The Model Thinker
Summary
Page teaches complex systems at Michigan and presents about thirty formal models with their mathematics, their assumptions, what each explains and where each fails. The models include linear regression, normal and power-law distributions, network structure, Markov chains, entropy, random walks, path dependence, cellular automata, game theory in several forms, collective action, mechanism design, signalling, learning models, spatial and hedonic competition, diffusion, and models of contagion and cascades.
The organising argument is that a single model is always wrong and that using many models on the same problem produces better results than refining one. He supports this with the mathematics of ensembles: the diversity prediction theorem shows that a collection's error equals average individual error minus the diversity of the predictions, which means adding a model that is individually worse can improve the collection if it errs differently. He calls this many-model thinking and argues it is what distinguishes people who reason well about complex systems.
The book is a textbook and it requires comfort with mathematics, though Page keeps the treatment accessible and each chapter stands alone. For investors and operators the value is having a formal vocabulary for situations they encounter, particularly the material on power laws, path dependence, network contagion and mechanism design. The limitation is breadth over depth: thirty models in one volume means each gets a sketch, and applying any of them seriously requires further reading. Page is also more confident about the practical value of formal modelling than the record of formal modelling in economics and policy supports.
Key takeaways 7
- Using many models on the same problem produces better reasoning than refining a single model, and the mathematics of ensemble error shows why.
- A collection's prediction error equals average individual error minus the diversity of the individual predictions, so a model that is individually worse can improve the group.
- Distinguishing whether an outcome is normally distributed or power-law distributed changes every decision that follows, and the two are frequently confused.
- Path dependence means that early events determine long-run outcomes, and it is a specific mechanism that can be tested rather than a general claim about history.
- Network structure determines how contagion spreads, so the same infectiousness produces completely different outcomes on different network topologies.
- Models should be judged on what they help a decision-maker understand rather than on whether they predict, since most social models predict badly.
- Formalising an argument reveals assumptions the verbal version conceals, which is the main practical reason to build a model at all.
Who should read it
Investors and analysts who want a formal toolkit of models for complex systems and the argument for applying several to the same problem.
Where it sits in the reference 2
Who recommended it 2
Similar books 6
Chosen by shared topics, the entries that hold them, and the people who recommended both.
