Foundation Model Forecasting
Amazon Chronos-2 + ARIMA Statistical Baseline - Oura Ring Biometrics, Post-HSCT
Foundation Model
Watchunavailable
ModuleNotFoundError: No module named 'torch'
RMSSD MAE
WatchN/Ams
90% PI coverage: N/A%
HR MAE
WatchN/Abpm
90% PI coverage: N/A%
Feb 9 Chronos Holdout
WatchNot run
Chronos retrospective holdout unavailable
Ruxolitinib Start
Info16. mar 2026
March Ensemble
ARIMA only
0 HIGH confidence anomalies | Feb 9 not detected
NIGHTLY FORECAST
1. Nightly RMSSD and HR - Chronos unavailable
Chronos could not be loaded in this run, so nightly foundation-model forecast charts were skipped. Statistical baseline outputs remain below.
HOURLY HR
2. Continuous HR - Chronos unavailable
Hourly Chronos analysis was skipped because the Chronos pipeline was unavailable.
ENSEMBLE
3. Statistical Baseline - ARIMA
Chronos was unavailable, so the consensus view is reduced to the statistical baseline only.
FEB9 RETRO
4. Feb 9 Retrospective Validation
Retrospective Chronos validation was skipped because the foundation model was unavailable.
Feb 9 Detection Summary
| Method | Series | Detected? | Residual |
|---|---|---|---|
| Chronos retrospective not available in this run. | |||
Prospective March ensemble HIGH-confidence anomalies:
None detected
Feb 9 in March ensemble consensus:
No
RUXOLITINIB
5. Pre vs Post Ruxolitinib Regime Analysis
Pre/post-ruxolitinib Chronos regime analysis was skipped because the foundation model was unavailable.
METRICS
Detailed Metrics
| Section | Metric | Value |
|---|---|---|
| data_range | ||
| start | 2026-01-08 | |
| end | 2026-07-02 | |
| n_nights | 178 | |
| ensemble_consensus | ||
| error | no overlapping forecasts | |
| statistical_hr | ||
| model | ARIMA(0, 1, 2) | |
| mae | 4.270 | |
| rmse | 5.493 | |
| coverage_90ci | 100.000 | |
| n_anomalies | 0 | |
| anomaly_dates | ||
| statistical_rmssd | ||
| model | ARIMA(0, 1, 1) | |
| mae | 1.448 | |
| rmse | 1.843 | |
| coverage_90ci | 92.900 | |
| n_anomalies | 1 | |
| anomaly_dates | 2026-03-09 | |
N=1 retrospective case study. All detection metrics are descriptive, not inferential. The model was trained and evaluated on a single patient's data. Validation requires an external multi-patient cohort.