fixed delays preferable?
Is the stepwise behavior of the average flow in the flushing model due to rouding issues?

Kind regards,

Guido

PS: If these questions have been touched before please point out because I have not been able to locate it in the forum.

--
Guido Wolf Reichert
Management Consultant
Selent, Germany

[LAUJJL]

Hi Guido

I think that the silence comes from the fact that your question has no real answer.

There are many ways a problem can be modeled, and there is no conventional method to do it.

Relying on conventions or on books is deadly, especially if you have to do deal with real life situations. You will have to find the best solution by yourself and be confident that you can do it provided that you allow yourself the time to learn.

Nobody will ever be able, especially through the internet to take into account all the circumstances you have to deal with: type of problem (stake, strategic or operational, time horizon, implementation problems etc..), type of client (knowledge, previous experiences, time available, type of interlocutors), price that you charge (if you charge not enough you will be constrained to deliver a low quality model) and scheduled time to accomplish your job (if you are engaged into a too tight schedule, you will be obliged to deliver a quick and poor model) etc… The only thing that you can do, is experiment and be patient, very patient…

About your precise problem, chose the solution that delivers the best utility to the client, which 'should be' (?) the aim of modeling.

Best regards.

JJ

[Administrator]

Personally I don't see any merit in discussing DT at all. As long as it is chosen based on the guidelines then it only becomes relevant when DT is too large.

For me it is far more important to make sure models are dimensionally correct. It seems a lot of model builders see this as an optional step. I also like to see when a builder has included other mechanisms to give confidence that the model is computationally correct (eg, mass-balance and reality checks).

Tony.

[LAUJJL]

Hi Tony.
Of course I share your point of view, but with some restrictions.

I have already in a last post, pointed out the exceptionally bad models published in the SDS by academics. In the last conference, on 16 Vensim models reviewed, two only were dimensionally correct, only one had a mass balance and of course there were no reality checks.
I have already in the past, made such reviews and this is the reason I have never attended any SDS,
although being highly interested by SD. if the papers are that bad, I suspect everything else to be at the same level of quality.

When I started SD, I studied of course reality checks, but never founding any publically available Vensim models with RC, nor in the accompanying models in the Vensim package, I concluded that it was just an added feature, like those you can see in other software packages that are not used by anybody, because it is not useful practically.

Bur over the time, having tested different modeling methods, I found out that all these methods founded on diagramming, had severe drawbacks that to my opinion are at the root of the poor development of SD.

I used more and more RC last year, but only to check already built models, which is to my opinion not the real interest of the tool.

I decided as I was starting a new model at the beginning of the year, to adopt a strict method of building model with RC as a definition tool, as explained in the Vensim user guide and in the joined paper. It is only after having read this paper some months ago, that I decided to try the method fully. I do not know why this paper is not referenced in the Vensim documentation. I might have decided to try the method already years ago.

By the way if you know places where you can show reality checks used, I would be really interested.
But I wonder if there is anybody building models this way. Since I use RC massively, I found a lot of bugs. I have never found so many bugs in the other parts of Vensim. This proves that RC are very rarely used and probably not been extended and improved since it was implemented in 1994.
Why? Was it tested and finally not found interesting by Ventana people and by clients?

I think that if the method could be really improved and used, it would be a formidable advantage for Vensim and for SD in general that to my opinion is actually badly used.

Since then I have a hard time trying to use the method, but I progress slowly. It needs a total philosophical approach shift that will be long to master and to be accustomed to. I do not use for instance diagrams anymore, even if I can still study the structure of the model once built.
But it would be too long to explain here and I have not made up my mind about the real practical utility of the approach and if so how it must be practically applied.

Do you know people that apply this approach?

