TL;DR
A hobbyist combined about 1.5 years of Garmin sensor data with daily chess results from Lichess to train a model that predicts whether they will win a game. A sparse logistic regression reached roughly 60% accuracy and highlighted REM sleep and Garmin's stress metric as the strongest predictors of short-term cognitive performance.
What happened
The author exported daily chess results (date, starting ELO, change in ELO) from Lichess and approximately 1.5 years of activity and sleep metrics from a Garmin watch. They engineered roughly 20 features tied to exercise and sleep, then tested several statistical models; a sparse logistic regression performed best in cross-validation, predicting win/loss outcomes at about 60% accuracy. Feature weights showed REM sleep duration and Garmin’s stress signal (approximated as inverse heart-rate-variability) were positively associated with better chess performance, while recent active calories and light-sleep duration were negatively associated. Deep sleep and sedentary time appeared to have little predictive power in this dataset. The author built a personal Garmin app that computes a daily probability of an ELO gain and updates weekly, but says Garmin’s platform restrictions prevent publishing the app broadly because sleep data and Health API access are limited for hobby projects.
Why it matters
- Wearable sensors can be combined with real-world cognitive outcomes to infer short-term mental performance, beyond standard fitness metrics.
- Signals commonly treated as uniformly positive for health (e.g., exercise, deep sleep) may not align with near-term cognitive clarity.
- Consumer sleep and stress metrics can reflect different physiological states — Garmin’s stress tracks sympathetic arousal that can correlate with alertness.
- Platform restrictions on device data and APIs can limit translating personal research into widely available tools.
Key facts
- Data sources: Lichess chess export (date, start ELO, ΔELO) and ~1.5 years of Garmin watch data.
- About 20 exercise and sleep-related signals were engineered and evaluated.
- Modeling approach: sparse logistic regression; cross-validated accuracy ≈ 60% for predicting win vs. loss.
- Top positively correlated features: REM sleep duration and Garmin’s stress metric (approximated as 1 / HRV).
- Features with little or no predictive effect in this dataset: deep sleep duration and sedentary duration.
- Negatively correlated features included active calories and light sleep duration.
- Author built a personal Garmin watch app that reports the probability the day’s ELO will rise and retrains weekly.
- Garmin limits: sleep data not available on-device for third-party hobby projects and Health API access is restrictive, preventing marketplace publication.
What to watch next
- Whether Garmin changes Health API or sleep-data access policies to allow third-party apps broader capabilities.
- Replication of these results on other users or larger datasets — not confirmed in the source.
- Further validation of which wearable-derived signals generalize as predictors of other complex cognitive tasks — not confirmed in the source.
Quick glossary
- REM sleep: A sleep stage associated with vivid dreaming and memory consolidation; often linked to problem-solving and creativity in research literature.
- HRV (Heart Rate Variability): A measure of variation in time between heartbeats; lower HRV is commonly associated with sympathetic arousal and stress.
- Logistic regression: A statistical method used to model the probability of a binary outcome (e.g., win vs. loss) from input features.
- ELO: A rating system that estimates a player’s relative skill level in games such as chess; changes reflect game outcomes against similarly rated opponents.
- Yerkes–Dodson law: A psychological principle proposing an empirical relationship between arousal and performance, often described as an optimal moderate arousal level for complex tasks.
Reader FAQ
How accurate was the model at predicting chess outcomes?
About 60% accuracy for win/loss classification, confirmed via cross-validation.
Which wearable signals were most predictive of cognitive performance?
In this analysis, REM sleep duration and Garmin’s stress metric were positively associated with better performance; active calories and light sleep showed negative associations.
Is this model generalizable to other people?
The author notes the model was trained on personal data and was custom to their patterns; broad generalizability is not confirmed in the source.
Can the Garmin app be distributed on the marketplace?
Not currently — the author says Garmin’s restrictions on sleep data and Health API access for hobby projects prevent marketplace publication.
Home Photos About RookLift 2024 · April 25 · 5 minute read Some mornings I’d wake ready for the world—I’d feel alert, clear-headed, present. Other days, I’d retrace my steps…
Sources
- Training my smartwatch to track intelligence
- Using the Daily Suggested Workout Feature on a Garmin …
- Gaming Activity | Esports Activity
- A New Look at Chess: Engage Your Cognitive Skills …
Related posts
- Chroma Explorer: Native macOS Client for Managing ChromaDB Vector Stores
- 9to5Mac Daily: Jan 14, 2026 — New Details on Apple–Google Deal and More
- Musk denies knowledge of Grok underage sexual images as California launches probe