Half-Life: Predicting Which Research Ideas Compound and Which Fade

Published in Vela Partners, 2026

Every hype cycle produces hundreds of “next big things”. Most fade within two years. A few, such as transformers and GRPO, become the foundation of a field and of the companies built on it. The usual way to tell them apart is momentum: how fast something is growing. But in their first months, a durable idea and a spike look identical.

Half-Life reads a research theme’s first months and predicts whether it will still be growing two years later.

Result 
+0.23 AUPRC over momentumdurable vs fad across 383 AI/ML research themes, 95% CI [+0.12, +0.33], p < 0.001
6 / 6pre-registered cross-field predictions correct
+0.16, p = 0.002transfers unchanged to biology (263k bioRxiv preprints)
Momentum subsumedadding momentum to the model adds nothing on top
0 look-aheadevery datum carries the date it became public; a leakage test gates every build

The pipeline

  1. Collects papers and eight other alternative-data sources, storing each record with the date it became public.
  2. Detects new research themes as they emerge, month by month.
  3. Measures each theme as it looked at the time: how persistently it appears, how many independent groups adopt it, how fast it grows, how widely it spreads.
  4. Predicts durable vs fad, and reports the theme’s current trajectory (accelerating, sustained, cooling, faded) and the researchers driving it.

Detector performance against the momentum baseline

Quant methodology

The project applies quant-finance backtesting discipline to research and alternative data. Each control is enforced in code, not merely described:

  • Point-in-time data (the ALFRED / real-time data approach). Every record carries a reference_date (what it is about) and a knowledge_date (when it became public). A month-T query sees nothing published after T.
  • A leakage canary: check_no_future_leakage runs on every panel build and fails it on any violation.
  • Survivorship-free universes: failed markets and dead ideas stay in every month’s universe.
  • Look-ahead leak hunting: arXiv venue and institution fields turned out to be backfilled after acceptance, and were removed.
  • Pre-registration: hypotheses, thresholds and all six field predictions were fixed before scoring.
  • Placebo and permutation tests: label-shuffle placebos and stratified permutation nulls, to rule out lift from a bug or from sector/year composition.

Transfer of the detector to biology preprints

Capital null test

Signal decomposition for LLM developer tools

In production at Vela

Vela Partners now runs an extended version in house as a monthly research-trends product: every idea in the field plotted as its share of papers over time, with a durability verdict, trajectory and the researchers behind it.

Vela Research Trends product

Rising-now view

Conversational interface over the trends data