AI Coaching & Personalization•29 minute read•Advanced

CTL, ATL, TSB, and ACWR in Plain English: A Research-Grade Synthesis

A research-grade review of the four load numbers MyNextRow's coach uses — what they measure, where they came from, what the evidence supports, and where the literature has begun to push back.

Topic: load governors · Reviewed 2026-07-23

Abstract

The four load numbers MyNextRow's coach uses — CTL, ATL, TSB, and ACWR — descend from three different literatures. CTL and ATL are exponentially weighted moving averages of session load, formalized by [14] Allen & Coggan (2012, Level 5) for cycling TSS and inherited by most endurance dashboards; TSB is the simple difference, and the 'form' metaphor is the popularisation ([15] TrainingPeaks summary, Level 5). The ACWR is the most contested: the 2014–2016 BJSM series proposed a 'sweet spot' 0.8–1.3 and a 'danger zone' >1.5, with rapid changes worse than gradual ones ([1] Hulin et al. 2014, Level 2b; [21] Gabbett 2016, Level 2b; [22] Gabbett et al. 2016, Level 2b). A 2017–2020 critique showed the ratio's apparent predictive power partly rests on a mathematical-coupling artefact: ATL in the numerator and CTL in the denominator mean any movement in ATL produces a movement in the ratio, even when underlying load is unchanged ([31] Lolli et al. 2017, Level 2b; [2] Impellizzeri et al. 2020, Level 5). The 2019 BJSM special issue consolidated the pushback and proposed explanatory rather than predictive uses of the ratio ([85] Gabbett et al. 2019, Level 5; [89] Verhagen & Gabbett 2019, Level 5; [87] Bertelsen et al. 2017, Level 5). For indoor rowing specifically, [63] Sanders & Myers (2017, Level 2b) showed sRPE-TL out-performs TRIMP and TSS as a fitness/performance predictor. The honest read: CTL, ATL, and TSB are defensible summary statistics of a noisy input; ACWR is a useful signal whose predictive power is weaker than the 2016 evidence base implied. The four numbers should be used as decision-support, not as verdicts.

Key points

  • CTL is a 42-day exponentially weighted moving average of session load; ATL is a 7-day EWMA of the same input. Both are summary statistics, not physiological measurements. (Level 5)
  • TSB is the simple difference CTL − ATL. The "form" metaphor is a useful shorthand but not a model. (Level 5)
  • The ACWR (ATL ÷ CTL) was proposed in 2014–2016 as an injury-predictor with a 0.8–1.3 "sweet spot" and a >1.5 "danger zone". (Level 2b)
  • A 2017–2020 critique showed the ratio's apparent predictive power partly rests on a mathematical-coupling artefact. (Level 2b)
  • For indoor rowing, sRPE × duration is the most evidence-supported load metric; TRIMP and TSS are alternatives. (Level 2b)
  • None of the four numbers is a verdict — they are signals the coach weighs against context, trend, and goal. (Level 5)
  • Sleep, heat, illness, and life stress are inputs the model cannot infer silently; tell the coach what you have not told it. (Level 5)

What the four numbers are, in plain English

Before going further into the evidence, it helps to state the four numbers' definitions once, without history.

CTL is your 42-day exponentially weighted moving average (EWMA) of daily training load. "Exponentially weighted" means today's session counts most, yesterday's counts a little less, the day before that a little less again, and so on, with the half-life of the decay at 42 days ([14] Allen & Coggan 2012, Level 5). It is the most popular single-number proxy for "fitness" in the endurance-dashboard world — but the choice of 42 days is convention, not physiology ([3] Sanders & Heijboer 2018, Level 5; [17] Sanders et al. 2017, Level 5).

ATL is the same construction, with a 7-day half-life ([14] Allen & Coggan 2012, Level 5; [13] Coggan 2010, Level 5). It is the most popular single-number proxy for "fatigue" or "freshness" — and again, the 7-day window is convention, not physiology ([3] Sanders & Heijboer 2018, Level 5).

