Occam’s Razor
When multiple explanations fit the evidence, prefer the one with the fewest necessary assumptions.
author
William of Ockham
Model type
About
Background
Attributed to the 14th-century philosopher William of Ockham, the razor is a parsimony heuristic. It does not claim the world is simple; it recommends starting with the simplest model that explains the data, then adding complexity only when evidence demands it.
Attributed to the 14th-century philosopher William of Ockham, the razor is a parsimony heuristic. It does not claim the world is simple; it recommends starting with the simplest model that explains the data, then adding complexity only when evidence demands it.
How it works
Competing hypotheses: list explanations that account for the same facts.
Parsimony test: rank by assumptions/complexity (entities, parameters, ad hoc fixes).
Evidence first: only choose the simpler option if predictive adequacy is comparable.
Update: if residuals or new data expose gaps, add warranted complexity.
Parsimony test: rank by assumptions/complexity (entities, parameters, ad hoc fixes).
Evidence first: only choose the simpler option if predictive adequacy is comparable.
Update: if residuals or new data expose gaps, add warranted complexity.
Use cases
Root-cause analysis: avoid multi-factor just-so stories when one mechanism fits observed failures.
Model selection: analytics and forecasting; penalise complexity (AIC/BIC/MDL) to reduce overfitting.
Commercial narratives: diligence storylines and board papers—pick the leanest hypothesis consistent with facts.
Product/design: ship the minimal viable mechanism; defer optionality until validated.
Process and policy: streamline rules—remove steps that lack demonstrable risk reduction.
Model selection: analytics and forecasting; penalise complexity (AIC/BIC/MDL) to reduce overfitting.
Commercial narratives: diligence storylines and board papers—pick the leanest hypothesis consistent with facts.
Product/design: ship the minimal viable mechanism; defer optionality until validated.
Process and policy: streamline rules—remove steps that lack demonstrable risk reduction.
How to apply
Define the question and enumerate plausible hypotheses.
State assumptions for each; note data requirements and implied mechanisms.
Test predictive adequacy on current evidence; reject underperformers.
Prefer the simplest remaining (fewest assumptions/parameters) that passes tests.
Monitor for anomalies; add complexity only to explain persistent residuals.
State assumptions for each; note data requirements and implied mechanisms.
Test predictive adequacy on current evidence; reject underperformers.
Prefer the simplest remaining (fewest assumptions/parameters) that passes tests.
Monitor for anomalies; add complexity only to explain persistent residuals.
pitfalls and cautions
Simplest isn’t always true: prefer evidence over elegance.
Ambiguous “simplicity”: depends on description language; use explicit penalties (e.g. AIC/BIC) where possible.
Premature closure: don’t ignore hidden variables or non-linear interactions when data support them.
Overfitting by subtraction: oversimplified models can miss tail risks and regime changes.
Ambiguous “simplicity”: depends on description language; use explicit penalties (e.g. AIC/BIC) where possible.
Premature closure: don’t ignore hidden variables or non-linear interactions when data support them.
Overfitting by subtraction: oversimplified models can miss tail risks and regime changes.