Keio University ML researcher building reproducible forecasting models for LLM agents, epidemics, and biosignals.
3
papers here
3
reviewed by Evolve
2
first-author
1
corresponding
—
works · OpenAlex
2
h-index · OpenAlex
Bibliometric figures come from OpenAlex (2026-10-02) and cover the author's entire public record, not just the papers listed here. Reproducibility signal comes from Evolve's own reviews.
Research profile
Output—
Not yet scored — verified output data is insufficient. This axis would measure publication volume, venue tiers, and preprint-to-published conversion; OpenAlex shows 18 works in the last three years, but no formal output score exists yet.
Impact—
Not yet scored — citation-impact data is too sparse for this nascent record (h-index 2, i10-index 0). The axis would measure citation counts, field-normalized influence, and uptake of the benchmarks this researcher releases.
Focus0/100
The focus score is 0/100, reflecting a deliberately dispersed portfolio: the OpenAlex field distribution splits roughly one-third each across machine learning (LLM agents and time-series forecasting), machine learning / epidemic forecasting, and machine learning (time-series and biomedical signal processing). Topic counts are similarly spread — multimodal ML applications (4), forecasting techniques (4), topic modeling (3), solar and space plasma dynamics (3), EEG and brain-computer interfaces (3) — indicating breadth rather than a single specialized line.
Collaboration28/100
The collaboration score is 28/100, based on only 3 distinct coauthors in the raw record, with Yuta Funayama and Ragamayi Puli the most frequent partners. All three catalog papers are compact-team arXiv preprints where Nagashima takes first authorship on two (one also corresponding) and senior last authorship on one — a small-team, lead-your-own-paper pattern rather than large-consortium work, which keeps the score low while preserving clear authorship credit.
Activity86/100
The activity score is 86/100, driven by a steep output trajectory: OpenAlex counts 2 works in 2023, 2 in 2024, 5 in 2025, and 11 in 2026, totaling 18 works in the last three years. The 2026 volume alone exceeds all prior years combined and includes three arXiv preprints in the Evolve catalog — a cadence consistent with an early-career researcher in a high-throughput preprint culture.
Reproducibility85/100
The reproducibility score is 85/100, a mean across 3 reviewed papers. Evolve rates two preprints at high readiness (TERN and the ROOSTER shared-alignment paper) and one at medium (Forecast Workflow Bench). The verdicts consistently flag single-seed runs and missing significance tests, yet the open benchmarks, uncontaminated evaluation protocols, and tooling-heavy design suggest reproducibility is an explicit design goal of this researcher's work.
Research summary
Shunya Nagashima is an early-career machine-learning researcher at Keio University (Japan) working on time-series forecasting and its evaluation, spanning LLM agents, epidemic forecasting, and biomedical signal processing. OpenAlex records 18 works over the last three years with an accelerating cadence: 2 in 2023, 2 in 2024, 5 in 2025, and 11 in 2026. Three 2026 arXiv preprints appear in the Evolve catalog. As first and corresponding author, Nagashima introduced Forecast Workflow Bench, a benchmark for budget-constrained LLM forecast-tool decisions in which GPT-6 Astra wins at a 2.5% budget while local models trail a free heuristic. As first author of TERN, Nagashima proposed a delta-rule fast-weight memory with a seasonal reference that wins uncontaminated on three influenza benchmarks, though Evolve notes the gains lack significance tests. As last author on the ROOSTER paper, Nagashima contributed a shared relative-alignment module linking forecasting with PPG-to-vital-sign reconstruction. Bibliometrics indicate a nascent citation record (h-index 2, i10-index 0), a compact collaboration pool (3 distinct coauthors), and strong reproducibility signals (mean readiness 85 across 3 reviewed papers). The portfolio is deliberately broad — fields split roughly one-third each across three areas — so the profile reads as a fast-moving, hands-on builder of forecasting methods and benchmarks rather than a narrowly focused specialist.
Collaboration pattern
Nagashima works in compact teams: the raw record lists only 3 distinct coauthors, with Yuta Funayama and Ragamayi Puli the recurring partners. Authorship patterns in the 2026 catalog show self-led projects — first author on two preprints (one also corresponding) and senior last author on a third. The low collaboration score (28/100) reflects small-team, arXiv-driven work rather than large multi-institution consortia.
Research keywords
Highlights
TERN: delta-rule memory wins uncontaminated on three influenza benchmarks
Forecast Workflow Bench: budgeted LLM-agent decisions under a cost ceiling
ROOSTER: one shared relative-alignment module across two signal domains
Timeline
- 2023First works appear in OpenAlex records: 2 publications.
- 20242 further works recorded, holding an early-career pace.
- 2025Output accelerates to 5 works as the forecasting research line takes shape.
- 2026Most productive year: 11 works, including three arXiv preprints in the Evolve catalog — first-authored TERN and Forecast Workflow Bench, and last-authored ROOSTER shared-alignment paper.
Papers in the Evolve catalog
Learning Where to Look: A Shared Relative-Alignment Module for Time-Series Forecasting and PPG-to-Vital-Sign Reconstruction
- first authorcorresponding
Forecast Workflow Bench: Evaluating Language-Model Decisions with Budgeted Forecast Tools
- first author
TERN: A Delta-rule Memory with a Seasonal Reference and Online Adaptation for Epidemic Forecasting
Frequent co-authors
About this profile
This profile is AI-structured by Evolve.Pub from public bibliographic data (OpenAlex) and Evolve's own paper reviews. It is not written or endorsed by the researcher unless the profile is claimed. Radar dimensions are computed from real bibliometrics with fixed formulas — raw figures are shown alongside; missing dimensions mean insufficient data, never zero. Errors? Claim the profile to correct or hide it.
- Bibliometrics come from OpenAlex only; identity is name-resolved from public bibliographic data.
- Output and impact axes are unscored; citation counts (h-index 2, i10-index 0) reflect a nascent record.
- All three reviewed papers are 2026 arXiv preprints; peer-review status is unknown.
- Evolve verdicts cover only the 3 catalog papers, not the full 18-work record.