Ventana Modeling Techniques

Ventana chooses its methods and techniques to help clients get maximum value from their experience. Ventana® technology refines client knowledge and applies it to help clients make smarter choices.

Refining client knowledge: triangulation

Ventana refines organizational experience and data by knowledge triangulation:

• Experience: client expertise, collected through interviews and recorded in Reality Check® libraries as expectations about how changes in one factor will affect others
• Data: observations of past activity, collected from text, spreadsheet or database records and recorded as measurements in time
• Model: descriptions of how actions and environment combine to affect performance, created from client input and recorded in documented model equations

Each knowledge source provides valuable insight, yet each has gaps, biases, and other possibilities for imperfection. Ventana triangulates among the three sources by cross-checking to identify discrepancies, finding the source of discrepancies with Vensim's Causal Tracing® technology, and resolving them. Combining the three sources through triangulation produces a fuller, more accurate picture than can be obtained from any one or two sources alone, and Ventana's tools make triangulation fast and efficient.

Cross-checking

Experience and Data
Parallel to model development, Ventana examines available data for consistency with client hypotheses about the reasons behind their historical results. This process often reveals the surprising impact of data collection and handling procedures on client understanding, and informs choices about storing, using and acquiring data. The data analysis also inspires deeper thinking about the true drivers of performance, as well as the role of measurements in organizational self-assessment.

Experience and Model
As the model develops, it is continually checked for consistency with client experience. Clients contribute their expertise through common sense cause and effect statements requiring no knowledge of the model or the mathematics. These statements can be stored in Vensim's Reality Check® format, allowing the model to be tested against the entire Reality Check library with a single click. As the model develops, model output is examined for consistency with expected behavior over a wide range of conditions. If the model behaves contrary to expectations, it is a signal to examine the model structure, the expectations, or both. The Reality Check library can grow over time, deepening and strengthening the experience base against which the model is compared.

Data and Model
A model must be able to reproduce historical outputs when given historical inputs. If during the development of a model, it is unable to reproduce history closely enough, it is a sign that the model, the data, or both must be examined to find the cause of the inconsistency. How close is "closely enough"? To be reliable, a model should contain every important effect, but for simple interpretation it should not contain the numerous effects that make little difference. Therefore, while model results should be close to exact, they will not generally be exact. Likewise, properly interpreted data provide a close approximation to the truth, but numerous factors can prevent data from giving a perfectly accurate picture. Ventana applies statistical hypothesis tests to determine when discrepancies between the model and the data are due to noise in the data, and when they are due to gaps in model structure and understanding.

Causal Tracing®

When cross-checking reveals conflicting or missing information, it signals an opportunity to both deepen understanding and improve the model. Vensim's patented Causal Tracing technology provides automated visualization of the causes of model behavior, making it feasible and efficient to track down the root cause of the discrepancy, even through complex models and data sets.

Resolution

Once the cause of a discrepancy is identified, the resolution is often obvious. When it is not, Ventana works with the client to judge the importance of the discrepancy and to determine the most expedient solution.

Applying Knowledge to Make Smarter Choices

A refined knowledge base is the fuel that powers better decisions. Through the best choices of model architecture, decision analysis techniques, and knowledge management tools, Ventana technology applies the knowledge base to help clients understand what choices are most likely to give maximum performance and why.

Model architecture

The model equations form the bridge from knowledge and data to decisions. Model structure must be chosen to include the client's decision options and measures of success, while remaining consistent with knowledge and data as described above. These requirements dictate the amount of detail and level of focus the model must address. Ventana draws on several analytical architectures, sometimes combining more than one in a single model, to meet these requirements efficiently.

System Dynamics
A widely applicable modeling framework, system dynamics emphasizes cause-and-effect understanding of why things change over time. System dynamics begins by describing the world as a number of stocks, representing the status at any given time of resources (capital, inventories, people, money) and of psychology (morale, satisfaction, awareness). Changes in conditions are represented as flows that increase or deplete each stock over time, calculated from the real world factors which drive them. System dynamics traditionally focuses at an aggregate level, where stocks describe overall or average properties rather than separately tracking each element of a large population.

Agent based modeling
Agent based models represent dozens to millions of individual agents, and separately calculate the status and actions of each one over time. Each agent's actions are determined by rules governing its response to its status, its surroundings, and the actions of other agents with which it is in contact. Overall properties are then observed from the population of individuals. Agent based modeling is a helpful way to understand overall behavior when the complexity of possible interactions among individuals makes an aggregate description difficult.

Single agent modeling
Single agent modeling calculates the status and actions over time of a single actor, independent of the actions of others. Single agent modeling can be viewed as system dynamics applied to a single individual, or as agent based modeling when individuals have little short-term influence on one another's behavior. This approach is most profitable when data are available on the actions of individuals over time.

Statc optimization
Some models require for each period of time a separate optimization specific to that point in time. These situations, such as repeated, short-term resource allocation or pricing decisions, are often amenable to widely used optimization approaches such as linear programming, integer programming, or network programming.

Calibration
Specific model settings that yield the most accurate results are found through automated searches over many possibilities. To avoid relying on trial-and-error, several optimization algorithms exist to steer the search process to the best solutions as quickly as possible while avoiding getting stuck on sub-optimal solutions. Ventana's tests have shown that for many realistic problems, standard algorithms such as the Powell algorithm with multiple starting points outperform alternative techniques such as simulated annealing, genetic algorithms, and simultaneous perturbation stochastic approximation (SPSA). However, these alternatives can be used to advantage in special circumstances.

Decision Analysis

Once a model has been built which is consistent with all available data and experience, a host of techniques can be applied to evaluating potential actions and discovering optimal strategies.

Dynamic optimization
The same automated search procedures used for calibration can be brought to bear on finding the combination of policies that maximize performance. These powerful techniques reveal strategic potential, including new opportunities not yet considered.

Monte Carlo simulation
This common technique explores future possibilities and uncertainty through hundreds or thousands of repeated simulations, representing unknowns as a pool of possible values from which values are drawn at random.

Estimation
Important but hidden factors can be estimated by combining data with an appropriate model. Maximum likelihood estimation is a general approach to this type of statistical inference, including specific procedures such Kalman filtering.

Scenario planning
By understanding potential results--and the reasons for them--in several possible future conditions, it is possible to define actions that hedge against a range of futures, and plan in advance for contingencies.

Real options
The real options framework evaluates investment decisions in uncertain conditions, identifying the actions that will produce the best average likely outcome while accounting for the fact that future information will be used in future decisions.

Knowledge management

The knowledge and data collected for a Ventana model tends to be drawn from all around the client organization, cutting across boundaries to produce a central source for comprehensive understanding. This central repository grows with use, providing an ongoing and improving record of organizational understanding -- and an ever-richer knowledge base, to sharpen ongoing strategy.

Reality Check® archives
Personal knowledge is elicited with expert facilitation in interviews and workshops, and is collected and archived in documented Reality Check libraries.

Documented data collections
Data are absorbed from multiple sources, with the basis and source of each data stream documented and immediately accessible.

Model result archives
Model output is stored together with all relevant inputs for later review, analysis, and comparison to other scenarios. Causal Tracing and other Vensim tools allow fast comparison and visualization of model results and the reasons for differences.

Natural language equation descriptions
The meaning, key assumptions, and source of each model equation is documented in plain language and is immediately accessible as part of the equation definition.

Nearly all of the above techniques are supported directly in the Vensim® modeling environment. The rest are easily connected to Vensim as external functions, or used in parallel to the main model, as expediency demands. Ventana's experience and tools allow for rapid testing and selection of techniques to best meet client needs.