Engineering Modeling

Correlation and Impact on Monte Carlo Analysis Results (5/8)

Karl - CB Expert

Share:

Print

Rate article:

4.0
Rate this article:
4.0

All the top dogs in the Monte Carlo Analysis spreadsheet universe have distribution-fitting capabilities. Their interfaces have common elements, of course, since they rely on (for the most part) the same PDFs in their arsenal of distribution-fitters. There are important differences, to be sure. It is hoped this comparison will illustrate pros and cons from a practical standpoint. Before going over our scorecard between Crystal Ball and ModelRisk, there is one more very important capability category begging for review: Correlation.

Correlation indicates a rough behavior, or a degree of relationship, occurring between two or more uncertain variables. It can be defined mathematically via a single value (or a few) but, being rough and not exact, it adds its own unique noise to the MCA world. I sometimes call it "a lazy causality." True, there might appear to be causality creeping out of the data. But any attempts to prove the causality (from a statistical hypothesis viewpoint) does not provide more confidence in a transfer function, like a polynomial derived from Design of Experiments (DOE). Still, we know this behavior is important so shouldn't our MCA software provide this capability between two random variable inputs? And how good should it be?

The answer is: It depends. (Don't you hate it when people say that?) Here is the primary reason that modeling correlation between uncertain variables is important: It affects the output variation, either by increasing or decreasing it. It does not affect the predicted mean. It does affect the predicted variance. (We can say this with a high degree of confidence. How much it affects the variance is another question. Along with the question of how much does it really matter.)

Here's how: Let us use the simplest type of additive model on the planet, the one-dimensional tolerance stack. There are three blocks stacked within the gap, making up our system. It is certain that the block widths will vary from assembly to assembly. But would they vary the same way (larger or smaller) for each assembly made? (See Figure 5-1.) That is, if Block 1 width is on the larger side (larger than the mean), would Block 2 width exhibit a tendency to be on the larger side? Since we are talking about blocks most likely randomly pulled from a box of blocks, any correlation would be minor. But what if they did? A positive correlation between Block 1 and Block 2 would accentuate any existing variation in the system output (the gap). That is, when Block 1 and Block 2 get wider at the same time, that has a tendency to influence the gap to its extreme minimum values (gets smaller more often). With positive correlations, there are fewer instances of "offsetting" variations (Block 1 is randomly wider while Block 2 is randomly smaller). Over many trials, correlation behavior accentuates variation in the output. Even though correlation may not exist for blocks (or bearings) and therefore not important to model (in the one-way clutch), it has been shown to rear its ugly head in other areas like finance. For some models, correlation is very important because it directly affect the risk predictions. It either shows them up as being too conservative or under-predicting the risk. Remember: Not the mean. But the standard deviation.

Correlations are traditionally calculated from paired data. Examine Figure 5-2. For all three plots, the axes match one uncertain variable (x1) against another (x2) and each point represents a pair of linked data points. The two variables were measured at the same time or for the same instance (very important). These scatter plots provide a sense for the magnitude of correlation. If there is no correlation between the two variables (as in the first plot), the random nature of their occurring values will "spread out" over the entire variable range space. However, as shown the next two plots with greater and greater positive correlations, order peeks out from the chaos. The paired data sets start to congregate around the diagonal of the plot. In the last plot, the correlation is so great that no data pairs are observed where x1 is low and x2 is high, as well as no pairs observed where x1 is high and x2 is low.

The most common calculation associated with correlation characterization is Pearson's product-moment coefficient. It was designed to indicate a degree of linear relationship between two variables. For values approaching +1, this indicates a positive correlation where both variables tend to increase or decrease together around their central tendencies. Values approaching -1 indicate a negative correlation where one variables tends to increase while the other tends to decrease. Pearson's coefficient uses the raw variable values but a similar calculation places the rank of those values (descending from greatest values to smallest) into the same formula. This version is called the Spearman's coefficient (or Rank Correlation coefficient, these are the values identified in Fig. 5-2).

