Research Like Stock Trading: 40K ICLR Papers Backtest Shows Chasing Hot Topics Works

The author draws an analogy between academic research and stock trading, introducing the concept of momentum : papers in rising research areas get accepted more easily because reviewers become familiar with the direction, senior students publish there, and big labs release benchmarks and open-source models that add “volume.” However, momentum has crashes — just as momentum strategies in finance can wipe out years of gains in months, a research direction’s acceptance premium disappears once it becomes crowded.

Data and Metrics

The analysis uses 42,123 ICLR papers (2017–2026) with known accept/reject decisions. Papers are assigned to 28 research directions via fixed phrase matching in titles/abstracts (a paper can belong to multiple directions). Three metrics are computed per direction per year:

Share : direction’s paper count divided by total conference papers. Example: LLM share rose from 2% (78 papers) in 2023 to 34% (4,716 papers) in 2026 — a 60× increase in three years.

Share change (momentum) : year-over-year share difference. Positive = heating up; negative = cooling down.

Relative acceptance-rate difference : acceptance rate of papers in the direction minus acceptance rate of papers not in the direction, for the same year. This controls for overall conference acceptance-rate fluctuations. A value of +19 means the direction’s acceptance rate is 19 percentage points higher than the rest.

Key Findings

Trend Self-Reinforcement Confirmed

LLM’s explosive growth (78 → 863 → 2,357 → 4,716 papers) mirrors the described mechanism: influential papers → reviewer familiarity → student follow-on → industry benchmarks. The charts show this clearly.

Acceptance Premium Vanishes at Peak Heat

In 2023, LLM papers enjoyed a +19 pp acceptance premium . By 2024 it dropped to +7, 2025 to 0, and 2026 to +0.13 (statistically indistinguishable from zero) . Multiple-comparison correction across 45 tests still shows a significant decline. The trend continues but alpha is eaten up .

Agent: One-Year 4× Growth, Premium Turns to Discount

Agent papers grew from 128 (2025) to 533 (2026) with 38% acceptance rate, but relative acceptance difference is -1 (CI crosses zero) — no advantage. Prompt engineering showed a similar trajectory: +16 pp premium in 2023 (21 papers), +2 in 2024, -11 in 2025 , -6 in 2026. Six-fold paper increase turned premium into discount.

Left-Behind Directions: Growth Lands in Rejections

GNN papers rose 65% (210 → 347) from 2023 to 2026, but acceptances only grew from 83 to 93 while rejections doubled (127 → 254). Acceptance rate fell from 40% to 27%. This illustrates alpha decay : many groups pursue the same idea, and later submissions become rejections. The decline is not statistically significant after multiple-comparison correction, so the author treats it as descriptive.

Counterintuitive: Niche Directions Aren’t Necessarily Small

ICLR itself grew 3.6× (3,792 → 13,704 decided papers) from 2023 to 2026. Among 20 directions with ≥30 papers in 2023, 18 grew in absolute count, but only LLM, diffusion models, and interpretability outpaced the 3.6× conference growth. The other 15 directions grew in count but shrank in share . Causal inference, continual learning, fairness all doubled in papers yet became more marginal. “Cold” often just means “losing to the market.”

Three Flavors of “This Direction Is Dead”

Share ebb but headcount still rising — e.g., diffusion models (642 → 898 papers, share 7.5% → 6.6%). People still enter, just slower than LLM.

Actual headcount decline — only GAN (128 → 58) and NAS show this over a PhD lifetime.

Headcount rising but all growth goes to rejections — GNN case above.

Backtest: Does Chasing Heat Pay Off?

Using only data available at each year-end, the top quartile of directions by share growth are labeled “hot.” Over the next three years, the hot group’s average relative acceptance difference is higher than the non-hot group for start years 2020–2023 (2023 start: +3 pp). 2019 is negative but only three eligible directions — not reliable. Chasing momentum is not wrong. However, both groups’ average relative acceptance differences are negative: entering an already-hot direction yields slightly less disadvantage , not a free pass. Hot directions give you “lose less,” not guaranteed acceptance.

Full 28-Direction Heatmap

All 28 directions’ annual share changes (green = heating, red = cooling) sorted by 2026 paper count. Cells show share change, not acceptance rate or scientific value.

What Can and Cannot Be Concluded

Data scope : ICLR accepted/rejected papers only, fixed-phrase matching, non-exclusive categories, not all submissions. 2025 missing 217 texts; key numbers unaffected. All findings are historical associations, not causal claims .

Cannot conclude :

“Late entrants get exploited” — no individual entry-time data.

“Crowding causes rejections” — no causal identification.

“GAN has no value” — no long-term contribution or industry data.

The single actionable takeaway :

For PhD students choosing a direction, the question isn’t just “is it hot?” but “does this direction’s growth land in acceptances or in rejections?”