A model is a simplified representation of a system, process, object or idea that preserves selected relationships so we can describe, explain, predict, compare or design more effectively. If you are asking “what is a model?”, the useful answer ranges from physical scale models and diagrams to mathematical equations, computer simulations, statistical models, conceptual frameworks and machine-learning models. A model is not reality itself. It is a purposeful representation built to answer a question.
A complete explanation of models must include both their power and their limits. Models work by leaving things out. A road map ignores the colour of every building so that routes become visible. A climate model cannot represent every molecule in the atmosphere, so it represents large-scale processes at chosen resolutions. A mathematical model can isolate a few variables to reveal a relationship that would be difficult to see in a fully detailed description. The simplification is not necessarily a defect; it is often the reason the model is useful.
This guide explains models from first principles: representation, abstraction, assumptions, variables, parameters, equations, diagrams, physical models, conceptual models, statistical models, simulations, machine learning, validation, verification, calibration, sensitivity analysis, uncertainty, explanatory models, predictive models, causal models, model selection, overfitting, underfitting and model limits. The goal is a durable mental model of models themselves: know what is represented, what is omitted, what assumptions make the representation work, how the model was tested and which questions it can legitimately answer.
What Is a Model? A Short Definition
A model is a representation constructed for a purpose. It may represent the structure of a bridge, the spread of a disease, the behaviour of an economy, the orbit of a planet, the flow of traffic or the relationship between study time and performance.
Models can be concrete or abstract. A globe is a physical model of Earth. An equation describing projectile motion is a mathematical model. A causal diagram is a conceptual model. A neural network used for image classification is a computational model.
The key idea is selection. A model preserves some features of a target system and ignores others according to the problem being solved.
Why Models Matter
Reality is often too large, small, slow, fast, expensive, dangerous or complicated to manipulate directly. Models create a manageable representation.
Engineers test bridge designs before building full-scale structures. Epidemiologists model disease transmission. Astronomers model stellar evolution across timescales impossible to observe directly. Teachers use number lines and diagrams to model mathematical relationships.
Models also allow counterfactual questions: what would happen if one parameter changed? These questions are central to science, engineering and decision-making.
Models Are Not Reality
A model can resemble its target closely and still remain a model. The distinction matters because every model contains idealisations, boundaries and approximations.
A globe preserves relative geography but not the exact texture of every landscape. A financial model may represent cash flows while ignoring organisational culture. A physics equation may assume no air resistance.
Confusing the model with reality creates model risk: conclusions may be applied beyond the conditions under which the representation is accurate enough.
Representation
Representation is the relationship between a model and the target it stands for.
Some models resemble their targets visually. Others represent through mathematical structure. A graph of predator and prey populations does not look like animals, but its variables and equations represent relationships among population sizes.
Representation depends partly on purpose. The same system may require different models for different questions.
Abstraction
Abstraction removes detail to expose structure.
A subway map distorts geographic distance so that station order and interchange relationships become clear. A free-body diagram removes colour, material texture and background scenery so that forces become visible.
Good abstraction removes irrelevant detail without removing the relationships needed for the task.
Idealisation
Idealisation deliberately represents a system as simpler or more regular than it is.
Physics often models surfaces as frictionless, gases as ideal or objects as point masses. These systems do not exist exactly in nature, yet the idealisation can reveal dominant relationships.
The value of an idealisation depends on whether neglected effects are small enough for the intended question.
Assumptions
Assumptions specify conditions under which a model operates.
A population model may assume a constant birth rate. A financial model may assume interest rates remain within a range. A classroom model may assume students have prerequisite knowledge.
Good modelling makes important assumptions explicit because hidden assumptions are often where transfer fails.
Variables
Variables are quantities or categories that can change within a model.
A motion model may include time, velocity and position. A business model may include price, demand and cost.
Choosing variables is already an act of theory because the modeller decides which features are relevant enough to represent.
Parameters
Parameters are quantities that shape model behaviour and are often treated as fixed within one run or scenario.
