How the forecast works
What you see
On Home, after your latest CGM reading, T1Q draws where your glucose is likely to go next: a line with a shaded range around it. The line is the model's middle estimate. The darker band spans the middle half of its estimates (25th to 75th percentile); the lighter band spans the 10th to 90th percentile, meant to hold about eight in ten outcomes. For the first two hours the forecast comes from the model described below. Past two hours, out to six, a fainter tail is drawn from your own history: from the past moments when your glucose sat near the same level, where it actually went next.
The forecast is a picture, for display only. It never triggers an alarm — T1Q's optional alarms fire only on CGM readings that have already happened — and it never suggests an insulin dose or any treatment. Always act on your meter or CGM reading and your care team's advice, not on a forecast.
What the model reads
Your recent CGM readings (the level, how fast it is changing over the last 5, 15 and 30 minutes, and how much it has moved), and — when your devices supply them — insulin and carbs logged in the last few hours, heart rate, steps and the time of day. When something is not recorded, the model is told it is unknown rather than given a zero.
How it was made
Gradient-boosted decision trees (XGBoost): one set predicts the middle of the range, one the band edges, at 30, 60, 90 and 120 minutes ahead. It was trained only on openly licensed public research datasets — 173 people with type 1 diabetes from eight studies — never on the data of people using T1Q. A model trained on your own data alone can take its place for your account only if it beats this one on your own held-out readings.
How accurate it is
Measured on people the model had never seen (cross-validation that holds out whole people, 113 people in the evaluation run), the average distance between the middle line and the real CGM reading was:
- 30 minutes ahead: about 14 mg/dL
- 60 minutes ahead: about 25 mg/dL
- 90 minutes ahead: about 32 mg/dL
- 120 minutes ahead: about 37 mg/dL
That beat simply assuming glucose stays where it is for about 95–98% of those people. The middle-half band held 47–49% of real readings, close to the half it aims for. The evaluation measured a slightly wider outer band than the one shown (5th to 90th percentile, aiming at 85%), and it held 82–84% — so expect the lighter band to miss a little more often than one time in five. Treat the range as a guide, not a guarantee. Error grows the further ahead you look, and it is larger around meals, exercise and missed data. On a new sensor, a compression low or a gap in readings the forecast can be wrong by much more.
The data it was trained on
With thanks to the researchers and participants behind these datasets:
- HUPA-UCM (Universidad Complutense de Madrid) — CC BY 4.0
- AZT1D (Mayo Clinic Arizona) — CC BY 4.0
- BrisT1D-Open (University of Bristol) — CC BY 4.0
- T1D-UOM (University of Manchester) — CC BY 4.0
- OU-DATASET — CC BY 4.0
- ShanghaiT1DM — CC BY 4.0
- MOVE-CGM (Ash, G., 2025, Harvard Dataverse, doi:10.7910/DVN/NC9JNB) — CC0 1.0
- DDATSwHR, Dataset Diabetes Adolescents Time Series with Heart Rate (Windesheim) — CC BY 4.0