Glide · Science
How the glucose forecast works
Evaluation numbers from June 2026 runs · page updated 29 July 2026
Glide projects the next three hours of glucose from recent CGM readings and the insulin and carbs you logged, applied at full textbook strength. The dashed line answers one question: “if I take no further action, where does this go?” That is deliberately not a prediction of what will happen — in real life you will act: correct a high, treat a low, eat. The gap between the line and what you intend to do is the whole point. Everything runs on the phone; it is a glanceable aid, never dosing advice.
Why we refuse to chase the accuracy score. A forecast tuned to minimize error on your own history learns your habits — including the habit of stepping in. It quietly predicts “you'll be fine” because you historically made it fine. That is circular, and useless for a tool whose job is to help you decide whether to step in. So Glide keeps the center line at plain physiology and puts all the doubt into a measured uncertainty cone instead.
The cone comes from measured errors
The app replays its own projector across your recent history, on your phone, and records how wrong it actually was at each horizon. The band you see is the 10th-to-90th percentile of those real errors — and it is calibrated separately for day and night, because measured on real use, daytime errors run two to three times wider. When insulin or carbs are on board the cone widens further (your configured ISF and carb ratio are population numbers, good to roughly ±35%), and a bolus logged without carbs widens only the top — it may have covered an unlogged meal.
The numbers as a table
| Horizon | Daytime p10 … p90 | Night p10 … p90 |
|---|---|---|
| 30 min | −42 … +54 mg/dL | −19 … +18 mg/dL |
| 60 min | −71 … +90 | −33 … +30 |
| 120 min | −86 … +110 | −47 … +43 |
Where the skill actually lives
Validated by rolling-forward testing — fit on past data, score on unseen following days, repeat — genuine forecasting skill concentrates in the first hour. Past roughly 90 minutes, no model we tested meaningfully beats “guess your typical level”: the far horizon mostly says where glucose tends to settle, not what it will do next. That shapes how to read it: trust the direction of the line for the next hour; read the far cone as “where things tend to land,” never as an early warning. We also stress-tested this with heavier machine-learned models on the same unseen data; they improved on the simple approach by 5–10% at best — the limit is the information in the signal, not the model.
What it can't do
- It is a “do nothing” projection. When you act, reality diverges from the line — by design. The gap is the decision.
- Past ~90 minutes it reads as “where things tend to settle,” not an early warning for a specific spike or dip.
- A bolus logged without its carbs projects a drop that won't happen (the meal was unlogged). The cone's top carries that doubt, but the center will look alarming. Logging carbs with the bolus fixes it.
- It assumes your configured ISF and carb ratio are right. If they're off, the dose response is off with them.
- Exercise, stress, illness, and pump-site issues are not inputs.
- It goes blank rather than guessing when the latest CGM reading is more than 20 minutes stale.
These numbers come from calibration and rolling-forward evaluation on one household's real CGM history — tens of thousands of readings and pump-imported boluses. We describe the results; the underlying personal data is never shared.
← The carb estimator, measured Can AI count carbs? The numbers, the bake-off, and the weak spots.Glide displays data from a third-party CGM and is not affiliated with or endorsed by Dexcom. Not a medical device; not medical advice. [email protected]