Best regards.
JJ

[gwr]

JJ,

thanks for the interesting posts albeit the first one comes to me like a chinese fortune cookie: It all depends on the purpose at hand and there is no theory of model building

So far my personal road of learning has been:

1. Kim Warren does have a point in the way he presents a strategic architecture in that it is (still) comprehensible by SD-laymen and yet goes beyond typical BSC/Strategy Map diagramming. A way to go if SD shall end up with the "masses" IMHO - without taking the CONSIDEO road of "banning" stocks & flows from view.

2. Unfortunately a lot of the simplicity in the presentation (and in the myStrategy models) that go with Kim's book relies on building a model more or less with discrete time in mind (month to month, year to year...) and if unit checking were to work properly a lot of the presentations would lose their charm.
(Nonetheless this still is an open issue for me: Might working with DT=1 in modeling such strategic architectures be a reasonable road to take in order to make modeling much simplier to understand?)

3. There seems to be no easy way to merge the approaches of sophisticated modeling (using RC, unit checks, varying DT) and nice presentation within the modeling-environment itself (e.g. one might need to show special structure just for presentation). Maybe the way molecules work (e.g. structure as an object as far as I have understood it) in Stella/ithink might help Vensim?

4. If two approaches are possible to model a problem (e.g. the "average flow problem" in my earlier post) I might choose the one that leaves the most options for different integration techniques. Personally I still find the difference between Runge Kutta and Euler quite significant and before I had thought of the typical recommendation that "it does not matter in business apps that much".

As for the RC-checks and the like I must admit that it has been Bob Eberlein's course at WPI that has convinced me that this is a useful feature in Vensim. Albeit I still do not know why it is not possible to have a "passive RC check" (e.g. :THE CONDITION: :IMPLIES: stock >= 0 ) is not possible to run without an active one. I would like to get a warning on stocks turning negative without having to activate this separately.

As for the usefulness of RC: RC are for checking "behavior - behavior" statements. But since we have build a model for most of the behavior producing stiuations in many cases I would need to use the optimizer to show that my model is able to reproduce the known behavior of some endogenous concept and if the model could do so THEN we would need to look what has happened on the implication side. Unfortunately this is not how RC work.

I must admit, JJ, that I have not quite understood the way your method works in modeling based on RC without using diagramming?

Kind regards,

Guido

[LAUJJL]

Hi Guido

I did not mean that there is no method building, but no a conventional one.
This means that the method will largely vary depending on the circumstances.
If Kim' method suits you, all the better.
Anyhow with the time and experience, any method applied is evolving.

I do not understand your emphasis on DT and integration method. There is a difference whether you use Euler’s or RungeKutta’s method, but this difference will lessen if you decrease the time step.
And the objective of SD is not to give precise results but an aggregated and holistic vision of a problem and it does not depend on the integration method you use. Rungekutta’ method is probably there for very specific technical problems.

You can use dt = 1 as long as it does not introduce an integration error. See the rule of setting the time step. It has been already exposed in a preceding thread. I have not Coyle’s book nearby, but he proposes a very simple rule based on existing time variable in the model, included for example in delays, smooths etc…

I do not see why there needs to be a necessity in Kim’s models to set a time step equal to 1. It should normally work with any time step smaller than 1, provided that the model is time step independent.
I have just looked at a mystrategy model about rivalry, you can change the time step that is set in this example to 0.125 with a time unit = one month. There is a time setting clock just besides the run button. So I do not understand your problem about time step setting in mystrategy. Kim will be certainly more able to understand your problem than me.

The great advantage of Mystrategy is its simplicity. It is even simpler than Vensim Ple. And it is always better to start with simple tools. Of course after some time, you may need to step up to a more advanced tool.

Unit checking is not at all sophisticated and very easy to implement as is mass balance. I do not see either the interest of making dt vary as a sophisticated method.

Rc on the contrary if used properly is a highly sophisticated method. This is one of the reason it is not used at all plus some others.
I have followed the same Eberlein’s course at the MIT years ago. It did not unfortunately exposed RC as a building method tool, but only as a testing tool, used after the model is built.

To see an example unfortunately very simplistic of building a model using RC, see the example in Vensim’s user guide chapter 14.

