How to Forecast Non-Obvious Innovations
Year: 2026
Type: preprint
Abstract: Companies forecast innovation with instruments built for forecasting sales: they extrapolate the attribute the market rewards today and call the slope the future. This is pattern recognition, and pattern recognition reads frequency in a remembered series. It is silent exactly where innovation happens — at the jump, when a lighter embodiment dissolves an obligation buyers had stopped noticing, an event absent from any series because it has not yet happened. This paper proposes path recognition as the alternative operation: reading the dependency structure that holds an embodiment in place — its obligations, its identity chain, the capabilities that could change it — rather than the frequency of its past. The paper separates four words the roadmap conflates (forecast, prediction, expectation, desire), places eight established forecasting families on the map of what they read and what they output, and states plainly where a fat-tailed environment breaks each of them, drawing on Kahneman's account of the heuristics that generate pattern recognition and on Taleb's account of the tails that defeat it. It surveys how six other fields already forecast under related conditions — meteorology, epidemiology, seismology, venture capital, military net assessment, and the deliberate culture-building studied in team research — and treats the innovating organisation as a system of stocks, flows and feedback, in the sense used in systems theory, so that an innovation jump becomes a leverage-point intervention rather than a parameter tweak. A specific model is proposed and specified for testing: four measurable variables, a coding protocol, a historical backtest and a prospective trial, each with a stated failure condition. Extreme counterfactuals are then run across five unrelated domains — telephony, books, drinking water, photography and home heating — to stress the model outside its country of origin, alongside three deeper company forecasts (Versuni, Apple, OpenAI), each with a jump, a set of directions to leave alone, a disappearance risk and a falsifier. The paper closes with the positioning dissection a company can run on itself, the forecast machine that runs it quarterly, and the limits of the whole claim. Every forecast produced here is a method-generated hypothesis; the tests, ahead of the argument, decide it. Author ORCID: 0009-0008-6255-7724. License: Creative Commons Attribution 4.0 International.
Author: Melo de Magalhães, Gonçalo
Language: English
Field: Decision Sciences
Domain: Social Sciences
Open Access status: green
OpenAlex ID: W7214532848
Cites:
Cited by:
Related to: