Simple Baseline Model¶
Artifact — instantiates Parsimony Filter
Provides a low-complexity model or design that more complex candidates must outperform or justify exceeding.
A Simple Baseline Model is a deliberately low-complexity model, forecast, or design, kept on hand as the bar every fancier candidate must clear. Its whole purpose is to convert "the complex version is better" from an assumption into a claim that has to be demonstrated: a concrete reference, scored on the same data and the same metric, that a sophisticated alternative must beat by enough to justify its added burden. Its defining trait is that it is a persistent artifact — a thing that sits in the pipeline as the number to beat — not an activity that chooses or removes anything. It neither ranks rivals nor cuts parts; it simply exists as the floor, and its existence is what makes every complexity claim answerable.
Example¶
A regional bakery chain asks its data team to build a machine-learning demand forecaster so each shop bakes closer to what it will sell. Before touching a model, the team stands up a baseline: tomorrow's demand for an item equals what that shop sold on the same weekday last week, nudged by a simple trend factor. It is almost embarrassingly plain, and it scores a mean absolute percentage error of 18% on held-out weeks.
Then comes the gradient-boosted model with weather, promotions, local events, and holiday calendars. It scores 16.5%. Without the baseline, 16.5% sounds like success and the model ships. With the baseline in place, the real question surfaces: is a 1.5-point gain worth a pipeline that needs weather feeds, retraining, and an on-call engineer when it breaks at 5 a.m.? The baseline reframes every candidate's score as a lift over the floor, and it encodes the adequacy the shops actually need — a forecast good enough to set the morning bake within tolerance. The baseline already clears that bar; the complex model must now show its extra machinery buys more than it costs.
How it works¶
- Build the dumbest defensible reference. Reach for the naive move the domain already trusts — last value, seasonal repeat, group mean, majority class — nothing that needs tuning or explanation.
- Score it on the candidates' terms. Evaluate it on the same held-out data and the same metric every alternative will be judged by, so comparisons are apples to apples.
- Publish it as the floor. Post the baseline's score where candidates are reviewed, and require each alternative to report its performance relative to the baseline rather than in the absolute.
- Refresh it as reality moves. Re-derive the baseline when the data distribution shifts, so it stays the best simple reference rather than a stale strawman.
Tuning parameters¶
- Baseline sophistication — how naive the reference is: pure persistence, a seasonal average, a shallow rule. A too-dumb baseline is trivially beaten and proves nothing; a too-clever one stops being a floor and becomes a rival.
- Scoring metric — the yardstick the baseline and all candidates share. It silently defines what "better" means, so a baseline strong on one metric can look weak on another.
- Refresh cadence — how often the baseline is rebuilt against current data. Frequent refresh keeps the floor honest but adds upkeep; rare refresh risks anchoring on a reference the world has left behind.
- Adequacy floor — the performance level the baseline is meant to embody as "good enough." Set it where the downstream decision actually needs it, not where the baseline happens to land.
When it helps, and when it misleads¶
Its strength is killing the reflex that complexity equals quality: a cheap, interpretable, always-available floor means no sophisticated model gets credit for beating nothing.[n1] It is standard practice in forecasting, machine learning, and operations precisely because a surprising number of elaborate models fail to beat it once validation and upkeep are counted.
Its central failure mode is baseline anchoring: a baseline that everyone trusts stays in place after evidence shows it is no longer adequate, quietly blocking the added complexity a changed problem now genuinely needs. The classic misuse is the rigged floor — choosing a weak, strawman baseline so a favored model clears it easily and the "lift" looks impressive. The guarding discipline is to make the baseline the best simple reference, not the worst one, and to re-derive it as the data shifts, so it keeps testing candidates rather than flattering them.
How it implements the components¶
simple_baseline— it is one: the concrete low-complexity reference, scored and posted, that every candidate must outperform to justify its added machinery.adequacy_criterion— its score operationalizes "good enough" as an actual number on the shared metric, so adequacy is a measured floor rather than an opinion.
It neither counts candidates' assumptions (assumption_list) nor sets the rule for when a richer model is worth reinstating (complexity_reintroduction_threshold) — that comparative and threshold work is Occam-Style Model Selection, which consumes this baseline as one of its candidates. And it removes nothing: it runs no necessity_test and keeps no removal_log of stripped parts — that surgical removal is Feature Pruning.
Related¶
- Instantiates: Parsimony Filter — the Simple Baseline Model is the filter cast as a standing artifact: the floor complexity must earn its way above.
- Sibling mechanisms: Feature Pruning · Occam-Style Model Selection · Lean Design Review · Scope Reduction Review · Assumption Audit · Minimum Viable Explanation · Model Complexity Penalty
Editorial Notes¶
Form Classification¶
Form family: Analysis, Modeling & Optimization
Rationale: Simple Baseline Model operates as an analytical, modeling, inference, comparison, or optimization procedure that derives insight or a solution because it provides a low-complexity model or design that more complex candidates must outperform or justify exceeding.
Independent corroboration: The frozen evidence defines Simple Baseline Model as 'Provides a low-complexity model or design that more complex candidates must outperform or justify exceeding', so its operative form is Analysis, Modeling & Optimization.
Nearest alternative: Representation, Specification & Plan — Simple Baseline Model includes features of a static representation, map, specification, schema, or prospective plan that externalizes information, but its defining operation is an analytical, modeling, inference, comparison, or optimization procedure that derives insight or a solution.
Review outcome: Independent reviewer agreement; medium confidence.
Origin Attribution¶
Primary origin: Statistics & Experimental Design
Origin pattern: Convergent development
Present-day reach: Universal
Rationale: Requiring complex models to beat a low-complexity comparator is statistical benchmark and model-selection discipline.
Related originating lineages:
- Data Science & Analytics — Naive baselines reveal whether machine-learning complexity adds predictive value.
- Engineering & Design — Reference designs provide a performance and cost floor for alternatives.
- Mathematics — Mathematical modeling, proof, and abstract-structure practice supplies a parallel or contributing lineage for the mechanism's defining operation: provides a low-complexity model or design that more complex candidates must outperform or justify exceeding.
- Organizational & Management Science — Organizational design, management, and operational governance supplies a parallel or contributing lineage for the mechanism's defining operation: provides a low-complexity model or design that more complex candidates must outperform or justify exceeding.
- Philosophy — Parsimony requires added complexity to earn explanatory value.
Review resolution: The blind reviewers agree that statistics_experimental_design is the primary origin and differ only on alternate origin disagreement, origin mode disagreement, encyclopedia synthesis disagreement. I preserve every independently explained alternate from both records rather than imposing a numeric cap. I retain convergent because the combined evidence shows independent disciplinary development. The broader reach of universal records portability separately from historical provenance; encyclopedia_synthesis=true preserves the affirmative synthesis judgment where either reviewer identified one.
Encyclopedia synthesis: The exact catalogued form synthesizes established practice rather than reproducing a single standard historical label.
Review outcome: Reconciled after independent review; high confidence.
Notes¶
The baseline is a reference, not a decision. It reports whether a candidate cleared the floor and by how much; it does not itself decide whether that margin is worth the cost. That judgment — the required lift, the add-back rule — belongs to the selection mechanisms (Occam-Style Model Selection) or the in-loop Model Complexity Penalty that consume it. Keeping the artifact separate from the decision is what lets a team improve the baseline without re-litigating every model choice at once.
[n1] A naïve or persistence forecast predicts that the next value equals the most recent (or same-season) observation. It is the canonical baseline in forecasting, and a model that cannot beat it has demonstrated no real skill — the reason a baseline is stood up before any complex candidate is trusted. ↩