💻 computer science

Predicting consumer-technology ownership without a diffusion history

This paper demonstrates that perceived attributes of consumer technologies, rated by both humans and frontier language models, can significantly improve the prediction of ownership prevalence compared to launch-age baselines, although this advantage diminishes for short-term forecasts where ownership trends remain static.

Original authors: Irina Vartanova, Niels Selling, Jennifer Viberg Johansson, Pontus Strimling

Published 2026-08-14

📖 4 min read☕ Coffee break read

CC BY 4.0

Original authors: Irina Vartanova, Niels Selling, Jennifer Viberg Johansson, Pontus Strimling

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

Imagine you are a detective trying to solve a mystery before the crime has even happened. In the world of technology, this is the ultimate challenge: predicting whether a brand-new gadget will become a household staple or gather dust in a drawer. Usually, experts try to guess this by looking at how similar products performed in the past, like trying to predict the weather by looking at yesterday's clouds. But what if the gadget is so new that nothing like it has ever existed? There are no clouds to look at. This is the "pre-launch" problem. To solve it, scientists use a framework called UTAUT2, which is basically a checklist of seven feelings people have about a new tool: Is it useful? Is it easy to use? Do my friends think I should have it? Is it fun? Does it cost too much? If we can measure these feelings, maybe we can predict the future without needing a crystal ball.

This paper is about testing a new kind of crystal ball. The researchers wanted to see if they could predict how many people would own a specific technology just by asking about its "personality traits" (the UTAUT2 feelings), without needing to know the product's own sales history. They gathered a list of 65 different gadgets, from smartwatches to kitchen tools, and asked two different groups to rate them: real humans and super-smart computer programs (AI language models). The humans were regular people surveyed online, while the computers were two of the most advanced AI models available, named Claude Opus 4.7 and GPT-5.5.

The team then built a mathematical recipe to see if these ratings could predict the actual ownership numbers. They played a game of "hide and seek" with the data: they took one gadget out of the list, tried to guess its popularity using the ratings of the other 64 gadgets, and then checked if they were right. The results were surprising. When the computers did the rating, they were much better at guessing the popularity than the humans were. The AI model Claude Opus 4.7 made errors that were 17% smaller than a simple guess based on how old the technology was, and it beat the human ratings by a significant margin. The GPT model was also better than the humans, though not quite as good as Claude.

However, the story has a twist. While the attribute ratings were great at predicting how popular a gadget is right now, they failed to predict how that popularity would change over time. When the researchers looked at how ownership numbers shifted between 2022 and 2025, the ratings didn't help at all. In fact, just guessing that "nothing will change" was just as accurate as using the fancy AI ratings. This suggests that while we can use these traits to take a snapshot of a product's current success, they aren't magic enough to tell us if a gadget will boom or bust in the next few years.

The researchers also had to be careful about a sneaky trick the AI might have played. Since the AI was trained on data from the internet, it might have "known" which gadgets were already famous and just pretended to rate them based on their traits, when really it was just remembering the answer. They ran special tests to check this, and while they couldn't prove the AI wasn't using its memory, the evidence suggests it was doing the work of reasoning rather than just reciting facts.

In the end, the paper shows a promising new way to forecast the future of tech. By asking an AI to rate a new product on things like "how fun it is" or "how easy it is to use," we can get a surprisingly accurate guess of how many people will own it. However, the core result of this study applies to technologies that are already on the market, where the AI can use its training data to reason about their traits. It is a very sharp tool for understanding products that have already been judged by the market, but whether this same method can truly predict the success of a gadget before anyone has ever seen or owned it remains the next big question. It's like having a very sharp camera that can tell you exactly what a car looks like today, but it still can't tell you if that car will win a race tomorrow.

Technical Summary: Predicting Consumer-Technology Ownership Without a Diffusion History

Problem Statement
The study addresses a critical gap in technology adoption forecasting: predicting the market penetration of a focal technology that has no observed adoption history. Traditional diffusion models (e.g., Bass model) rely on fitting S-shaped curves to existing sales trajectories, a method impossible for pre-launch products or genuinely new innovations lacking comparable analogues. Existing pre-launch approaches, such as analogical forecasting or expert/consumer surveys, suffer from selection bias, scalability issues, and the overestimation of stated intentions versus actual behavior. The authors ask whether a compact set of perceived attributes, evaluated before launch, can predict the population ownership prevalence of a held-out technology without relying on its own historical trajectory.

Methodology
The research employs a cross-sectional design using data from 65 consumer technologies introduced between 2002 and 2018.