A population-growth rate, friction coefficient or learning-rate parameter changes the model’s dynamics.
Parameters may come from direct measurement, fitting to data, expert judgment or assumptions.
State Variables
State variables describe the current condition of a dynamic system sufficiently for the model to predict how it evolves under its rules.
A simple pendulum model might use angle and angular velocity. A reservoir model might use current water volume.
State representation is important because two systems with the same observed output may behave differently if their hidden state differs.
Physical Models
Physical models are tangible representations such as scale models, prototypes and mock-ups.
Architects use building models, aerospace engineers use wind-tunnel models and product designers use prototypes.
Scaling is not trivial. Forces, flows and material behaviour may not scale linearly, so physical models require dimensional reasoning.
Conceptual Models
Conceptual models organise ideas and relationships without necessarily using equations.
A water-cycle diagram, food web or causal-loop diagram is a conceptual model.
These models are powerful for teaching and planning because they make structure visible before numerical detail is added.
Mathematical Models
Mathematical models express relationships using equations, functions and probability distributions.
They can describe motion, growth, heat transfer, population change, risk and many other processes.
Mathematics makes assumptions precise and allows logical consequences to be derived systematically.
Statistical Models
Statistical models represent uncertainty and relationships in data.
Linear regression models an expected relationship between variables. Probability distributions represent variability. Hierarchical models represent structure across groups.
A statistical model does not automatically identify causation. It describes a probabilistic relationship under assumptions.
Computational Models
Computational models implement rules or equations in software so complex systems can be simulated numerically.
Weather forecasting, fluid dynamics, astrophysics and engineering all rely heavily on computation.
Computer output is not inherently more trustworthy than hand calculation. The code can correctly solve the wrong model or incorrectly solve the intended one.
Simulations
A simulation evolves a model through time or repeated scenarios to explore behaviour.
A traffic simulation can represent individual vehicles, road rules and signals. An epidemic simulation can represent contacts and transmission.
Simulation is especially useful when analytical solutions are unavailable, but it inherits every assumption built into the model.
Agent-Based Models
Agent-based models represent individual agents with rules for behaviour and interaction.
Large-scale patterns can emerge from many local interactions. Examples include traffic, markets, ecosystems and social behaviour.
Agent-based models are valuable for studying heterogeneity and networks but can contain many parameters and assumptions that are difficult to identify uniquely from data.
Mechanistic Models
Mechanistic models represent processes believed to generate an outcome.
A biochemical model may represent reactions, enzyme activity and concentrations. A mechanical model may represent forces and material properties.
Mechanistic models can support intervention because they explain how changing one component should affect the system.
Phenomenological Models
Phenomenological models describe patterns without fully representing the underlying mechanism.
A curve may fit observed data extremely well even when the deeper process remains unknown.
These models can be useful for prediction but may transfer poorly when conditions change because the correlation is not mechanistically grounded.
Toy Models
Toy models deliberately simplify a problem to a small number of components.
Their purpose is often conceptual clarity rather than realistic prediction.
A toy model can expose a mechanism, paradox or threshold that becomes hidden in a high-dimensional simulation.
Scale Models
Scale models preserve selected geometric relationships while changing size.
They are useful in architecture, engineering and education.
Not every physical process scales with geometry alone. Fluid flow, heat transfer and structural loads may require dimensionless numbers or special materials to preserve similarity.
Model Inputs and Outputs
Inputs are variables, parameters or boundary conditions supplied to the model. Outputs are quantities the model calculates or predicts.
An output is only as meaningful as its input quality and model structure.
Garbage in, garbage out is incomplete because even excellent inputs can produce poor outputs if the model itself is wrong.
Boundary Conditions
Boundary conditions define what happens at the edges of the modelled system.
A heat-transfer model may fix temperature at a surface. A fluid model may specify inflow speed. An economic model may define which markets are included.
Changing boundaries can change the answer dramatically, so they are part of the model rather than background trivia.
Initial Conditions
Initial conditions specify the starting state of a dynamic model.
