Opifex

Inside Return's model

Built on forty years of research.

Return's model draws on more than forty peer-reviewed papers across seven research domains. Every coefficient — the weight Return gives to sleep, to HRV, to noise, to weekends — is anchored to a citation, not a guess. Here's where the numbers come from.

40+
Peer-reviewed papers
7
Research domains
15
Health signals
100%
On-device

Research foundations

Seven fields, one model.

01

Self-Determination Theory

The backbone. Developed by Deci & Ryan over four decades, SDT holds that well-being depends on three psychological needs — autonomy, competence, relatedness. Return's entire model is organized around the first: the sense that what you're doing is yours to choose. The weekend effect (β = +0.60), the contextual category shifts, and the daily emotion term all trace back to this literature.

  • Ryan & Deci (2000). Self-Determination Theory and the facilitation of intrinsic motivation.
  • Chen et al. (2015). Basic psychological need satisfaction across four cultures.
  • Reis et al. (2000). Daily well-being: autonomy, competence, and relatedness.
  • Ryan et al. (2010). Weekends, work, and well-being.
  • Milyavskaya et al. (2011). Balance across life domains.
  • Vansteenkiste et al. (2013). Autonomy support and dependence support.
  • Campbell et al. (2021). Within-person variability in autonomy.
  • Holding et al. (2020). Daily autonomy and affect.
02

Sleep Science

Sleep is the strongest physiological predictor in Return's model (β = +0.40). The app uses a multidimensional sleep score — duration, efficiency, blood oxygen, respiratory rate — with the optimal band set at seven to nine hours. Priors come from large cohort studies and consensus recommendations.

  • Hirshkowitz et al. (2015). National Sleep Foundation's sleep time duration recommendations.
  • Buysse (2014). Sleep health: can we define it? Does it matter?
  • Watson et al. (2015). Recommended amount of sleep for a healthy adult.
  • Cappuccio et al. (2010). Quantity and quality of sleep and incidence of type 2 diabetes.
  • Ohayon et al. (2017). National Sleep Foundation's sleep quality recommendations.
03

Heart Rate Variability

HRV is the second-strongest physiological predictor (β = +0.30). It reflects autonomic nervous system balance — a reliable biomarker for stress resilience and self-regulation. Return reads SDNN from Apple Watch and includes a one-day lag term (β = +0.20) to capture delayed cognitive and emotional effects.

  • Thayer et al. (2012). A meta-analysis of HRV and neuroimaging studies.
  • Laborde et al. (2017). HRV and cognitive function: a systematic review.
  • Forte et al. (2019). HRV as a marker of self-regulation.
  • Kim et al. (2018). Stress and heart rate variability: a meta-analysis.
  • Umetani et al. (1998). Normal values of HRV across decades of age.
  • Almeida-Santos et al. (2020). HRV reliability.
  • Koenig et al. (2016). HRV and personality traits.
04

Stress & Perceived Control

Physiological stress is computed from resting heart rate, HRV, respiratory rate, and walking heart rate — then weighted and lagged (β = −0.35 same-day, −0.20 previous-day) to capture carry-over. Environmental noise gets its own coefficient (β = −0.25): noise is a known contributor to cortisol release and sleep disruption.

  • Brosschot et al. (2016). Generalized unsafety theory of stress: unconscious perseverative cognition.
  • Münzel et al. (2024). Environmental noise and cardiovascular health.
05

Menstrual Cycle & Physiology

Return optionally reads menstrual cycle data from Apple Health to adjust baselines — hormonal cycles measurably affect HRV, sleep, and perceived autonomy, and ignoring them produces biased predictions. This feature is opt-in and, like everything else, runs on-device.

  • Schmalenberger et al. (2019). How to study the menstrual cycle: practical tools and recommendations.
06

Physical Activity & Well-being

The behavioral-autonomy term (β = +0.25) is a composite of exercise minutes, active energy, distance, standing hours, and mindfulness — weighted by their established links to psychological need satisfaction. Noise deducts from the score. The prior is anchored to meta-analyses of physical activity and self-determination.

  • Teixeira et al. (2012). Exercise, physical activity, and self-determination theory: a systematic review.
  • Hagger & Chatzisarantis (2014). An integrated behavior-change model for physical activity.
07

Methodology — EMA & Multilevel Modeling

Ecological Momentary Assessment — short, repeated check-ins in daily life — gives Return a within-person time series. The autoregressive term (β = +0.35, AR-1) captures the fact that yesterday's autonomy predicts today's. Ridge regression with a Bayesian prior handles the cold-start problem when you have only a week of data, following Dora's 2024 recommendation for small-sample multilevel models.

  • Bolger & Laurenceau (2013). Intensive longitudinal methods.
  • Hamaker et al. (2015). Within-person and between-person effects in longitudinal data.
  • Dora et al. (2024). Ridge regression in single-person multilevel models.

How the model learns

Three stages. All on your phone.

Return uses a single-person multilevel model with twenty-six features: sixteen core signals, nine missingness indicators (so the model knows when data is imputed), and one intercept. Training happens in three stages as you accumulate data.

Stage 0

Population Prior

Day 0 — no data yet

Twenty-six coefficients derived from published literature. You get a reasonable prediction from day one — not a personalized one, but not a guess either.

Stage 1

Ridge OLS

7–30 days of data

After a week, Ridge-regularized regression fits against your own observations. The prior still pulls the model toward published values, but your data starts to bend it. λ = max(0.1, 10/n) adapts to your sample size.

Stage 2

Personal Correction

30+ days of data

A second layer (ΔBeta) captures how your coefficients deviate from the population. Out-of-sample temporal validation confirms the model isn't just memorizing your recent past.

No cloud calls. No telemetry. The entire training pipeline — feature extraction, z-standardization, Ridge solve, residual correction, temporal cross-validation — runs locally on your iPhone. Your model belongs to you and lives with you.