feature marts

ML readiness assets

The project exposes feature and label tables that could support future modeling experiments, while deliberately stopping short of deployed predictions, recommendations, readiness scores, or coaching.

Readiness boundary

These are analytical feature assets. They make model inputs, labels, lineage, and missing-data behavior easier to inspect, but this site does not train models, serve predictions, or convert the outputs into athlete guidance.

Feature library

Existing marts already separate features from outcomes.

Route prediction features

mart_route_prediction_features

Run-route grain with activity date, route id, and route distance bucket.
Route shape and terrain fields including segment count, average grade, altitude range, ascent, and descent.
Run outcome labels for completion distance, duration, average pace, and average heart rate.

Weekly training features

mart_weekly_training_features

Completed-week grain inherited from mart_weeks.
Current and prior-week run count, distance, duration, active-week flag, and resting heart rate.
Next-week run count, distance, duration, and active-week labels for future forecasting experiments.

Recent training context

mart_run_sessions, mart_days

Prior 7-day and 28-day run counts before each session.
Prior 7-day and 28-day distance before each session.
Prior 28-day average resting heart rate for session-level context.

Route and history features

mart_routes, mart_run_sessions

Stable route identity derived from representative H3 path and distance bucket.
Lifetime route run count and historical averages for distance, pace, duration, and heart rate.
Prior same-route run count and prior same-route averages before the target run.

Health context

mart_days, silver_health_days

Same-day resting heart rate, HRV, HRV status, sleep score, and sleep duration.
Rolling resting heart-rate context for daily and session-level analysis.
Payload availability flags so missing Garmin health endpoints remain inspectable.

Labels and outcomes

mart_route_prediction_features, mart_weekly_training_features

Route-level labels: completion distance, duration, average pace, and average heart rate.
Weekly labels: next-week runs, distance, duration, and active-week flag.
Labels are stored for transparent offline experiments, not live recommendations.

Potential ML use cases

Future work would start with descriptive, offline experiments.

Route pace prediction

Estimate expected pace for a known route using route shape, prior route history, recent training context, and same-day health context.

Weekly training load forecasting

Use completed weekly rollups and prior-week lag fields to test simple forecasts of next-week distance, duration, or active-week status.

Completion and duration estimation

Estimate likely run duration or completed distance for recurring routes without turning the result into a coaching recommendation.

Descriptive clustering and segmentation

Group routes, weeks, or sessions by observed behavior to explain patterns such as route types, training blocks, or recovery-heavy weeks.