Experience Curve Model¶
Quantitative model — instantiates Experience Curve Cost Reduction
Fits the power-law between cumulative volume and unit cost into a single learning rate and a forecast — and flags when the curve is flattening and extrapolation should stop.
Once the decline is known to be real, someone still has to say how steep it is and how far it goes. Experience Curve Model is the quantitative fit that answers both. It regresses unit cost on cumulative production volume as a power law — the shape that gives the experience curve its signature "cost falls a roughly constant percentage each time cumulative output doubles" — and reads two things off it: a learning rate (the decline per doubling) that compresses the whole history into one number, and a forecast of where unit cost lands at future volumes. Its defining discipline is the second job: it watches its own residuals for the curve going flat and marks the point past which the fit no longer holds, so the forecast comes with a built-in stop sign rather than an open-ended promise.
Example¶
A battery manufacturer wants to know whether its pack cost will reach a price target if it keeps scaling. It plots historical cost per kilowatt-hour against cumulative GWh shipped on log-log axes, and a straight line emerges — the hallmark of an experience curve. The slope implies a learning rate of roughly 18% per doubling: each time cumulative output doubles, cost per kWh falls about that much. Extending the fit, the model projects cost dropping from ≈$180/kWh today toward the ≈$120 target after the next two doublings of volume.
But the model does not simply run the line to zero. It imposes a cost floor near the raw material and cell-chemistry limit — say ≈$70/kWh — and watches whether recent points are already bending above the fitted line. When the last few quarters start landing above the curve, the plateau flag trips: the easy learning is being exhausted, and the forecast beyond that point is flagged as unreliable. What the team carries forward is not "cost will keep halving" but "cost falls at ≈18% per doubling until roughly here, then a redesign is needed to reach a new curve."
How it works¶
- Choose the experience axis. Cost is regressed on cumulative volume (units, hours, or output to date), never on calendar time — time is what confounds learning with everything else.
- Fit the power law. On log-log axes the relationship is a straight line; its slope is the learning rate (percent decline per doubling), the single parameter that summarizes the curve.
- Forecast within a horizon. Project unit cost at future cumulative volumes, but only as far as the fit is trustworthy, and against an imposed floor so the power law can't predict impossible costs.
- Monitor for saturation. Track residuals for the curve flattening; when recent points drift above the line, mark the plateau and stop extrapolating past it.
Tuning parameters¶
- Learning-rate source — fit the slope from your own data or borrow an industry rate. Fitting is honest but needs history; borrowing lets you model a new line but imports someone else's conditions.
- Experience-axis basis — whether a "doubling" is counted in units, labor-hours, or output value; the choice changes the rate and what a forecast even means.
- Extrapolation horizon — how far past observed data to project. Further reach is more useful and more dangerous, because plateau risk grows with distance.
- Cost floor / asymptote — whether a physical or material lower bound is imposed so the curve bends toward a limit instead of toward zero.
- Fit window — which slice of history is fitted. Fitting only the early, steep ramp overstates the rate; including the mature tail tempers it.
When it helps, and when it misleads¶
Its strength is turning a scatter of cost history into a decision-grade rate and forecast, and — unlike a naïve trend line — building in the knowledge that learning curves end. The plateau monitor is what keeps it from over-promising, and the cumulative-volume axis is what keeps it from crediting learning for what is really economies of scale or a cheaper input this quarter.
Its failure modes cluster around extrapolation. Wright's law is an empirical regularity, not a physical guarantee — it describes many production histories but promises nothing, so projecting it past a technology's limits or fitting only the steep early segment produces confident nonsense.[n1] Input-price swings and scale economies can inflate the apparent learning rate, and the tidy single number invites false precision. The classic misuse is running it backwards: fitting the flattering early slope and extrapolating it to justify an aggressive cost commitment or bid already decided on. The discipline that keeps it honest is to fit on mature data, impose a floor, carry the plateau flag into every forecast, and re-fit as new volume lands rather than freezing the first line.
How it implements the components¶
Experience Curve Model fills the curve-fitting-and-forecast components of the archetype — the ones a quantitative model produces:
unit_cost_learning_curve— its core output: the fitted power-law relating unit cost to cumulative volume, reduced to a learning rate and a projection.saturation_and_plateau_monitor— the model watches its own residuals for the curve flattening and marks the point past which extrapolation is no longer trustworthy.
It fits and forecasts but does not establish that the decline is real learning rather than scale or case mix (experience_attribution_model, case_mix_normalization — Cumulative Volume Cohort Analysis), nor watch the curve live across sites (quality_and_safety_guardrail, cross_site_learning_comparison — Learning Rate Dashboard).
Related¶
- Instantiates: Experience Curve Cost Reduction — it is the quantitative core that turns a validated curve into a rate and a bounded forecast.
- Consumes: Cumulative Volume Cohort Analysis — supplies the confound-controlled cost series the model can honestly fit.
- Sibling mechanisms: Cumulative Volume Cohort Analysis · Learning Rate Dashboard · Production Learning Log · Playbook Revision Cadence · Time-and-Motion Study · Yield and Defect Pareto Review · Standard Work Revision · Setup Reduction Workshop · Simulation Drill Ladder · After-Action Review
Editorial Notes¶
Form Classification¶
Form family: Analysis, Modeling & Optimization
Rationale: Experience Curve Model operates as a computation, comparison, model, or analytic representation used to infer, estimate, or choose because it fits the power-law between cumulative volume and unit cost into a single learning rate and a forecast — and flags when the curve is flattening and extrapolation should stop.
Independent corroboration: The frozen evidence defines Experience Curve Model as 'Fits the power-law between cumulative volume and unit cost into a single learning rate and a forecast — and flags when the curve is flattening and extrapolation should stop', so its operative form is Analysis, Modeling & Optimization.
Review outcome: Independent reviewer agreement; high confidence.
Origin Attribution¶
Primary origin: Organizational & Management Science
Origin pattern: Convergent development
Present-day reach: Multi-domain
Rationale: The generalized experience curve was formalized in industrial strategy and operations management as cumulative volume reduces unit cost.
Related originating lineages:
- Aviation & Aeronautics — Aircraft production supplied Wright's earlier empirical learning-curve formulation.
- Economics & Finance — Economic learning-by-doing theory materially shaped interpretation and forecasting. Learning-by-doing models materially formalized the productivity and cost relationship in economics.
- Engineering & Design — Manufacturing learning curves and Wright's law independently shaped empirical power-law estimation from production experience.
Review resolution: Both reviewers agree that organizational_management is primary. I retain aviation_aeronautics, economics_finance, engineering_design only as formative origin lineages; convergent is appropriate because the same operational pattern arose through parallel professional lineages. Reach is multi_domain because the structure transfers across several fields but is not a near-universal human pattern, an applicability judgment kept separate from provenance. Encyclopedia synthesis is false because the artifact is already established enough that encyclopedia-specific synthesis is not required. No unresolved historical ambiguity remains after reconciling the secondary fields.
Review outcome: Reconciled after independent review; high confidence.
Notes¶
The model is only as honest as the attribution beneath it. Handed a raw cost series, it will happily fit a beautiful curve to what is actually scale or a falling input price — so it belongs on top of Cumulative Volume Cohort Analysis, not in place of it. Fitting first and asking "is this learning?" afterward is how a plausible curve becomes a costly forecast.
[n1] Wright's law — the empirical observation, first documented for airframe production, that unit cost tends to fall by a roughly constant percentage for each doubling of cumulative output. It is a descriptive regularity across many industries, not a physical law, which is exactly why a model built on it must bound its extrapolation and impose a floor. ↩