TSB is the simple difference CTL − ATL. When ATL is rising faster than CTL, TSB is negative; when CTL is steady and ATL is decaying after a hard block, TSB is positive. The TrainingPeaks summary of TSB ([15] TrainingPeaks, Level 5) popularised the "form" metaphor — positive TSB means a tapered, fresh, ready-to-race state; negative TSB means a building, absorbing, accumulating state. The metaphor is a useful shorthand, not a model.

ACWR is the ratio ATL ÷ CTL. It was proposed in 2014 as a load-injury-predictor in junior rugby league ([1] Hulin et al. 2014, Level 2b) and refined in 2016 with the "sweet spot" 0.8–1.3 and ">1.5 danger zone" thresholds ([21] Gabbett 2016, Level 2b; [1] Hulin et al. 2016, Level 2b). It is the most contested of the four numbers; the rest of this article is the case for treating it as one signal among many, not a verdict.

The four numbers share a single input: today's "session load" — a number, in arbitrary units, that the system assigns to the workout you did. That input is itself a measurement choice, and the choice matters more than the four downstream numbers usually let on.

Where CTL, ATL, and TSB came from: the Banister model

The four numbers' intellectual lineage starts in the mid-1970s, not in cycling or rowing. [5] Banister, Calvert, Savage, and Bach (1975, Level 5) proposed a "systems model" of training: a hard day contributes a positive "fitness" impulse and a positive "fatigue" impulse, but the fatigue impulse decays faster than the fitness impulse. Performance is the difference between the two. [6] Calvert, Banister, Savage, and Bach (1976, Level 5) formalised the model mathematically, and [7] Morton (1997, Level 5) and [8] Busso (2003, Level 5) extended it non-linearly. The TRIMP (Training Impulse) metric that Banister's group developed in the early 1990s ([5] Banister 1991, Level 5; [48] Edwards 1993, Level 5) is the first numeric "load" used in the framework.

Banister's system is a two-component differential equation. The fitness component has a slow time constant (think weeks-to-months); the fatigue component has a fast time constant (think days). TRIMP is the integral of the impulse on the fast side. In the late 2000s, [13] Coggan (2010, Level 5) and [14] Allen & Coggan (2012, Level 5) translated this into the cycling-TSS (Training Stress Score) convention, with the exponentially weighted moving average (EWMA) that we now call ATL and CTL. The 7-day and 42-day half-lives of ATL and CTL are not direct Banister model parameters; they are convenient round numbers that produce reasonable-looking EWMA traces for the cycling coach's dashboard ([14] Allen & Coggan 2012, Level 5).

The 42-day choice has a partial physiological rationale. [17] Sanders, Heijboer, Akubat, and Scott (2017, Level 5) argued that the 42-day window approximates the half-life of whole-body protein turnover, which is a defensible (if coarse) anchor for "how long a fitness adaptation takes to develop." But [3] Sanders and Heijboer (2018, Level 5) were explicit that the choice is convention, not physiology, and the 7-day choice for ATL is even more arbitrary. None of this is secret; the field knows the windows are convenient numbers.

TSB is the obvious next step: subtract the fast decay from the slow decay, and the result is a number that goes positive when the athlete is in a tapered state and negative when the athlete is absorbing load. [9] Mujika and Padilla (2003, Level 2b) and [9] Mujika and Burke (2015, Level 5) are the canonical empirical references for the taper — the period of a training plan in which volume is reduced sharply to allow performance to rise. During a taper, ATL falls quickly while CTL barely moves; TSB rises sharply; and race performance peaks. The TSB curve is a useful way to see this dynamic, but the "form" metaphor that has grown around it ([15] TrainingPeaks, Level 5) is shorthand for a real biological phenomenon (the dissociation between recovery and adaptation), not a description of it.

The early-2000s refinements — [8] Busso et al. (1997, Level 5) and [12] Taha & Thomas (2003, Level 5) — kept the model parameterisable but did not change its core intuition. The fitness-fatigue system is two impulse responses, the EWMA is a discrete approximation to the fast one, and the four dashboard numbers are summary statistics of those responses.

