More accurate "projected usage at reset" for the 7-day limit
Hi! The 7-day (7d) limit projection in the dashboard and extension is now more accurate and stable. Here is what was wrong, how we fixed it, and how much it improved.
🔍 Why we made this change
The old projection assumed "your usage rate over the last few hours stays constant for the entire 24 hours until reset." That caused two problems.
- Near-constant "will hit 100%" warnings during the day — Most people sleep at night. But applying your busy daytime rate straight through the nights of the remaining days (a 7d window usually has 3–6 days left) meant that after a bit of daytime use it often predicted "will reach 100%+ at reset," scaring you far more than reality.
- Jumpy projections — Because utilization is recorded in whole-percent steps, a single new data point could swing the projection by more than 6%p at once. The number changed every time you looked, so it was hard to trust.
🛠️ How we fixed it
① We factored in your daily activity rhythm
Instead of counting the remaining time as plain "hours left," we convert it into "how many *active* hours are actually left." Late-night sleeping hours are heavily discounted while your busy hours count in full — in other words, we subtract the hours you are asleep from the projection.
- This rhythm is not a generic average — where possible it is personalized to *your own* usage pattern, learned from your recent history (the times of day you are actually active). (If you do not have enough data yet — under ~2 weeks of activity — we temporarily use the all-users average rhythm, then automatically switch to your own pattern as data accumulates.)
- Safety: this adjustment only ever lowers over-prediction. It never predicts higher (more alarming) than before.
② We made the usage-rate estimate far more robust
- Instead of just the last 6 hours, we look at about 48 hours, weighting recent data more (EWMA) to filter out noise.
- We compute the rate per *active* hour, so mixed-in night hours do not distort it.
- We also reflect your weekday/weekend rhythm as a personal pattern. For example, if you binge on weekdays and rest on weekends, we automatically discount the upcoming weekend.
📈 How much better is it
Validated on 6,078 real anonymized usage cases:
- Daytime over-prediction down ~63% — For projections taken while you are actively using during the day, the average amount they overshot the true reset value dropped by about 63%. The "scaring you by going over 100%" effect is now about a third of what it was.
- Early-window jumpiness down ~70% — The 6%p-per-datapoint swings are far steadier now. The worst single jump shrank from ~68%p to ~13%p.
- Overall error improved — Error versus actual usage at reset is noticeably lower.
- We also took care to improve results for night owls, overseas users, and people who do not use it on weekends.
This improvement is already live on the dashboard, and will roll out to the extension in an upcoming update. You should be able to trust the reset projection a bit more now. As always, feedback is welcome. 🙌