Some systems are highly sensitive to small differences in initial conditions, producing diverging trajectories over time.
This is one reason long-range prediction can become difficult even when the governing equations are known.
Calibration
Calibration adjusts model parameters so model output agrees sufficiently with observed data for an intended purpose.
A calibrated model is not automatically validated. Many parameter combinations can sometimes fit the same historical data.
Calibration should therefore be separated from independent testing.
Verification
Verification asks whether the model or computer implementation correctly solves the equations or rules it was intended to solve.
Code verification can test whether algorithms are implemented correctly. Solution verification estimates numerical error.
A verified simulation can still represent reality poorly if the underlying model is inappropriate.
Validation
Validation asks whether the model represents the real target system well enough for its intended use.
Validation compares model outputs with independent observations, experiments or benchmark cases.
Validation is purpose-specific. A model adequate for estimating average temperature may be inadequate for predicting local extremes.
Verification Versus Validation
Verification asks, “Did we solve the model right?” Validation asks, “Did we choose and represent the right model for the purpose?”
The distinction is essential in computational science and engineering.
A program can perfectly implement equations that omit the mechanism driving the real system.
Uncertainty Quantification
Uncertainty quantification studies how uncertainty in inputs, parameters and model structure affects outputs.
Instead of one number, a model may produce ranges, distributions or sensitivity intervals.
Representing uncertainty honestly is often more useful than adding decimal places to a fragile point estimate.
Sensitivity Analysis
Sensitivity analysis asks how much outputs change when assumptions, inputs or parameters change.
If a tiny change in one uncertain parameter reverses the conclusion, the model is fragile around that decision.
Sensitivity analysis helps identify which measurements or assumptions deserve the most attention.
Scenario Analysis
Scenario analysis compares outcomes under different plausible sets of assumptions.
It is useful when probabilities are difficult to assign but several futures need to be explored.
Scenarios are not predictions. They are conditional stories: if these assumptions hold, this is how the model behaves.
Model Identifiability
Identifiability asks whether available data can uniquely determine the parameters or structure of a model.
Two different parameter combinations may generate almost identical outputs. In that case, historical fit cannot tell us which internal explanation is correct.
Identifiability matters because apparently precise parameter estimates can be misleading when the data do not contain enough information to distinguish alternatives.
Residuals
Residuals are differences between observed values and model predictions.
Residual patterns reveal model failure. If errors increase with time, curve systematically or cluster in one subgroup, the model may be missing structure rather than experiencing random noise.
Residual analysis is therefore more informative than one overall accuracy number because it shows where and how the model is wrong.
Extrapolation
Extrapolation uses a model outside the range of data or conditions on which it was developed.
A relationship that is approximately linear between 10 and 30 degrees may not remain linear at 200 degrees. A demand model learned in stable conditions may fail during a crisis.
Extrapolation should be treated as a stronger claim than interpolation because the model is being asked to operate where evidence is thinner.
Benchmarking
Benchmarking compares a model against simple baselines, established methods or standard test cases.
A complex model is valuable only if it improves something that matters: predictive accuracy, computational speed, explanatory insight or robustness.
Benchmarks prevent sophistication from being mistaken for progress. If a simple rule performs equally well, the extra complexity needs another justification.
Model Error Decomposition
Model error can come from several places: noisy observations, uncertain input values, numerical approximation, wrong parameter values, missing mechanisms or an unsuitable model form.
Separating these sources matters because each requires a different repair. Better sensors do not fix a missing mechanism, and a more accurate solver does not fix a biased dataset.
A useful modelling workflow therefore asks not merely how large the error is, but what kind of error it is and whether it changes the decision the model supports.
Model Stress Testing
Stress testing deliberately exposes a model to extreme, unusual or boundary conditions to discover where its assumptions fail.
A financial model may be tested under sharp interest-rate changes. An engineering model may be tested near material limits. An AI system may be tested on rare subgroups or adversarial inputs.
Stress tests do not prove universal safety, but they reveal fragility that average-case validation can miss and help define the model’s safe operating envelope.