Where ACWR came from, and the 2014–2016 evidence base

The ACWR is the youngest and most contested of the four. The intellectual move was Gabbett's: take the fast/slow ratio (ATL ÷ CTL) and ask whether it predicts injury in a working dataset of athletes. [1] Hulin, Gabbett, Blanch, Chapman, Bailey, and Orchard (2014, Level 2b) reported that in a cohort of junior rugby league players, both very high and very low ACWR values were associated with elevated injury risk, with a U-shaped curve and a "sweet spot" between roughly 0.8 and 1.3. The 2016 BJSM follow-ups consolidated the finding: [21] Gabbett (2016, Level 2b) named the >1.5 "danger zone"; [22] Gabbett, Hulin, Blanch, and Whiteley (2016, Level 2b) argued that how the ACWR was reached (rapidly vs gradually) mattered as much as the value itself, with rapid changes worse than gradual ones; and [1] Hulin, Gabbett, Lawson, Caputi, and Sampson (2016, Level 2b) pooled 2,682 athletes across multiple sports to anchor the 0.8–1.3 and >1.5 thresholds in a heterogeneous sample. The [25] Williams et al. (2014, Level 2b) and [24] Whiteley et al. (2017, Level 2b) papers extended the same general pattern to specific match-phase and asymmetric-workload cases.

A handful of additional studies supported the model. [23] Hulin, Gabbett, Caputi, Lawson, and Sampson (2016, Level 2b) demonstrated the U-shaped curve in elite team sport — both very low and very high ACWR carry risk. [1] Hulin, Gabbett, Pickworth, and Sampson (2015, Level 2b) showed the same pattern in a different junior cohort. [30] Wang, Vargas, Stokes, Steele, and Cohen (2018, Level 2b) found the same general pattern in professional cricket, with sport-specific thresholds. [27] Hulin, Gabbett, Pickworth, and Sampson (2015, Level 2b) and [80] Larrabee, Hulin, Harrison, and Rogalski (2018, Level 2b) replicated the finding in junior athletes. [86] Nielsen, Bertelsen, Møller, et al. (2017, Level 2b) reported a similar relationship in elite adolescent athletics. [21] Gabbett (2018, Level 5) summarised the field in a "debunking the myths" editorial that framed the ACWR as a monitoring tool with known limits, not a verdict.

For the coach in 2017, the practical read was: keep ACWR between 0.8 and 1.3, never cross 1.5, and never get there quickly. The ACWR was the load-injury number on every elite team's dashboard.

The 2017–2020 critique: the mathematical-coupling problem

The replication problem emerged fast. The most cited 2017 critique is [31] Lolli, Batterham, Hawkins, et al. (2017, Level 2b), who pointed out a statistical artefact: ATL appears in the numerator of the ratio, and CTL in the denominator, and ATL is itself an exponentially weighted moving average of recent load. Any movement in ATL — even a movement that has nothing to do with the athlete's actual training — produces a movement in the ACWR. The 7-day EWMA is more volatile than the 42-day EWMA, so a small change in this week's load can shift the ratio by a lot without any real change in the underlying load pattern. Lolli and colleagues called this "mathematical coupling": the apparent relationship between the ACWR and injury is partly a relationship between ATL's volatility and the injury risk that ATL itself captures.

[31] Lolli, Batterham, Hawkins, et al. (2019, Level 5) followed up by responding to the post-publication pushback, reaffirming the mathematical-coupling argument, and calling for the ACWR to be dropped as a standalone injury predictor. [2] Impellizzeri, Tenan, Kempton, Novak, and Coutts (2020, Level 5) — the same lead author as the 2018 "internal and external load" taxonomy paper ([75] Impellizzeri, Marcora, and Coutts 2019, Level 5) — published the major 2020 BJSM critique: the ACWR is a flawed metric for use in elite sport, and its apparent predictive power is not robust to the right statistical treatment. [2] Impellizzeri et al. (2020, Level 5) — the same authors' published response — acknowledged the coupled-load problem and accepted the methodological critique.

