Bass Diffusion Model
Bass Diffusion Model
One-line summary: A quantitative model of new-product adoption driven by two forces — innovators who adopt independently (coefficient p) and imitators who adopt as a function of how many already have (coefficient q), scaled by a market potential M — and the workhorse for fitting and forecasting S-curves.
The insight
The Bass model decomposes the adoption rate into an innovation term (p, external influence — advertising, independent decision) and an imitation term (q, internal influence — word-of-mouth, network effects), bounded by M (the ultimate market). q dominates steepness (see curve-steepness-and-adoption-velocity); p governs the slow early phase. Maximum-likelihood estimation outperforms OLS on fit and one-step forecasts and yields standard errors.
The crucial practical caveat — the heart of "how much growth is left" — is that p, q, and especially M are badly identified. In an empirical study of EVs, literature values of p/q "exhibit dramatic variations," ad-hoc estimates are "poorly conclusive," and the estimated p is highly sensitive to the assumed market potential M. So the parameter that most determines remaining upside (the ceiling) is the hardest to pin down.
Evidence
- From 2026-06-16-academic-research-technology-adoption-s-curves: innovators adopt independently (
p), imitators adopt on contact with existing users (q); MLE beats OLS on fit and forecasts (Schmittlein & Mahajan 1982). - From 2026-06-16-academic-research-technology-adoption-s-curves: for EVs, literature
p/q"widely variable," ad-hoc estimates "poorly conclusive," andp"highly sensitive to the assumed market potential M" — an explicit warning to forecast users (Massiani et al. 2015). - From 2026-06-16-autoresearch-tony-seba-technology-disruption-s-curves: an illustrative AI fit uses p≈0.01, q≈0.8 (very high imitation), projecting ~77% of internet users by 2030 (Dr Li — illustrative, not calibrated).
Design implications
- Treat
M(the ceiling) as a range, not a point — the remaining-runway estimate inherits its uncertainty (see s-curve-position-in-stock-evaluation). - A high
q/pratio flags a steep, network-effect-driven curve worth catching early.
Contradictions / tensions
- A simple logistic/Bass doesn't always fit the S well; piecewise/fuzzy-logistic variants can outperform on short or noisy data (Tseng et al. 2014, via 2026-06-16-academic-research-technology-adoption-s-curves).