Does use of flow measurements actually solve the autocorrelation problem?

[tomfid]

Problem: there's likely to be autocorrelation in the residuals for stock measurements, if the model doesn't filter to adjust for driving noise.

Proposed solution: use flow measurements instead (production starts and finishes rather than WIP inventory, for example).

Question: (to what extent) does this actually work?

In an open loop system, it should work, because the flow is a first difference and therefore whiter than the stock. But in the closed loop version, flows = f(stocks), so it seems like there should still be autocorrelated disturbances in the flows arising from the noise previously integrated into the stocks.

Is some of the real-world advantage here that the flow measurements will capture some of the driving noise influencing the stocks?

[hyunjimoon]

Isn't all argument related to prior/posterior/stock/flow relative? #103

We can't get free from autocorrelation (we are injecting process noise, after all), but stocked flow is lesser evil (as it is one step before the stock..?) was my understanding. From the diagram below I abstracted the inventory model to trace process noise's effect on estimation, so using stocked pping, ping as observed data may have lower autocorrelation than using SS, S.

[tomfid]

What is EEOR here? Looks like the expectation process is 2nd order? Is this for extrapolation, or something else?

I'm not sure the expectation adjustment rates need to match the stock adjustment rates - seems like the relationship might vary depending on what demand noise looks like.

[hyunjimoon]

My subconsciousness analysis report:

• I thought taking expectations over different adjustment times would have the effect of a layered low pass filter; e.g. if adj info frac1 is 7 and adj info frac2 is 30.5, perhaps we could capture multi-seasonality. Long and short-term trend at the very least like this Aki's birthday casestudy because the structure of gaussian process' correlation time and process noise's autocorrelation all seem similar to me...

• EEOR is expected expected order delay. The idea of ever-growing hierarchical model motivated the above model where stock's stock and preparing the preparing can be imagined. In two layer model, WIP is Inventory's stock as production start rate prepares production rate. Deeper layer would have longer delay to reach to the delivery state which is why I connected that with the longer smoothed order rate (expected expected order rate). As such, in n^th layer model, we would inductively have hierarchical stock $S_n$, preparing rate $p_n$_ing, expectations $E_n$OR (order rate).

• ARIMA statistical analysis differences the real world time series until white noise is excavated. I cannot justify enough yet, but I thought clear time series decomposition and frequency-based allocation to different levels of resources at hand is a must for efficient management.

[hyunjimoon]

In other words, seasonalities statisticians extract from real data may be generated from adjustment mechanisms (balance loop) with different aggregation level and time horizons, represented as both information and material adjustment fraction: E_n_OR and S_n+p_n_ing.

Again, the aggregation level, time step, and time horizon are all resolution (dynamic and spatial aggregation and separation in #34, #120)

[hyunjimoon]

Hierarchy in forecasting as information adjustment and preparing stocks as material adjustments collaborate to react to "the others" where "one-the other" relationships are overlayed in a recursive manner. Question of how forecast accuracy affect supply chain slide gives us reference as below:

• Aburto, L. and Weber, R. (2007) Improved supply chain management based on hybrid demand forecasts
• Disney, S.M. and Lambrecht, M.R. (2010) On replenishment rules, forecasting, and the bullwhip effect in supply chains. 2008th
• Forrester, J.W. (1958) ‘Industrial dynamics: A major breakthrough for decision makers’
• Hu, Y., Zhang, J. and Xu, Z. (2007) ‘Asymmetric demand information’s impact on supply chain performance and relationship under price-only contract’
• Lyneis, J.M. (2000) ‘System dynamics for market forecasting and structural analysis’, System Dynamics Review
• McCullen, P. and Towill, D. (2002) ‘Diagnosis and reduction of bullwhip in supply chains’
• Moran, F.V. and Barrar, P. (2006) ‘Supply Chain Dynamics: Structural Causes of the Bullwhip Effect’
• Ouyang, Y., Lago, A. and Daganzo, C.F. (2006) ‘Taming the Bullwhip Effect: From Traffic to Supply Chains’
• Sterman, J.D. (1989) ‘Modeling Managerial Behavior: Misperceptions of Feedback in a Dynamic Decision Making Experiment’
• Sterman, J.D. (2000) Business Dynamics: Systems Thinking and Modeling for a Complex World. McGraw-Hill Higher Education.
• Terwiesch, C. et al. (2005) ‘An Empirical Analysis of Forecast Sharing in the Semiconductor Equipment Supply Chain’, Management Science
• Torres, O.C. and Moran, F.V. (2006) The Bullwhip Effect in Supply Chains - A review of methods, components and cases

[hyunjimoon]

Could you send me the vensim model? I have questions about your model such as:

• is shrinkage observability multiplicative effect?
• is inflow of expected loss rate hidden?
• where does driving noise enter?

[hyunjimoon]

@tomfid Would the following from your https://metasd.com/2012/05/bathtub-statistics/ one reason why you feel less welcoming for flow-based measurements?

Derivatives amplify noise, and it’s easy for other things to go wrong, like getting the reporting lags misaligned.