Excel and MR provide functions to calculate these coefficients plus one more. When data1 and data2 are the ranges of the paired data cells, the formulas are:

  1. Pearson's Coefficient → =CORREL(data1, data2)
  2. Spearman's Coefficient → =VoseSpearman(data1,data2)
  3. Kendall's Tau Coefficient → =VoseKendallsTau(data1,data2)

Most decent MCA spreadsheet packages offer correlation capabilities using Spearman's because it is simple to accurately implement (relative to Pearson's) between random variables characterized by different distributions. In essence, the sampled variables retain the intended PDF behavior. At the same time, individual trial values from one variable are paired up with trial values from another variable, such that a calculated Spearman's coefficient approaches the desired correlation value (defined by the user).

ModelRisk stands apart from the crowd with a more sophisticated way to enforce correlations termed Copulas. The reason for this distinction is that correlation coefficients alone do not provide enough flexibility for certain classes of correlated variables. It is possible that two different correlation patterns discerned in scatter plots provide the same correlation coefficient. For particular finance or insurance simulations, this distinction is quite important. Without more flexibility to capture the appropriate correlation behavior, the analyst may either under-estimate or over-estimate the risk. And, as with distribution-fitting, MR offers criteria that indicate goodness-of-fit for the five closed-form Copulas in the MR arsenal. Again, this offers more flexibility to the user instead of simply plugging one single coefficient value that pairs up PDFs. (As a side note, MR also has an empirical data-fitting version for Copulas. CB does not.)

Now that the theoretical underpinnings of Correlations and Copulas have been discussed, let us use the Duke Basketball Scores to implement the behavior practically in our MCA spreadsheet alternatives.

Comments

Collapse Expand Comments (0)
You don't have permission to post comments.

Oracle Crystal Ball Spreadsheet Functions For Use in Microsoft Excel Models

Oracle Crystal Ball has a complete set of functions that allows a modeler to extract information from both inputs (assumptions) and outputs (forecast). Used the right way, these special Crystal Ball functions can enable a whole new level of analytics that can feed other models (or subcomponents of the major model).

Understanding these is a must for anybody who is looking to use the developer kit.

Why are analytics so important for the virtual organization? Read these quotes.

Jun 26 2013
6
0

Since the mid-1990s academics and business leaders have been striving to focus their businesses on what is profitable and either partnering or outsourcing the rest. I have assembled a long list of quotes that define what a virtual organization is and why it's different than conventional organizations. The point of looking at these quotes is to demonstrate that none of these models or definitions can adequately be achieved without some heavy analytics and integration of both IT (the wire, the boxes and now the cloud's virtual machines) and IS - Information Systems (Applications) with other stakeholder systems and processes. Up till recently it could be argued that these things can and could be done because we had the technology. But the reality is, unless you were an Amazon, e-Bay or Dell, most firms did not necessarily have the money or the know-how to invest in these types of inovations.

With the proliferation of cloud services, we are finding new and cheaper ways to do things that put these strategies in the reach of more managers and smaller organizations. Everything is game... even the phone system can be handled by the cloud. Ok, I digress, Check out the following quotes and imagine being able to pull these off without analytics.

The next posts will treat some of the tools and technologies that are available to make these business strategies viable.

Multi-Dimensional Portfolio Optimization with @RISK

Jun 28 2012
16
0

Many speak of organizational alignment, but how many tell you how to do it? Others present only the financial aspects of portfolio optimization but abstract from how this enables the organization to meets its business objectives.  We are going to present a practical method that enables organizations to quickly build and optimize a portfolio of initiatives based on multiple quantitative and qualitative dimensions: Revenue Potential, Value of Information, Financial & Operational Viability and Strategic Fit. 
                  
This webinar is going to present these approaches and how they can be combined to improve both tactical and strategic decision making. We will also cover how this approach can dramatically improve organizational focus and overall business performance.

We will discuss these topics as well as present practical models and applications using @RISK.

Reducing Project Costs and Risks with Oracle Primavera Risk Analysis

.It is a well-known fact that many projects fail to meet some or all of their objectives because some risks were either: underestimated, not quantified or unaccounted for. It is the objective of every project manager and risk analysis to ensure that the project that is delivered is the one that was expected. With the right know-how and the right tools, this can easily be achieved on projects of almost any size. We are going to present a quick primer on project risk analysis and how it can positively impact the bottom line. We are also going to show you how Primavera Risk Analysis can quickly identify risks and performance drivers that if managed correctly will enable organizations to meet or exceed project delivery expectations.

