Research Pulse — 2026-06-15-timesfm

# Research Pulse — 2026-06-15 (timesfm)
_Brown Biotech tooling track · manual curation by Demis_

## TL;DR

[Das et al., Google Research, ICML 2024 · arXiv 2310.10688](https://arxiv.org/abs/2310.10688) — **TimesFM 2.5** (200M decoder-only time-series FM, 16k context, XReg covariates, 30M quantile head) 가 zero-shot point + 10-quantile forecast을 `pip install timesfm[torch]` 한 줄로 deploy 가능하게 만들었다. **BB stack에 "temporal foundation module"** 이 생긴 것 — biostatx · strict-omics · peptide-service 의 시간축 분석을 ARIMA/Prophet → FM baseline + LoRA 패턴으로 전환할 수 있는 인프라. **scGPT/Stack이 single-cell 축의 foundation이라면, TimesFM은 temporal 축의 foundation.**

### 왜 지금 — 2026-06-15

- v2.0 → **2.5 업그레이드** (Sept 2025): 200M (500M↓), 16k context (2048↑), 30M quantile head, **frequency indicator 제거**
- **XReg covariate** (Oct 2025) — BB의 "적은 N + 풍부한 메타데이터" 임상/랩 데이터에 fit
- **LoRA fine-tuning example** (Apr 2026) — HuggingFace + PEFT, biomedical adaptation 진입장벽↓
- **AGENTS.md + SKILL.md** (Mar 2026) — agent-native
- Google 1P 제품 통합: **BigQuery ML · Google Sheets · Vertex Model Garden** (production-validated)

### Headline numbers

| 항목 | TimesFM 2.5 | 비교/비고 |
|---|---|---|
| Params | 200M | 2.0의 500M에서 축소 |
| Context | **16k** | 2.0의 2048에서 8x↑ |
| Horizon | up to **1k** | 30M continuous quantile head 경유 |
| Quantile | 10개 (10th–90th) | continuous head |
| Pretraining | 100B+ time-points | 공개 + 합성 |
| License | Apache 2.0 | open |
| GitHub stars | 20.8k | active dev, 326 commits |
| Dist | PyPI + HF + BigQuery + Sheets + Vertex | 5+ distribution channel |

### Architecture — decoder-only + patched input

1. Continuous time-series → patched (각 time-step이 sub-series token)
2. **Decoder-only transformer** (GPT-style causal mask) — 200M
3. 출력: point forecast + 10-quantile (mean + 10th–90th)
4. **XReg (optional linear head)**: known covariates (dose, treatment, age) 추가 → 적은 N에 강함

### Code (canonical)

```python
import timesfm, numpy as np

model = timesfm.TimesFM_2p5_200M_torch.from_pretrained(
    "google/timesfm-2.5-200m-pytorch"
)
model.compile(timesfm.ForecastConfig(
    max_context=1024, max_horizon=256,
    normalize_inputs=True,
    use_continuous_quantile_head=True,
    force_flip_invariance=True,
    infer_is_positive=True,
    fix_quantile_crossing=True,
))
point, q = model.forecast(
    horizon=12,
    inputs=[np.linspace(0, 1, 100), np.sin(np.linspace(0, 20, 67))],
)
# point: (n_series, horizon)
# q:     (n_series, horizon, 10)
```

### BB 통합 — 6 actions

| Track | Action | 강도 |
|---|---|---|
| **biostatx** | cohort trajectory zero-shot benchmark (TimesFM vs Prophet vs ARIMA) | 🟢 HIGH |
| **strict-omics** | time-course RNA-seq (treatment arm covariate) — TimesFM + XReg PoC | 🟢 HIGH |
| **peptide-service** | PK/PD curve + dose covariate via XReg | 🟡 MED |
| **research-intelligence** | paper citation trajectory forecast (meta-time-series) | 🟡 MED |
| **genox-site** | longitudinal cohort patient-stratification pre-screener | 🟡 MED |
| **business-pipeline** | lab throughput + ticket volume forecasting (운영 메트릭) | 🟢 HIGH |

### ⚠️ Caveats

- Pretraining data는 **biomedical 아님** (economic/web/energy 위주). Zero-shot이 domain-tuned classical model을 항상 이기진 않음 → **foundation baseline + LoRA fine-tune** 패턴
- **16k context** — fine for biomarker streams, problematic for whole-genome per-gene
- **Irregular sampling** (clinical time-point) 처리 — 명시적 지원 없음. 전처리 layer 필요
- **LoRA fine-tune 예제는 economic data** 기준 — biomedical covariate에 잘 작동하는지 BB가 자체 검증 필요

### CTA — Brown Biotech로

Biomedical 시계열 데이터의 ARIMA/Prophet → **Foundation Model + LoRA** 전환. Brown Biotech는 TimesFM 2.5의 **XReg covariate + LoRA** 패턴을 biostatx / strict-omics / peptide-service 의 시간축 lane에 통합합니다 — IPF progression, biomarker trajectories, PK curves 같은 longitudinal 데이터에 **zero-shot FM baseline → 도메인 fine-tune** 워크플로우를 제공합니다. **임상/랩의 시간축 데이터, BB가 FM으로 정량화합니다.**

**Resources:**
- 📄 [arXiv:2310.10688](https://arxiv.org/abs/2310.10688) — ICML 2024 paper
- 💻 [github.com/google-research/timesfm](https://github.com/google-research/timesfm) — 20.8k ⭐, Apache 2.0
- 🤗 [HF: google/timesfm-2.5-200m-pytorch](https://huggingface.co/collections/google/timesfm-release-66e4be5fdb56e960c1e482a6) — checkpoint
- 🧠 [timesfm-forecasting/SKILL.md](https://github.com/google-research/timesfm/tree/master/timesfm-forecasting) — agent skill
- 🛠 [LoRA fine-tuning example](https://github.com/google-research/timesfm/tree/master/timesfm-forecasting/examples/finetuning) — HF + PEFT
- 📚 [bb-wiki/concepts/timesfm.md](/Users/ocm/openclaw/workspace/bb-wiki/concepts/timesfm.md) — full concept page with 4-섹션 judgment layer

*#ToolingTrack #TimesFM #FoundationModel #TimeSeries #XReg #LoRA #BrownBiotech #Biostatx #StrictOmics #PeptideService*