A passive RC is only checked each time any number of active RC are checked. See example joined.
I do not think that it is possible to get a warning independently from the RC.
But I do not see the real problem with it, unless you do not want to activate any RC checking at all. It is all a question of habit. If you get accustomed to RC checking, it makes no difference whether you set a stock non negative constraint as active with a time >= 0 input or as passive with no input at all.
All these are details questions and model building problems are elsewhere.

It is perfectly possible to implement calibration in RC. In the model I am building presently I check with several RC that a time smoothing constant is calibrated to some exogenous inputs. I will be able to keep that checks all along the modeling process and will always be able to check that for some reason that time constant becomes incoherent with an assumed behavior of the model.

About the non diagramming method, it would be too long to explain here. I am still testing the method and it will still take a long time until I have an advice based on sufficient practice.

RC and quality of models is not much of a concern among SD people at least as I see quality.
The notion of quality is anyhow very subjective.

I have already had the occasion to realize that in the past on this forum and on the SDS forum. But it is probably that I see things differently than most SD people. I build models and I use them.

Best regards.
JJ

[gwr]

Hi JJ,

thanks for pointing at chapter 14 in the User's Guide to Vensim and to hinting at how to turn passive RC to active RC and on to how to have RC work as a check to the fit of the model to data. I wondered about checking the fit when reading the most interesting article by A. Graham about "fit constrained Monte Carlo" (obviously parameters are constrained by a certain level of fit to data that has to be satisfied so this should be considered when drawing random samples from possible outcomes...) - but I figure this can probably not be done by using RC but rather has to be done externally in a spreadsheet?

As to the "Discussion about DT" being quite irrelevant I cannot help but to note that this is NOT the main point in - hopefulle constructively - critizising Kim Warren's approach to strategic model building.

Please let me once more make this clearer:

1. I am having trouble in using an arithmetic average for a stock (Kim regrets that stocks in SD do not report their content during the period but only at its end) to calculate an instant accounting number which remains an approximation of essentially an accumulation over a period.

As you have pointed out, JJ, one can change DT to say 0.125 in Kim's myStrategy models. I was well aware of it but that is not the point. One needs to ask what it will change in Kim's models? Here the answer is: Not much!

In the models I have seen the average of a stock is calculated for one period (a year, a quarter ...) by the using a DELAY FIXED for one period. Thus the average stock is calculated by:

AvgStock in one period = [Stock(t) + Stock(t-1)]/2

And that is the point: Even if DT is set to smaller values, the additional accuracy in calculating stock values will not be used, as values inbetween are never used in the calculation! This will play out if there is nonlinearity in the stock.

2. Whatever you do, the discrete approximation of the integration of a flow (even in accounting ) is a SUM. But there is no sum in the calculation as it is approximated by a multiplication of averages. This is definitely not a robust way to do it since this will not work if the other parameter that is multiplied with the average content of the stock is changing also, e.g. if the frequency of purchases for a customer changes along with the the stock of customers. (This is actually a flaw in the diagramms in Strategic Management Dynamics as Kim has recognized himself and will be changed in further editions) as there the charts for price etc. are shown not as step function but as continuous interpolations.

3. Tony and you, JJ, have pointed out, that accuracy in numbers is not that important and thus DT is a minor consideration, if one sticks to the rule of thumb to set DT at between 1/4 or 1/10 of the smalles time constant in the model. But that is a further point against what Kim has put forward as an argument for using averages in stocks in the process: "to have clients recognize their numbers".

The only way to achieve a higher accuracy is to make the integration of the flows more accurate. Kim does not doe this as the flows are not integrated or added. So if taking averages for the content in a period, this should be WITHIN DT. And essentially this is what more sophisticated Integration methods like Runge Kutta are trying to achieve.