.

 

Modeling Time-Series Forecasts with @RISK


Making decisions for the future is becoming harder and harder because of the ever increasing sources and rate of uncertainty that can impact the final outcome of a project or investment. Several tools have proven instrumental in assisting managers and decision makers tackle this: Time Series Forecasting, Judgmental Forecasting and Simulation.  

This webinar is going to present these approaches and how they can be combined to improve both tactical and strategic decision making. We will also cover the role of analytics in the organization and how it has evolved over time to give participants strategies to mobilize analytics talent within the firm.  

We will discuss these topics as well as present practical models and applications using @RISK.

The Need for Speed: A performance comparison of Crystal Ball, ModelRisk, @RISK and Risk Solver


Need for SpeedA detailed comparison of the top Monte-Carlo Simulation Tools for Microsoft Excel

There are very few performance comparisons available when considering the acquisition of an Excel-based Monte Carlo solution. It is with this in mind and a bit of intellectual curiosity that we decided to evaluate Oracle Crystal Ball, Palisade @Risk, Vose ModelRisk and Frontline Risk Solver in terms of speed, accuracy and precision. We ran over 20 individual tests and 64 million trials to prepare comprehensive comparison of the top Monte-Carlo Tools.

 

Excel Simulation Show-Down Part 3: Correlating Distributions

Escel Simulation Showdown Part 3: Correlating DistributionsModeling in Excel or with any other tool for that matter is defined as the visual and/or mathematical representation of a set of relationships. Correlation is about defining the strength of a relationship. Between a model and correlation analysis, we are able to come much closer in replicating the true behavior and potential outcomes of the problem / question we are analyzing. Correlation is the bread and butter of any serious analyst seeking to analyze risk or gain insight into the future.

Given that correlation has such a big impact on the answers and analysis we are conducting, it therefore makes a lot of sense to cover how to apply correlation in the various simulation tools. Correlation is also a key tenement of time series forecasting…but that is another story.

In this article, we are going to build a simple correlated returns model using our usual suspects (Oracle Crystal Ball, Palisade @RISK , Vose ModelRisk and RiskSolver). The objective of the correlated returns model is to take into account the relationship (correlation) of how the selected asset classes move together. Does asset B go up or down when asset A goes up – and by how much? At the end of the day, correlating variables ensures your model will behave correctly and within the realm of the possible.

Copulas Vs. Correlation

Copulas and Rank Order Correlation are two ways to model and/or explain the dependence between 2 or more variables. Historically used in biology and epidemiology, copulas have gained acceptance and prominence in the financial services sector.

In this article we are going to untangle what correlation and copulas are and how they relate to each other. In order to prepare a summary overview, I had to read painfully dry material… but the results is a practical guide to understanding copulas and when you should consider them. I lay no claim to being a stats expert or mathematician… just a risk analysis professional. So my approach to this will be pragmatic. Tools used for the article and demo models are Oracle Crystal Ball 11.1.2.1. and ModelRisk Industrial 4.0

Excel Simulation Show-Down Part 2: Distribution Fitting

 

One of the cool things about professional Monte-Carlo Simulation tools is that they offer the ability to fit data. Fitting enables a modeler to condensate large data sets into representative distributions by estimating the parameters and shape of the data as well as suggest which distributions (using these estimated parameters) replicates the data set best.

Fitting data is a delicate and very math intensive process, especially when you get into larger data sets. As usual, the presence of automation has made us drop our guard on the seriousness of the process and the implications of a poorly executed fitting process/decision. The other consequence of automating distribution fitting is that the importance of sound judgment when validating and selecting fit recommendations (using the Goodness-of-fit statistics) is forsaken for blind trust in the results of a fitting tool.

Now that I have given you the caveat emptor regarding fitting, we are going to see how each tools offers the support for modelers to make the right decisions. For this reason, we have created a series of videos showing comparing how each tool is used to fit historical data to a model / spreadsheet. Our focus will be on :

The goal of this comparison is to see how each tool handles this critical modeling feature.  We have not concerned ourselves with the relative precision of fitting engines because that would lead us down a rabbit hole very quickly – particularly when you want to be empirically fair.

RESEARCH ARTICLES | RISK + CRYSTAL BALL + ANALYTICS