Baseline Comparison¶
Comparative method — instantiates Counterfactual Comparison
Compares actual outcomes with a pre-action baseline, expected trend, benchmark, or no-action projection when direct controls are unavailable.
Baseline Comparison builds the missing counterfactual out of the focal unit's own history rather than out of some other unit. When there is no control group to hold against the treated case, you ask what this unit was on track to do — its pre-action level, its extrapolated trend, an external benchmark it tracks, or a no-action forecast — and you read the gap between that projected path and what actually happened as the estimated effect. Its defining move, and its whole vulnerability, is that the comparison lives entirely inside one time series: the baseline is a projection of the same thing forward, so the argument is only as good as the claim that the unit would truly have followed that line if nothing had been done.
Example¶
A procurement team at a manufacturing firm renegotiates its freight contracts and wants to report the savings. There is no parallel company to compare against, so they use a baseline: last year's freight spend, escalated by the shipping-rate index everyone in the sector faced, projected across the same twelve months. That projected line is the "no-action" world — what the bill would plausibly have been had they left the old contracts alone. Actual spend came in below it, and the difference over the year is the reported saving. The team is careful to align the windows (same seasonal peak, same volume mix) so the two paths are comparing like with like, and to note that part of the apparent drop is just softer shipping rates the index already accounts for — not their negotiation. The output is a saving figure expressed as actual minus projected baseline over the fiscal year, with the escalator assumption stated on its face.
How it works¶
The method is a subtraction between two lines over a shared window. First, pick the baseline form: the last observed level held flat, an extrapolated trend, a benchmark the unit tracks, or a modeled no-action forecast. Second, project that baseline across exactly the period the actual outcome covers, adjusting for anything (seasonality, known external shocks, volume) that would have moved the unit regardless of the action. Third, take the difference. The credibility work is almost entirely in the baseline: a flat "last year" line quietly assumes no underlying trend, while a fitted trend assumes the trend would have continued. Naming which assumption the baseline encodes is the difference between a defensible estimate and a flattering one.
Tuning parameters¶
- Baseline form — flat prior level, extrapolated trend, external benchmark, or modeled forecast. Richer forms capture more of what would have happened anyway but add modeling assumptions that can themselves be wrong.
- Trend model — how the pre-period is extrapolated (linear, seasonal, index-linked). More structure fits history better but risks projecting a pattern that would have broken.
- Alignment window — how much pre-period defines the baseline and how the post-period is matched to it. Too short a window misses slow effects; too long lets unrelated shocks contaminate the projection.
- Normalization — deseasonalizing, volume-adjusting, or indexing to a shared external driver so the two paths differ only by the action.
When it helps, and when it misleads¶
Its strength is availability: every unit has its own past, so a baseline can be drawn when no control group exists at all, which is why finance, operations, and program reporting lean on it constantly. It makes the implicit "compared with what" explicit and cheap.
Its central failure mode is that a single unit's own trajectory is a weak counterfactual. Regression to the mean alone will make a metric measured just after an extreme reading appear to improve, and a naïve before-after credits that reversion to the action.[n1] Secular trends, one-off shocks, and seasonality all masquerade as effect. The classic misuse is baseline cherry-picking — choosing the base year, or the flat-versus-trend form, that makes the action look best. The guarding discipline is to fix the baseline and its assumptions before seeing the result, to state on its face which "would have happened anyway" assumption the baseline encodes, and to explicitly deduct any trend or shock the action did not cause.
How it implements the components¶
actual_path— records the focal unit's realized outcome and the context (period, volumes, conditions) needed to align it to the projection.reference_baseline— the projected pre-action level, trend, benchmark, or no-action forecast is the comparison anchor this mechanism specializes in.time_horizon— because the baseline is an extrapolation, the comparison is inherently a rate over a defined window; the horizon is a first-class choice, not a detail.outcome_comparison— the gap between actual and projected baseline is the estimated difference the action made.
It does not build its baseline by weighting other units (comparison_unit_matching) or validate it with placebo tests (sensitivity_probe) — that construction-and-falsification is Synthetic Control Method, the nearest twin; Baseline Comparison anchors on one unit's own extrapolated trend over a time horizon instead.
Related¶
- Instantiates: Counterfactual Comparison — Baseline Comparison supplies the counterfactual when only the focal unit's own history is available.
- Sibling mechanisms: Matched Case Comparison · Synthetic Control Method · Scenario Contrast · Counterfactual History Review · What-If Analysis · Control Group Comparison · A/B Test
Editorial Notes¶
Form Classification¶
Form family: Analysis, Modeling & Optimization
Rationale: Compares actual outcomes with a pre-action baseline, expected trend, benchmark, or no-action projection when direct controls are unavailable, making its operative form a computation or analytic transformation that produces an inference, comparison, or optimized result.
Independent corroboration: The frozen evidence defines Baseline Comparison as 'Compares actual outcomes with a pre-action baseline, expected trend, benchmark, or no-action projection when direct controls are unavailable', so its operative form is Analysis, Modeling & Optimization.
Review outcome: Independent reviewer agreement; high confidence.
Origin Attribution¶
Primary origin: Statistics & Experimental Design
Origin pattern: Single lineage
Present-day reach: Multi-domain
Rationale: Quasi-experimental and time-series evaluation estimates a counterfactual from a focal unit's pre-action level, trend, or no-action projection.
Related originating lineages:
- Economics & Finance — Econometrics develops interrupted time-series and forecast-counterfactual comparisons.
- Public Administration & Policy — Program evaluation uses baselines where direct control groups are unavailable.
Review resolution: Statistics remains primary through before-after, trend, and quasi-experimental counterfactual construction. Econometrics and program evaluation are material applications of that established single methodological lineage.
Review outcome: Reconciled after independent review; high confidence.
Notes¶
Baseline Comparison is the honest default when nothing stronger is feasible, but it is also the mechanism most easily dressed up as more than it is. A single before-after line is not a controlled study; the moment a comparable untreated unit or a donor pool becomes available, Matched Case Comparison or Synthetic Control Method will give a far more defensible counterfactual, and the baseline should be demoted to a sanity check.
[n1] Regression to the mean — the statistical tendency for an extreme measurement to be followed by one closer to the average, for reasons unrelated to any intervention. Pre/post baselines taken right after an unusually bad (or good) reading are especially prone to crediting the action for a reversion that would have happened anyway. ↩