EWMA Load Trend
Also known as: Acute:Chronic Load Smoothing, Load Trend Chart
An exponentially weighted moving average (EWMA) of weekly AU, computed nightly. Acute load uses a 7-day window; chronic load uses a 28-day window. The chart is descriptive only — it shows whether your load is rising, falling, or stable over the last 12 weeks.
Formula
acuteAU_7d = EWMA(weeklyAU, alpha = 2 / (7 + 1))
chronicAU_28d = EWMA(weeklyAU, alpha = 2 / (28 + 1))
trend in {rising, falling, stable} [computed from the last 7-day slope of acuteAU_7d]Example
Last 4 weeks: 1,800 -> 2,100 -> 2,300 -> 2,200 AU. Acute (7d) tracks the most recent week, chronic (28d) the rolling average. If acute climbs noticeably faster than chronic, the chart labels the trend "rising."
How Afitpilot Uses This
A nightly batch job (compute_load_trends.py) reads each athlete's weekly load_summary docs and writes a load_trends/current doc with the smoothed values and trend label. The coach drawer renders a 12-week chart with neutral styling: single colour, no severity badges, explicit "descriptive only — does not predict injury risk" legend.
Why descriptive, not predictive
| Who / Context | Value | Note |
|---|---|---|
| Original ACWR research (2016) | Spike >1.5 = injury risk | Influential but methodologically criticised — population mean dressed up as individual signal |
| Impellizzeri (2020) | Don't use ACWR for individual prediction | Comprehensive critique; load monitoring still useful, but for description not prognosis |
| Impellizzeri (2023) | Same conclusion, stronger statistical case | Re-examines original cohorts and finds the predictive signal disappears once methodology is corrected |
| Afitpilot policy | No risk badges, no automated alerts | The moment a colour ships, it gets read as a risk signal regardless of the disclaimer |
Known Limitations
- •Per Impellizzeri (2020/2023), the literature does not support acute:chronic ratios for individual injury prediction. The chart deliberately does not surface ACWR as a number, because doing so invites misinterpretation.
- •The 7d / 28d window choice follows convention but has not been validated for self-coached non-elite athletes. We're treating it as v1 and will revisit after 3 months of production data.
- •Trend classification (rising/falling/stable) is computed from a 7-day slope and can flip noisily across small load fluctuations week-to-week.
- •Weeks with very few completed sessions (gaps) underweight chronic load — the EWMA does not distinguish between "low load week" and "didn't train this week."
Science Context
EWMA-based load smoothing was popularised by Williams et al. (2017) as an improvement on rolling-average ACWR. The smoothing is mathematically sound — it weights recent observations more heavily without abrupt drops at the rolling-window edge. The contested part is what to do with the resulting numbers. Afitpilot's stance follows Impellizzeri 2020/2023: ship the chart for description, do not infer individual risk from it, and never colour it as if you could.