A Theory-Driven Generalist Forecaster Cuts MAE by One Third Versus Bass and Classical Diffusion Models on 21 Technology Adoption Curves

This record contains a preprint on technology adoption forecasting using The First Pattern (TFP), a theory-driven generalist forecasting algorithm.

TFP v2.2 is a single, fixed-configuration generalist forecaster that processes time series without domain-specific tuning. In this study it is evaluated on 21 US household technology adoption curves spanning four adoption regimes (Early, Growth, Mature, Saturated) and 840 rolling-origin evaluation windows, and compared with classical diffusion models (Bass, Gompertz, Logistic) as well as SimpleTheta and Naive persistence.

Across the full panel, TFP reduces mean absolute error (MAE) by approximately one third relative to Bass (34% improvement) and SimpleTheta (35% improvement), with 95% entity-level block bootstrap confidence intervals excluding parity. To ensure that diffusion-model fitting failures do not inflate results, we also report matched-window comparisons on the 513 windows where Bass converges; TFP still improves MAE by 34% versus Bass on this subset.

TFP wins on 16 of 21 technologies (76%). The five technologies where Bass outperforms TFP, Colour TV, Internet, Cellular phone, Shipping containers, and NOx pollution controls, share a consistent pattern of extremely rapid, textbook S-curve adoption, where parametric diffusion assumptions are especially well matched and TFP’s conservative adaptation can lag. Naive persistence remains exceptionally strong in the Saturated regime (95%+ adoption), which constrains the achievable gain for any method and provides a useful sanity check.

Notably, TFP uses the same frozen configuration validated across 11 diverse forecasting domains including epidemiological forecasting, energy load, and retail demand. No adoption-specific tuning occurred after the v2.2 configuration was frozen. These results position TFP as a candidate generalist baseline for diffusion studies, complementing rather than replacing parametric adoption models for the fastest adoption curves. Code for TFP is available from the author on request. Supplementary materials and related resources are linked at the end of the manuscript.