power of many

September 12, 2019

Scott E Page is Leonid Hurwicz Collegiate Professor of Complex Systems, Political Science, and Economics at University Of Michigan, an external professor at The Santa Fe Institute, a Guggenheim Fellow, and a Fellow of American Academy of Arts and Sciences. He is the Author of The Model Thinker and The Diversity Bonus.

It is a long-known fact that models—systematic structures that let us organize information—help us make better sense of things. However, in today’s age of big data, one particular model may not be an all-encompassing one—it may not take into consideration all the data, its potential applications, and estimated outcomes. In this case, a many-model strategy seems to be a more logical and effective approach, as it makes use of the best of multiple models based on what the situation demands.

Where is the wisdom we have lost in knowledge? Where is the knowledge we have lost in information?

– T S Eliot (The Rock)

Eliot wrote before the era of big data. Had he not, he might well have added a third question: where is the information we have lost in data? Amended or not, Eliot’s queries raise a key concern for any manager within, or leader of, an organization. How, when we can gather, scrape, and accumulate so much data and information do we use them to construct the knowledge necessary to make better choices? In other words, how do we become wise with so much information at our fingertips?

Eliot’s categorization (with the inclusion of data) can be written formally as The Wisdom Hierarchy (DIKW). At the bottom sits data—raw, unstructured bits, transactions, and observations. Above that lies information—this is data organized into categories like sales, inflation rates, or quarterly profits. Knowledge consists of causal and correlative relationships among pieces of information. The equation force equals mass times acceleration (f=ma) is knowledge. So is supply = demand. Wisdom, which

rests at the top of the hierarchy, requires combining and engaging a body of relevant knowledge.

According to the diagram, wisdom requires logical coherence (knowledge), empirical grounding (knowledge derived from information and data), and breadth. If knowledge takes model form, then the breadth requirement implies using what I call many model thinking. In many model thinking, an individual, team, or organization, applies a collection of diverse models and constructs a dialogue among them. It offers a useful method for making wise choices.

Let us unpack the parts of the hierarchy below wisdom. Knowledge refers to understandings of causal and correlative relationships. In some cases, knowledge takes the form of scientific laws. More often, we describe knowledge within a model. Models simplify so that we can reason clearly. An economist draws supply and demand curves and in order to infer how shifts occur in either influence price and quantity. An epidemiologist assumes a contact structure among people and a virulence of a disease and from these can predict disease spread. A business strategist may evaluate an industry using Porter’s five factor model.

Next, consider the importance of empirical grounding in data and information. Here is where big data and thick observation enter. We now have access to abundant data which we can use to evaluate, fit and calibrate models. We can also do thicker observations to check if inferences drawn from the big data contains large gaps.

Even with data, the models will make mistakes. That is the cost of simplification. As George Box wrote, “All models are wrong.” By definition, they must be. They are models. They are not reality.

The inclusion of many models obliges us to engage and evaluate multiple logics. In any given situation, some logics will be more wrong than others. Some will be also be more right.

By using many model thinking, we can achieve wisdom. Wisdom emerges in part due to more expansive coverage: If each model has distinct blind spots, many models combined will have fewer. It also happens because the inclusion of many models obliges us to engage and evaluate multiple logics. In any given situation, some logics will be more wrong than others. Some will be also be more right. Sometimes we may consider many models and discover it is best to base our decision on just one or two. Most often though, we will find that a parliament of models works best. As Ray Dalio writes in his book Principles, “in an idea meritocracy, a single CEO is not as good as a great group of leaders.”

As an example of how to apply many-model thinking, imagine a clothing manufacturer deciding on a location for a new production facility. The manufacturer might use a factor rating model that identifies the most important factors (dimensions) and then assigns each potential location a score from one to ten. Factors might include geographic distance to distribution center, production costs, labor costs, taxes, and employee preferences.

Many-model thinking would not stop there. It might also construct a variety of weighted factor models that each assigns different factors and weights on those factors. In other words, it would create many models of the same type. This could be accomplished by having multiple teams choose a set of factors and assign relative weights to each.

If these teams are populated by diverse people—in their social identities, experiences, and professional roles—their models likely identify distinct sets of factors producing non overlapping blind spots. For example, if one team included someone with experience in politics that team might include local political support as a factor. While this factor may be less important than production and labor costs or distance to markets, it could be dispositive. Recently, public dissent caused Amazon to abandon a much publicized decision to locate a second headquarters in New York City. Had they used many-model thinking, they might have anticipated the public reaction and either made a different choice or taken actions to compensate those affected.

Using multiple weighted factor models provides variation within a class of models. More expansive many-model thinking includes multiple types of models. Our clothing manufacturer might include center of gravity models in which the major cities, labor markets, distribution centers, and scores, are each given a value, represented as a weight. The model then finds a location in the center of those weights.

The center of gravity model leaves out relevant variables like production costs, labor costs, access to utilities that underpin the weighted factor models. It places almost complete emphasis on place. If this model produces a similar answer to the factor models, then the manufacturer can place more faith in both models’ choice. If two ways of seeing a problem point to the same solution, that solution is more likely to be correct.

The ensemble of models guarantees a more robust choice. This same logic applies almost every complex business decision. All models are wrong. But many models, used in combination, offer

If the preferred locations in the factor model score well below average in the center of gravity model, then some unpacking of the models is necessary. Why does one family of models prefer location A? And why does the other prefer location B? The logic of models is often transparent. Without too much effort we can unravel the reasoning and understand why the models lead to different choices. It may turn out that the factor models place substantial weight on labor costs and that coastal regions (which are far from the center of gravity) have high unemployment. If so, the center of gravity model has exposed a risk. Labor costs can change over time. Locations stay fixed. Labor costs may be low now but may not in the future.

The factor models and the gravity models can be thought of as “inside the box,” They are standard models for plant location decisions. Full-throated many-model thinking advocates applying less standard models. The manufacturer might, for example, apply a macro economic model that tries to predict the firm locations decisions of other firms—not just the decisions of direct competitors but of all firms. If other firms seek low labor costs (and they likely do), then the key factor for locating on the coast may be eroded. What the center of gravity models revealed as a potential risk may be an almost sure thing.

Or, the manufacturer might construct a real estate valuation model. Here, they might act as if they were a real estate investment trust (REIT) and estimate the expected discounted value of a property in each location. By applying these four types of models as well as variations within each type, the decision makers guarantee a diversity of representations. One model sees the decision as about attributes, another about location, a third about anticipating macro-economic trends, and the fourth about land values.

We can interpret the contributions of many model thinking within the Wisdom Hierarchy. The models vary in the data that they use, the information, that they create, and the knowledge that they produce. The ensemble of models guarantees a more robust choice. This same logic applies almost every complex business decision. All models are wrong. But many models, used in combination, offer the hope of wisdom.