To sum it up: Accounting flows should be integrated to calculate the average flow over a period - this will also allow to use the flow for balance sheet stocks. To achieve higher accuracy - if this is considered an issue - one should set DT to smaller values or use a different method of integration. The aproach chosen in Strategic Management Dynamics IMHO is not a robust one and in my opinion diverges (unnecessarily) from common practise in SD-modelling. Calculating averages for a period in a stock will actually make varying DT useless and you should be aware of this.

Kind regards,

Guido

[gwr]

The Formula for the average content in the stock has been corrected to:

AvgInStock = [Stock(t) + Stock (t-1)]/2

Sorry for (previous) "sloppiness"...

Guido

[LAUJJL]

Hi Guido

First, you can check everything using RC. I can check reference modes, whether these are real or assumed. One possibility instead of using reference graphs is to use reference points that are easier to explain and to figure out.

About your difficulty with Kim's book, you cannot expect a text book to solve all the problems one can find in real life!

I too do not agree with some of Kim's method.
For instance he supposes in rivalry that competitors will not react to the change in policies which is true if the competitor who changes his policies has a very small market share, so as not to trigger any reaction from the competition.

He studies too Ryan air in his last book. But I know two persons who have negotiated with Ryan air. One was the president of Strasbourg's commercial chamber who manages the airport and the other was the manager of Lahr's airport not far from Strasbourg in Germany. None of them worked with Ryan air, for different reasons. One had to face a suit from Air France and the other had not enough resources to satisfy Ryan air’s subventions demand.
Ryan air development driving force is public subventions provided each time a new route is started. This generates a kind of bubble extremely dangerous that can burst anytime or that necessitates special policies avoiding these drawbacks. But I imagine that Ryan air is not willing to show the percentage of its turnover made of subventions and probably Kim had not the possibility to put this in the model.

But all SD books face this problem more or less.

About the way stocks are managed, I did not find anything especially annoying in Kim’s book, but I do not have any stock to manage in my business (car and truck rental) and I am not highly concerned with inventories.
But I still do not figure out what is exactly your problem.

The best would be to join a highly simple model that illustrates better your problem than a long explanation. Maybe I did not study well the models that you may have posted before about that problem. But there is always a way to solve a problem, one has just to take the time to think about it.

For me one has to choose a time step that corresponds to the level of details chosen. If you want to trace how the stocks vary during a period, than you have to reduce the time step.

But I think that the most current error done in modeling is to put too much material than needed or if one does it, not knowing exactly the purpose of doing it. This is why I advocate a strict top down approach, starting from the highest possible aggregated point and stepping down only if one knows why one does it.
That would mean starting first with a higher time step with no integration problem, using the model and decreasing the time step only when one has identified the problem generated by a higher time step. But one of the problem’s with this approach is that it is only the client that can decide if the model suits his needs and the corollary is that the client must participate closely to the model building, which for a consultant like you may be impossible to achieve.
Best regards.
JJ

[LAUJJL]

Hi Guido

I understand more or less your problematic.

You want to model accounting which is a process of summing up discrete events on a specific period with a continuous tool like SD to make the numbers provided by the model similar to the accounting numbers.
You can do that only by keeping the period used in the accounting equal to the time step in the model, calculating each period rate as a variation between the period end stock and the period beginning stock, rate = (stock at the end of the period – stock at the beginning of the period) / time step and introducing these rates as exogenous data.
But if you want to make these rates endogenous, you will have to make adjustments and you will lose the exact numbers. But anytime you make a model, it is only a simplified representation of the reality helping to see the big picture, especially in strategic issues, and there is no need for that to get so close to the reality.
Otherwise you get too many details in your model and it is no more useful to anything being too complex.

Depending on your clients, depending on their being results oriented or thinking oriented, you must decide if you make a model results oriented or helping to think, the latest being the real utility of modeling. You can eventually try to explain that to your client if he is evolved enough.

