What predicts my sleep?

• Question Verbatim from You: “What predicts my sleep?”
• Interpretation of the Question (an attempt to make the question slightly more precise, computable and scientific)
  • In your question you said “my sleep”. We are measuring that as two things your OuraOura makes a ring worn on the finger that measures sleep, heart rate, heart rate variability and skin temperature overnight, and scores each night. More at ouraring.com ring records every night: total sleep and deep sleep.
  • “What predicts” we are interpreting as: of everything logged about your day — sets lifted, steps, the ring's daytime stress reading, and on days you logged food, calories, protein, alcohol, caffeine and the time of your last meal — which go with more or less sleep that night, each with the others held steady.
  • Precisely: one model for every night with workouts logged, one more for the nights with food logged too, and a third for 2026 that adds the steam room.

Partly, based on the analysis below, and within its assumptions and limits.

• Days with more high stress on the ring were followed by less sleep: about 5 minutes less per hour of high stress.
• Days with more steps were followed by less sleep: about 3 minutes less per 1,000 steps.
• Both hold once everything else tested is allowed for. Nothing else does, and nothing predicts deep sleep.

• Data sources:
  • Devices:
    • OuraOura makes a ring worn on the finger that measures sleep, heart rate, heart rate variability and skin temperature overnight, and scores each night. More at ouraring.com ring data: nightly total and deep sleep, daily steps and daytime stress
    • StrongStrong is a phone app for logging weightlifting workouts in fine detail: every exercise, and the weight and reps of every set. More at strong.app app data workout log, exported 2026-10-01
    • TrainerizeTrainerize is an app personal trainers use to send clients their workouts and to log what was done. More at trainerize.com app workout log, exported 2025-11-01: the sets for 2024-11-02 to 2025-10-17, when the StrongStrong is a phone app for logging weightlifting workouts in fine detail: every exercise, and the weight and reps of every set. More at strong.app app was not in use
    • MyNetDiary food log to 2024-11-03 and MacroFactor food log from 2024-11-04: calories, protein, alcohol, caffeine and the time of each entry
    • Steam-room sessions, 2026, from your chat with PerSci Bot
  • Textual sources:
    • Your conversations with PerSci, 2026-10-01 to 2026-10-27: when each trainer started, that every set you log is taken to failure, and what the TrainerizeTrainerize is an app personal trainers use to send clients their workouts and to log what was done. More at trainerize.com session ratings were.
• Dates: 2024-03-15 to 2026-09-28: 895 nights with workouts logged, 539 of them with food logged too; 200 nights in 2026 for the steam room.
• People and Organizations Named or Anonymized:
  • Named (in your template names, your notes, or what you told us): None
  • Anonymized (shown by alias, inside your own notes and quotes too): None

What happened in your day, each compared with your usual for that weekday and season.

• Sets lifted
• Steps
• High daytime stress (OuraOura makes a ring worn on the finger that measures sleep, heart rate, heart rate variability and skin temperature overnight, and scores each night. More at ouraring.com)
• On days with food logged: calories, protein, alcohol, caffeine, time of last food
• 2026 only: the steam room

Your sleep the night that began that evening.

• Total sleep
• Deep sleep

Registered on 2026-10-28 before it ran, after you had settled two choices: daytime stress goes in, with the night before's sleep held steady, and the steam room gets its own 2026 model because you had seen its result already.

Daytime stress: about a 1-in-5000 chance of a link this strong by luck alone (p = 0.0002, raw); allowing for every link in its model, 0.001. Steps, in the food model: corrected 0.003.

For p-values, corrections for testing many questions, effect sizes and the smallest detectable differences, see The statistics behind this below.

• Stress and sleep come from one ring. The same heart-rate sensor reads both, so some shared error is possible. The night before's sleep is held steady to stop a bad stretch continuing from passing for stress.
• Stress does not show in the food model (−0.9 min, not significant) — but that model has fewer nights and could only have seen about 5 minutes, so it does not rule it out.
• Caffeine is weak here. The test could only have seen a large effect, likely because coffee was often not logged.
• Calories and protein move together (r 0.86), so their separate estimates are loose.
• Sets were already seen (the lifting-and-sleep question), and the steam room on its own page.
• Nights before 2024-03-15 are not used: workouts were not logged then.
• This describes a pattern in your data. It is not medical advice.

• Did I have better body comp progress when I had been sleeping more? — sleep against the same six-week body composition changes
• Did I have better body comp progress (more lean muscle, less fat) when I got more training stimulus? — weekly sets against the same six-week body composition changes
• Did I sleep better when I worked out harder? — the other direction: the day's sets against that night's sleep
• Of my sleep, training and steps, what goes with lean gain and fat loss? — sleep, sets and steps in one model, on the same blocks
• What does my DEXA data look like? — every scan, the measurements these changes are computed from
• What does my Oura data look like? — every night's sleep, described without any question asked of it
• What does my Strong App data look like? — every logged session and set

One least-squares fit per outcome per model. Every series has its centred 29-day rolling mean removed and is centred within epoch × weekday; the night before's sleep is a fitted control, never tested. Each coefficient is tested by Freedman-Lane permutation with the reduced model's residuals shifted circularly by whole weeks within each run of nights (a null that keeps runs of good and bad nights), 5,000 draws. Holm within model A and within model B; model C (2026, steam) is reported outside any correction. Model A: 895 nights; B: 539; C: 200.

Detectable: A real change this big would just pass the test, about half the time. One about 1.4 times as big would pass it about 4 times in 5. A smaller real change could easily go unseen, so 'no change found' means 'no change this big'.