How Training Load Works
Follow the exact path from recorded heart rate to daily TRIMP, ATL, CTL and TSB.
One daily heart-rate load series
Completed activity TRIMP feeds the fitness/fatigue history. PSS remains a separate power-load metric: there is no PSS fallback and no user switch that makes this CTL/ATL series power-based.
For local calendar day d, let L[d] be the sum of available TRIMP from activities starting on that day. The implementation uses:
ATL[d] = ATL[d−1] + (L[d] − ATL[d−1]) / 7
CTL[d] = CTL[d−1] + (L[d] − CTL[d−1]) / 42
TSB[d] = CTL[d−1] − ATL[d−1]ATL means Acute Training Load, CTL Chronic Training Load, and TSB Training Stress Balance. The smoothing coefficients are exactly 1/7 and 1/42, not 1 − exp(−1/N). TSB describes the balance entering the labeled day, before that day’s load.
Day boundaries, initialization and gaps
The account timezone determines an activity’s local start date; an activity crossing midnight is assigned as a whole to its start day. Multiple sessions are summed. Calendar dates, including daylight-saving transitions, advance the daily recurrence once; it is not a rolling 24-hour series. The complete rebuild starts at the earliest recorded activity with ATL and CTL equal to zero. The first 41 days carry low-confidence context. A chart-range change does not reset this state.
A real rest day uses L = 0 and both averages decay. An imported activity without valid TRIMP is marked incomplete. Available TRIMP from other sessions still contributes, possibly leaving a daily numerical zero with an incomplete-data flag. The application cannot identify an activity that was never imported; a missing import can therefore look like no recorded work.
Verified example
Starting at zero, a first day with 70 TRIMP gives ATL = 10, CTL = 1.666667 and TSB = 0. A following rest day gives ATL = 8.571429, CTL = 1.626984 and TSB = −8.333333. Two same-day sessions of 30 and 40 TRIMP give the same first-day load as one 70-TRIMP session.
Recalculation and interpretation
Importing older activities, fixing source data, changing applicable thresholds or changing timezone can change historical inputs and later daily values. Recalculation is versioned and uses the applicable persisted state or rebuilds from the natural beginning when a compatible seed is missing. A partial history understates the model even if its arithmetic is correct.
This model describes accumulated recorded load, not a direct measurement of fitness, fatigue, injury probability or medical readiness. Readiness workload uses a different series and must not be substituted here. Open Training Load in the app to inspect the history.
Sources and implementation scope
Morton, Fitz-Clarke and Banister (1990), impulse-response performance modeling provides background, not a specification of every EffortSignal rule. The formulas, windows, quality gates and examples above describe EffortSignal 0.1.13, checked September 10, 2026. Product-specific thresholds and rounding should not be attributed to the background source.
Content checked September 10, 2026 · EffortSignal release 0.1.13