Jorge Quitério Search
Technology

Estimating the saturation level of an S-curve from early data

Technologies are often described by S-curves: slow initial growth, rapid growth, then saturation. Fitting an S-curve to early data is unreliable, because the saturation level is barely identifiable until the curve bends. A simulation of 400 fits shows the spread of the estimates.

The S-curve is the most popular shape in technology management. A new technology improves slowly, then rapidly, then slowly again as it approaches a limit. The shape is widely accepted as a description. Fitting it to data is considerably more difficult.

The difficulty is that the early part of an S-curve looks exactly like an exponential. Until the curve starts to bend, the data contain almost no information about where it will stop. Yet the ceiling, the market size or the performance limit, is usually the number the decision depends on. This article examines how much of an S-curve must be observed before the ceiling can be estimated.

A forecaster estimating the eventual market share of electric cars from the years when they were 2% of new sales faced this problem: the early data fitted a ceiling of 20% as well as one of 80%. The same difficulty was widely visible in early 2020, when logistic curves fitted to the first weeks of COVID-19 case counts gave estimates of the final number of cases that changed by large factors from one week to the next.

The logistic curve entered technology forecasting through Fisher and Pry (1971), whose substitution model describes a new technology replacing an old one as a logistic in time. Richard Foster's Innovation: The Attacker's Advantage (1986) established the performance S-curve in strategy, with a warning that incumbents near the top of their curve are vulnerable to a newcomer starting another. Theodore Modis spent decades fitting logistic curves and, with Alain Debecker, quantified how uncertain those fits are by running some 35,000 fits on simulated data (Debecker and Modis, 1994). Their rule of thumb: with at least half of the curve observed and data precise to better than 10%, the ceiling is known within 20% with 90% confidence. Modis (2007) is a candid account of the strengths and weaknesses of the method. De Weck's course adds a caution of its own, mentioned in figures of merit and their theoretical limits: the empirical evidence for ubiquitous performance S-curves is thinner than the textbooks suggest.

The logistic

The standard S-curve has three parameters: the ceiling MM, the growth rate α\alpha and the midpoint t0t_0,

y(t)=M1+e−α(t−t0). y(t) = \frac{M}{1 + e^{-\alpha (t - t_0)}} .

Well before the midpoint, e−α(t−t0)e^{-\alpha(t - t_0)} is large and y(t)≈Meα(t−t0)y(t) \approx M e^{\alpha(t - t_0)}: an exponential with growth rate α\alpha. The ceiling and the midpoint appear only in the combination Me−αt0M e^{-\alpha t_0}. Any larger MM with a later t0t_0 fits the early data equally well. This is not a weakness of a particular fitting method; it is a property of the curve.

Where the information about the ceiling lies

The ceiling shows up through curvature, the gradual slowing of growth. Statistically, the information about MM in a data point is tiny while y≪My \ll M and grows as yy approaches M/2M/2. A forecaster who fits a logistic to data from the first 10% of the curve is mostly fitting noise to a parameter the data do not contain.

The figure below fits hundreds of noisy datasets. Before the midpoint, fitted ceilings scatter over an order of magnitude; around the midpoint, with small noise, they settle within roughly 20%.

The true curve has a ceiling of 100, a midpoint at year 20, and an adjustable level of multiplicative noise.

  1. Early fit. With data to year 14 (about 18% of the ceiling) and 5% noise, press "Fit 400 datasets". Only about a quarter of the fitted ceilings land within ±20% of 100; the middle 80% range from about 60 to over 400. The exponential fit (dashed) is as good a description of the data as the logistic.
  2. Fit at the midpoint. Move to year 20 (half the ceiling). With 5% noise, close to 90% of the fits land within ±20%, in line with Debecker and Modis' rule of thumb.
  3. Higher noise. Keep year 20 and raise the noise to 10%. In my simulation the share within ±20% drops to around 63%, less than the rule of thumb suggests. Their result also depends on how many points are available and how errors are weighted, so any single rule should be applied with caution.
  4. Later fit. Move to year 25 with 10% noise: about 95% of the fits are within 20%. The ceiling becomes identifiable only when the estimate is less urgently needed.

Non-identifiability before the midpoint. For t≪t0t \ll t_0, y(t)=Meα(t−t0)(1−eα(t−t0)+…)y(t) = M e^{\alpha(t-t_0)}\big(1 - e^{\alpha(t - t_0)} + \dots\big). Changing MM to kMkM and t0t_0 to t0+ln⁡k/αt_0 + \ln k/\alpha leaves the leading term unchanged and alters only the correction eα(t−t0)e^{\alpha(t - t_0)}, which is of the order of y/My/M. When all observations satisfy y/M≲0.1y/M \lesssim 0.1, those corrections are smaller than 5% noise, so the data cannot distinguish ceilings that differ by a factor of several.

Before the bend, a fitted S-curve is effectively an exponential plus an assumption about the ceiling. When the ceiling matters, it should be estimated from other information: physics for a performance limit, market size for adoption, or the incumbent technology for substitution. For electric cars, that information includes the total number of new cars sold each year and the share of buyers for whom charging is practical, both of which can be measured long before the adoption curve bends. The early data are still useful for the growth rate, since that is the parameter they identify.

References

  1. J. C. Fisher and R. H. Pry, "A simple substitution model of technological change", Technological Forecasting and Social Change, 3, 75-88, 1971.
  2. R. N. Foster, Innovation: The Attacker's Advantage, Summit Books, 1986.
  3. A. Debecker and T. Modis, "Determination of the uncertainties in S-curve logistic fits", Technological Forecasting and Social Change, 46(2), 153-173, 1994.
  4. T. Modis, "Strengths and weaknesses of S-curves", Technological Forecasting and Social Change, 74(6), 866-872, 2007.
  5. O. L. de Weck, Technology Roadmapping and Development: A Quantitative Approach to the Management of Technology, Springer, 2022.