Model Fit
Model fit measures how well a model reproduces observed data according to a chosen criterion.
Good fit can indicate useful structure, but it can also result from overfitting noise.
Fit should therefore be evaluated on data not used to estimate the model whenever possible.
Overfitting
Overfitting occurs when a model learns noise or accidental detail in training data rather than patterns that generalise.
A highly flexible curve can pass through every observed point yet predict new points poorly.
Regularisation, simpler models and independent testing help reduce overfitting.
Underfitting
Underfitting occurs when a model is too simple to capture important structure.
A straight line cannot represent a strongly curved relationship adequately.
The goal is not maximum complexity or minimum complexity, but enough structure to capture relevant patterns without fitting noise.
Bias–Variance Trade-Off
Simple models can have high bias because they miss structure. Flexible models can have high variance because they respond strongly to limited training data.
The bias–variance trade-off describes this tension in predictive modelling.
The best balance depends on data volume, noise and the cost of different errors.
Model Selection
Model selection chooses among candidate models.
Criteria may include predictive performance, simplicity, interpretability, computational cost, causal plausibility and robustness.
No single criterion is always best. A scientific explanation may favour a mechanistic model while an operational forecast may favour predictive accuracy.
Parsimony
Parsimony means avoiding unnecessary complexity.
If two models explain the same evidence equally well, the simpler model may be easier to test, interpret and generalise.
But parsimony is not permission to ignore real complexity. Evidence should determine how much structure the model needs.
Interpretability
Interpretability concerns how easily people can understand the relationships represented by a model.
A small decision tree or linear model may be easy to inspect. A large neural network can be difficult to explain at the level of individual predictions.
Interpretability matters more in some applications, especially when users need reasons, accountability or debugging.
Explanatory Models
Explanatory models aim to reveal why a phenomenon occurs.
They often represent mechanisms, causal relationships or unifying principles.
A model can explain well without maximising point prediction if it captures the structure that makes the phenomenon intelligible.
Predictive Models
Predictive models aim to estimate unknown or future outcomes accurately.
They may use relationships that are stable enough for forecasting even when causal mechanisms are incomplete.
Prediction should be tested prospectively or on held-out data, not judged only by historical fit.
Causal Models
Causal models represent how interventions are expected to change outcomes.
They differ from ordinary predictive models because correlation is not enough: the model must specify causal direction and confounding assumptions.
Causal graphs, structural equations and potential-outcome frameworks are common approaches.
Decision Models
Decision models connect uncertain outcomes with actions, costs and values.
They may combine probabilities, utilities and constraints to compare choices.
A decision model can be mathematically correct while still producing a poor decision if the objectives or values were defined badly.
Machine-Learning Models
Machine-learning models estimate patterns from data rather than relying only on manually specified rules.
Linear models, decision trees, support-vector machines, neural networks and transformers are all model families.
Training changes model parameters to reduce an objective function on examples. The resulting system remains a model: a learned representation that approximates relationships relevant to a task.
Neural Networks
Neural networks combine layers of weighted transformations and nonlinear functions.
They can represent complex patterns in images, language, audio and structured data.
Their flexibility makes them powerful but also increases requirements for data, evaluation, robustness testing and interpretability.
Large Language Models
Large language models estimate patterns in sequences of tokens using large neural networks trained on extensive text and related data.
They can generate fluent explanations, code and dialogue because they model statistical and semantic relationships at scale.
A language model is not a database of verified facts. Its outputs remain generated predictions conditioned on training and current context.
Foundation Models
Foundation models are large models trained broadly and adapted to many downstream tasks.
Their reuse changes the modelling workflow because one pretrained representation becomes a starting point for many applications.
This creates efficiency but also inherited risk: limitations or biases in the base model can propagate into many systems.
Model Drift
Model drift occurs when the relationship between inputs and outcomes changes after deployment.
A fraud detector trained on last year’s behaviour may weaken when attackers adapt. A demand model may fail after a market shock.