Being a consultant in SD, is not easy because you have to take into consideration the problem at hand and the type of client too with his expectations (generally wrong), phantasms, incorporate all his perceptions of the problem that stand at different unknown level of details to make him believe that the model really represents reality and finally generate a model that is as complex and difficult to understand than reality.
This is why I believe in top down modeling, each detail being taken into consideration in proper timing and if necessary and monitoring closely the utility expected and observed at each additional step so as to stop the modeling process when progressing does not add any more added sufficient enough utility. This type of modeling requires a close participation of the client and delivers a value all along the modeling process and not at the end when the model is finished.
The shorter the modeling step the better but the more work it requires.

I do not pretend that this method can be applied by everybody, probably in too few cases, because of client’s availability and acceptability. This method requires a lot of experience from the modeler too, using RC and managing the client.

In fact it is simply the method advocated in the Vensim modeling guide, progress step by step from the simpler to the more sophisticated, using eventually the RC method explained in Chapter 4 and in the RC_justification file I posted.
It is too important to use extensively the intermediate versions generated by this top down approach to help chose the next step.
If there is a doubt about the next step to chose, I build a model for all the candidates and I choose the model that generates the best added utility. It is too if possible preferable to choose between the candidates by other simpler methods that will not require building multiple models.

Best regards.

JJ

[gwr]

JJ,

thank you indeed for being engaged in this discussion it certainly is quite rewarding - especially since it got me to take out my notes from Bob's course again and to remember the nice features of Reality Checks.

As to your understanding my posts: It was actually a quite "trivial" but principal question that got me thinking about this - averages in stocks and how to model accouting figures (as these - especially cash flow - usually are among the main performance measures touched in Strategy Dynamics and in reality). I have enclosed a paper by Eric Melse about dynamic accouting that quite nicely touches theses questions but esentially you can start with a quick look at Vensim's chapter 7 of the Modeling Guide (Financial Modeling and Risk). Quite "naturally" accouting figures are represented as flows and become integrated. So we have what Melse calls the (integrated) ex post values and the current (ex ante) values of the simulated, most recent flows - a perspective that gets lost if the flows are calculated using averages of what is in the stock.

So why diverge from this "natural" way of modeling accouting flows without a "real need". I agreed that accuracy in numbers may in many, many cases not be important at all considered what SD has to teach us and that if it were a need then lowering DT or choosing another integration method would be the way to go in "traditional" SD. I showed that Kim's way of calculating the average in the stock prevents DT from helping in accuracy if the stock changes in a nonlinear way.

I have come to find that in the world of business and its messy details (unlike in public policy modeling...) there appear to be at least some occasions that call for a hybrid approach in modeling mixing discrete (noncontinuous) events with the overal continuity of the SD model. As you pointed out, JJ, at some times one would need to bring in accouting flows as a pulse if one wanted to be true to reality and if this kind of accuracy were needed. Nevertheless, showing and reporting theses flows to the client is a bit "nasty" as most people in business do not like the idea of thinking like an electrical engineer who is at home in the world of continuous signals... For those (poor?) non-engineering souls we might want to report "average flows", e.g. a flow that matches the model's or the accounting period. So this was the reason for my trying to model this transformation as either the difference in the stock using a pipeline delay or as a stock being "flushed out" each period like a counter being reset. Unfortunately later method is not working with Runge Kutta so I (still) wonder whether the method using a pipeline delay is the preferable one? (Any suggestions for other reasons, e.g. model performance etc.?)

Finally, thank you again for reminding me about the function of RC as guidelines for a model to evolve around (chapter 14 of Vensim's User's Guide). That is a very nice feature indeed and also RCs allow you to develop a "battery" of tests and sensitivity experiments that can be ported from on stage or version of a model to the other. I will certainly intensify my usage of RC.

@Tony and the Vensim developers: Vensim 5.10e unfortunately (at least my installation) will not change to "TrialPrio" instead of "Range:" when the variable type "Reality Check" is choosen. I have not yet tested whether is is just a "GUI-error" or whether the whole functionality is missing...

Kind regards,

Guido