Data Collection:
- Human Ratings: A 2022 Prolific survey of US adults ( $n=678$ ) assessed awareness and ownership of 65 products. Respondents rated a subset of known products on six UTAUT2 constructs: performance expectancy, effort expectancy, hedonic motivation, social influence, facilitating conditions, and price value.
- LLM Ratings: The same attribute battery was elicited from two frontier language models, Anthropic Claude Opus 4.7 and OpenAI GPT-5.5, in 2026.
- Follow-up: A 2025 survey re-measured ownership to analyze short-term changes.
Model Specification:
- The primary outcome is ownership prevalence (percentage of respondents owning the technology).
- A sign-constrained binomial elastic-net regression predicts ownership using four core UTAUT2 attributes (performance expectancy, hedonic motivation, effort expectancy, social influence) and a log-age covariate.
- Constraints were applied to ensure coefficients aligned with theoretical expectations (non-negative).
Evaluation Strategy:
- Cross-Sectional Prediction: A leave-one-technology-out (LOO) procedure evaluates the model's ability to predict the ownership level of a technology excluded from the training set. Performance is measured by Mean Absolute Error (MAE) in percentage points.
- Change Prediction: A secondary analysis attempts to predict the change in ownership between 2022 and 2025, using 2022 ownership as an anchor.
- Baselines: Comparisons are made against a "no-change" baseline, a "predict-training-mean" baseline, and a covariate-only baseline (log technology age).
Validation Checks:
- A knowledge-cutoff check compared earlier and later model releases to detect if attribute ratings were influenced by prior knowledge of product success/failure.
- Permutation tests ensured that performance gains were due to attribute signal rather than model flexibility.

Key Results

Cross-Sectional Ownership Prediction: The attribute-based models significantly outperformed the log-age-only baseline.
- Human Raters: Reduced MAE from 14.44 pp (baseline) to 11.95 pp.
- GPT-5.5: Reduced MAE to 7.55 pp.
- Claude Opus 4.7: Achieved the lowest error at 7.09 pp, representing a 51% reduction over the baseline.
- The improvement of the LLM ratings over human ratings was statistically significant, with Opus 4.7 outperforming human raters by 4.87 pp.
Short-Term Change Prediction: The attribute battery failed to improve predictions for the 2022–2025 change in ownership. The "anchor-only" model (2022 ownership + log age) performed best (MAE 2.56 pp), and adding attributes slightly increased error. The attribute signal did not capture the dynamics of ownership movement over this short window.
Rater Differences:
- LLMs exhibited greater variance in ratings across technologies compared to humans, who showed score compression (likely due to small sample sizes for niche products).
- Discrepancies were noted in "social influence," where LLMs interpreted the construct as market coverage while humans interpreted it as social network visibility.
- The knowledge-cutoff check found no evidence that the LLMs' attribute ratings were driven by prior knowledge of specific product outcomes, though the check was limited to newer products not in the main training set.

Key Contributions

Attribute-to-Ownership Mapping: The study demonstrates that a mapping from perceived attributes to population ownership prevalence is stable across heterogeneous technologies, allowing for the prediction of ownership levels for held-out products without historical adoption data.
LLM as Raters: It provides empirical evidence that frontier language models can serve as effective substitutes for human survey respondents in rating technology attributes, often yielding more accurate predictions of ownership prevalence than human raters in this specific context.
Scope of Prediction: The work delineates the boundaries of attribute-based forecasting, showing it is effective for predicting static ownership levels (the "level" of adoption) but currently ineffective for predicting short-term dynamic changes (the "movement" of adoption) without historical trajectory data.

Significance and Claims
The paper positions itself as a "first step" toward attribute-only adoption forecasting rather than a complete solution. Its primary claim is that the mapping from perceived attributes to population ownership is stable enough to predict the ownership of a held-out technology, establishing a precondition for forecasting genuinely new products.

The authors modestly conclude that while the model successfully reproduces the adoption levels of technologies already on the market, it has not yet demonstrated the ability to forecast a technology before anyone owns it. The inability to predict short-term changes marks the current limit of the attribute signal. The study suggests that while the "level" of adoption can be inferred from attributes, the "movement" over time remains a challenge requiring wider time windows or technologies with larger ownership shifts. The deployment illustration (predicting 2027 ownership for 11 new products) serves as a proof-of-concept for how the model would be applied to future innovations, pending future validation against actual ownership data.

Technical Summary: Predicting Consumer-Technology Ownership Without a Diffusion History

More like this