Models therefore need monitoring and revalidation rather than one-time approval.
Distribution Shift
Distribution shift occurs when the data encountered in use differ from the data on which the model was developed.
Changes in population, measurement, policy or environment can create shift.
Robust models should be tested across relevant subgroups and plausible future conditions, not only on an average historical sample.
Model Risk
Model risk is the possibility of loss or error caused by an inappropriate model, incorrect implementation or misuse.
Complex models can create false confidence because outputs look quantitative and precise.
Governance should therefore document purpose, assumptions, validation evidence, limitations and conditions that trigger review.
Model Governance
Model governance defines who builds, reviews, approves, monitors and retires models.
Independent validation can reveal assumptions overlooked by the development team.
Version control and change logs ensure decisions can be traced to the model actually used at the time.
Model Documentation
Good documentation states the target system, intended use, inputs, outputs, assumptions, equations or architecture, training data where relevant, evaluation methods and known limitations.
Documentation supports reproducibility and safe handover.
A model that cannot be explained operationally to its next maintainer is difficult to govern responsibly.
Model Updating
Models should be updated when new data reveal systematic error, when the environment changes or when the intended use expands.
Updating is not always retraining. It may involve changing parameters, structure, inputs or the decision threshold.
Every update can introduce new behaviour, so changes should be revalidated.
Model Retirement
A model should be retired when its purpose disappears, its assumptions no longer match the environment or a better replacement becomes available.
Retirement is different from simply stopping use. Systems should identify downstream dependencies, archive documentation and preserve the ability to reconstruct past decisions.
Good model governance therefore covers the entire lifecycle: commission, build, validate, deploy, monitor, update and retire.
Model Ensembles
An ensemble combines multiple models.
Averaging or voting can improve predictive performance because different models make different errors.
Ensembles can also reduce interpretability and may fail together if all models share the same flawed assumptions or data.
Model Comparison
Comparing models is more informative than evaluating one in isolation.
A complicated model should outperform simple baselines meaningfully enough to justify extra complexity.
Baseline models reveal whether sophistication actually adds value or merely adds computation.
Null Models
A null model represents what would be expected without the mechanism of interest.
In ecology, networks and statistics, null models help identify whether observed structure exceeds what simple randomness or constraints would produce.
They are powerful because explanation requires contrast: what pattern would exist if the proposed mechanism were absent?
Surrogate Models
A surrogate model approximates a more expensive simulation or process.
Engineers may train a fast statistical model on outputs from a slow physical simulation to support optimisation.
Surrogates require validation within the region where they will be used because extrapolation can fail sharply.
Digital Twins
A digital twin is a computational representation linked to data from a physical asset or system and updated over time.
Digital twins are used in manufacturing, infrastructure and operations to monitor state, simulate scenarios and support maintenance.
The label should not imply perfect duplication. A twin remains a model with selected variables, sensors and assumptions.
Models and Decisions
Models become consequential when outputs influence actions. A risk model may change who receives extra review; an engineering model may determine a safety margin; a forecast may trigger inventory decisions.
Decision use changes the standard of evidence. An approximate model that is harmless for classroom illustration may be unacceptable when a small error can cause large financial or safety consequences.
The correct question is therefore not only “Is the model accurate?” but “Is it accurate enough, in the right places, for this decision and its costs of error?”
Models in Education
Education uses models to represent learning progressions, assessment constructs, student misconceptions and system behaviour.
A model of a learner should not be mistaken for the learner. Scores and classifications summarise selected evidence under conditions.
Good educational models remain revisable and preserve uncertainty rather than turning one result into a permanent identity.
Models in Science
Scientific models help connect theory with observable phenomena.
Scientists build, test, compare and revise models as new evidence appears.
A model can be useful even when known to be literally false in some respects, provided its idealisations are appropriate for the problem.
Models in Engineering
Engineering models support design before construction.
They represent loads, material behaviour, control systems, heat transfer, reliability and other properties.
Safety-critical modelling requires conservative assumptions, validation and explicit margins because errors can have physical consequences.
