A model is a representation built to preserve selected features of reality for explanation, prediction, comparison, control or communication.
A map, equation, diagram, simulation, economic forecast, scientific theory, machine-learning model, scale model and decision tree are all models. They differ in purpose, structure, assumptions, scale, fidelity, uncertainty and validation.
Quick answer: how should models be categorised?
- Purpose: descriptive, explanatory, predictive, prescriptive, communicative?
- Representation: verbal, mathematical, visual, physical, computational?
- Structure: static, dynamic, deterministic, stochastic, agent-based, networked?
- Scale: microscopic, local, system-wide, global?
- Assumptions: what is simplified, fixed or omitted?
- Fidelity: which features of reality are preserved accurately?
- Uncertainty: where and how is uncertainty represented?
- Validation: how is the model tested?
- Domain: which system or phenomenon does it represent?
- Use: teaching, simulation, forecasting, control, decision support?
This article applies How to Categorise Anything to models as purposeful representations.
1. Separate model from reality
A model is never the thing itself. It preserves some features and omits others.
2. Separate model from data
Data record observations; models organise, transform or explain relationships among observations.
3. Separate model from theory
A theory may explain why a phenomenon occurs, while one or more models instantiate that theory for specific cases.
4. Purpose should be primary
A model designed for explanation may not be ideal for prediction, and a predictive model may not reveal mechanism.
5. Descriptive models summarise structure
They organise what is observed without necessarily explaining cause.
6. Explanatory models represent mechanism
They aim to show how components and relationships produce outcomes.
7. Predictive models estimate future or unseen states
Forecast accuracy and calibration matter more than narrative elegance.
8. Prescriptive models recommend action
Optimisation and decision models connect predictions to choices under constraints.
9. Communicative models support understanding
Teaching diagrams and conceptual maps can be useful even when they omit technical detail.
10. Physical models preserve geometry or mechanism
Scale models, prototypes and mock-ups can reveal spatial or mechanical relationships.
11. Mathematical models express relationships formally
Equations make assumptions and dependencies explicit enough for calculation.
12. Computational models execute rules over time
Simulations can represent dynamic systems too complex for closed-form solutions.
13. Statistical models represent patterns and uncertainty
They estimate relationships from data and often quantify variance or probability.
14. Machine-learning models learn mappings from data
Their internal representation may be less interpretable even when predictive performance is strong.
15. Static models represent one state or equilibrium
They do not explicitly model change through time.
16. Dynamic models represent transitions
They track how states evolve under rules, inputs or feedback.
17. Deterministic models give fixed outputs
Given the same inputs and state, the same output follows.
18. Stochastic models include randomness
Probabilities are part of the model rather than treated solely as noise.
19. Agent-based models represent local actors
System behaviour emerges from interactions among individual agents and rules.
20. Network models emphasise relationships
Nodes and edges preserve connectivity that ordinary tables may hide.
21. Scale defines model resolution
A microscopic model and a population-level model can both be valid while answering different questions.
22. Granularity affects tractability
More detail can improve fidelity but increase computational cost and data requirements.
23. Assumptions should be explicit
Fixed parameters, omitted variables, boundary conditions and idealisations define where a model can be trusted.
24. Simplification is not automatically weakness
A simpler model can be better when the omitted detail does not affect the question.
25. Fidelity is feature-specific
A model can preserve timing well while distorting geography, or preserve averages while missing rare events.
26. Validation should match intended use
Predictive models need out-of-sample testing; mechanistic models may require experiments and qualitative consistency too.
27. Verification and validation differ
Verification asks whether the model was implemented correctly; validation asks whether it represents reality well enough for its purpose.
28. Calibration tunes parameters
Calibration should not be confused with independent validation.
29. Uncertainty can enter through parameters
Input values may be poorly known or variable.
30. Structural uncertainty comes from model form
Different plausible models can represent the same system differently.
31. Scenario models preserve alternative futures
They can be more honest than one precise forecast under deep uncertainty.
32. Models have domains of validity
A model that works under normal conditions may fail under extreme or novel conditions.
33. Model drift matters
Relationships between inputs and outcomes can change as systems, behaviour or policy change.
34. Model versioning is essential
Predictions and decisions should remain traceable to the model and data version used at the time.
35. AI models need task-specific governance
Classification, generation, ranking and forecasting models carry different evaluation and risk profiles.
36. Interpretability is a separate property
A model can be accurate but difficult to explain, or simple and transparent but less predictive.
37. A practical model record
- model ID;
- purpose;
- representation type;
- domain;
- scale;
- structure;
- assumptions;
- inputs;
- outputs;
- uncertainty;
- validation method;
- fidelity limits;
- owner;
- version;
- validity period.
38. Model categories should improve selection
Classification should help users choose the right model for explanation, prediction, teaching or control.
39. Every model is selective
The question is not whether a model omits reality, but whether it omits the wrong parts for the intended job.
40. The deeper idea
Models are disciplined compressions of reality.
To categorise a model well is to know what it is for, what it preserves, what it assumes, how it was tested and where it should stop being trusted.
Final answer
Categorise models by purpose, representation, structure, scale, assumptions, fidelity, uncertainty, validation, domain and use. Separate model from reality, theory and data, and preserve model version and domain of validity so predictions and explanations remain interpretable.