Models in Economics
Economic models simplify households, firms, markets and institutions to study relationships among incentives, prices, production and policy.
Their assumptions are often strong because human systems are heterogeneous and adaptive.
Economic models are most useful when assumptions and scope are stated clearly rather than presented as universal laws.
Models in Climate Science
Climate models represent atmosphere, oceans, land, ice and energy flows using physical equations and parameterisations.
They are evaluated against observations, historical climate and physical constraints.
Uncertainty comes from future emissions, internal variability, parameter values and structural model differences rather than one single source.
Models in Medicine
Medical models estimate diagnosis, prognosis, treatment effects and physiological processes.
Clinical use requires validation in the relevant population because disease prevalence and measurement practices vary.
Model output supports, rather than replaces, clinical judgment and patient context.
Common Misconceptions About Models
- “A model is a miniature copy of reality.” Many models are abstract and intentionally selective.
- “More detail always makes a model better.” Extra detail can increase uncertainty and obscure the relationship of interest.
- “If a model fits past data, it is correct.” Overfitting and alternative structures can produce similar historical fit.
- “Verification proves the model is realistic.” Verification checks implementation; validation checks representation for a purpose.
- “A model prediction is a fact.” Predictions are conditional on assumptions and inputs.
- “Machine-learning models discover causation automatically.” Predictive learning does not remove the need for causal assumptions.
- “One model should answer every question.” Different purposes often require different representations.
- “Model uncertainty can be eliminated.” Uncertainty can be reduced and quantified but rarely disappears completely.
A Practical Model Checklist
Ask what target system the model represents and what decision or question it was built for.
Identify variables, parameters, boundary conditions, initial conditions and important assumptions. Ask what the model leaves out.
Then inspect evidence: how was it calibrated, verified and validated? On which data, populations or regimes has it been tested? What conditions would make it unreliable?
How to Learn Models Properly
Start by building a simple model of a familiar system. Use a diagram or equation to represent only the relationships needed for one question.
Then compare predictions with observations. Record where the model works and where it fails.
Finally, revise the model or build an alternative. The deepest lesson is that modelling is iterative: represent, test, diagnose, revise and retest.
Frequently Asked Questions About Models
What is a model in science?
A scientific model is a representation of a target system used to describe, explain, predict or investigate phenomena.
Why do models simplify reality?
Simplification removes irrelevant detail so important relationships can be represented and analysed.
What is validation?
Validation tests whether a model represents the real system adequately for its intended use.
What is verification?
Verification checks whether equations, algorithms or code correctly implement and solve the intended model.
Can a model be useful if it is not perfectly true?
Yes. Most models are idealised. They can be useful when the omitted details do not materially affect the question being asked.
What is overfitting?
Overfitting occurs when a model learns noise or accidental detail in development data and performs poorly on new cases.
What is a machine-learning model?
It is a model whose parameters are learned from data to perform tasks such as prediction, classification or generation.
When should a model be updated?
When data, environment, measurement systems or intended use change enough that previous validation no longer supports current decisions.
Authoritative Starting Points
For a rigorous overview of scientific modelling, see the Stanford Encyclopedia of Philosophy entry on Models in Science and its related entry on Scientific Representation. For computational modelling, the National Academies report on verification, validation and uncertainty quantification provides a clear distinction between solving a model correctly and validating its representation of a physical system.
What Is a Model? The Complete Idea
A model is a purposeful representation. It compresses reality into a form that makes selected relationships visible, calculable or testable.
The strongest understanding keeps several distinctions clear: a model is not reality; simplicity is not automatically weakness; fit is not proof; verification is not validation; prediction is not causation; and complexity does not guarantee accuracy.
Once those distinctions are stable, modelling becomes a disciplined way to think. Define the target, choose the purpose, state the assumptions, represent only what matters, test against evidence, quantify uncertainty and revise when the model fails. The value of a model is not that it contains everything. It is that it contains enough of the right structure to help us see what reality is doing.