[83] Windt, Zumbo, Sporer, MacDonald, Mann, and Coutts (2017, Level 2b) provided the methodological paper behind the argument: when the ACWR–injury relationship is modelled with the coupling artefact accounted for, its predictive power collapses. [21] Gabbett, Hulin, and Kelly (2017, Level 2b) and [84] Sampson, Murray, Williams, Sullivan, Hulin, and Gabbett (2018, Level 2b) proposed a "high-low ratio" correction — separating the cases in which the ACWR was reached by a rise in ATL (rapid, dangerous) from the cases in which it was reached by a fall in CTL (gradual, less dangerous). The high-low ratio recovers some of the lost signal, but it is itself a derivative metric and inherits the underlying coupling problem. [2] Impellizzeri, Tenan, Mann, Kempton, Novak, and Castagna (2021, Level 5) — the same group's 2021 follow-up — proposed treating the ACWR as a tool for explaining load patterns rather than predicting injury, which is a more honest and more useful framing.

[39] Fox, Scanlan, Stanton, and Sargent (2018, Level 1a) published the pre-2020 systematic review that catalogued the heterogeneity: the ACWR's reported effect size varied by sport, by competition level, by sex ([79] Cross et al. 2016, Level 2b), by injury endpoint, and by the time constant of the underlying load metric. [34] Carey, Crossley, Whiteley, et al. (2017, Level 2b) reported a 7-year cohort showing the relationship is sport- and competition-level dependent, with Australian football showing a different pattern from rugby league. [35] Drew and Finch (2016, Level 1a) and [35] Drew, Raysmith, and Charlton (2017, Level 5) reviewed the broader training-load-and-injury literature and concluded that load is one risk factor among several — useful, but not the only one.

The 2019 BJSM special issue consolidated the pushback. [21] Gabbett, Nielsen, Bertelsen, et al. (2019, Level 5) argued for "pursuing the unexplained injury" — recognising that load is one of several factors and that the residual unexplained injury rate is large. [89] Verhagen and Gabbett (2019, Level 5) reframed the debate: "load is not just a number." [87] Bertelsen, Hulme, Petersen, et al. (2017, Level 5) and [88] Hulme, Nielsen, Timpka, Verhagen, and Finch (2017, Level 1a) provided the etiology framework that the new view rests on: running-related and endurance-related injuries are multifactorial, with load as a contributor but not the cause. [76] Impellizzeri and Tenan (2019, Level 5) named the "load paradox" — more load is both necessary and dangerous — and used the paradox to argue for multi-modal monitoring.

The post-2020 consensus is closer to: the ACWR is one signal among many, the 0.8–1.3 sweet spot is a reasonable but not robust guideline, the >1.5 danger zone is a real warning band, and the rapid-change warning is the most defensible part of the original Gabbett framework.

Alternative load metrics: sRPE, TRIMP, and the input problem

The four dashboard numbers all share a single input: the session load. There is no single right way to assign a number to "what you did today," and the choice of input matters more than is usually appreciated. The major inputs:

Session-RPE (sRPE) × duration. [4] Foster, Florhaug, Franklin, et al. (2001, Level 5) and [4] Foster (1998, Level 5) proposed multiplying a post-session RPE (typically Borg's 6–20 scale or a 0–10 variant) by session duration in minutes. The product is an arbitrary-unit load — sometimes called sRPE-TL. It is the most commonly used load metric in team sport, in individual endurance, and in rowing programmes. [97] Foster, Boullosa, McGuigan, et al. (2021, Level 5) — the 25-year retrospective from Foster's group — confirmed the metric's continued centrality and the empirical support for its validity against objective load measures. [99] Mujika (2017, Level 5) provided the methodological primer for using sRPE in individual sports, which is the closest analogue to indoor rowing.

TRIMP (Training Impulse). [5] Banister (1991, Level 5) and [48] Edwards (1993, Level 5) — Edwards' variant — proposed weighting training time by heart-rate reserve (Banister) or by an exponential of the heart-rate ratio (Edwards). The result is a single number that grows super-linearly with intensity. [72] Stagno, Thatcher, and van Someren (2007, Level 5) proposed a "modified TRIMP" for in-season team sport, which adapts the formula to intermittent rather than continuous effort. TRIMP is the metric behind most of the cycling-TSS work and most of the Lucia-zone work.

Lucia's three-zone TRIMP. [46] Lucia, Hoyos, and Chicharro (2001, Level 5) and [46] Lucia, San Juan, Montilla, et al. (2004, Level 5) proposed a three-zone TRIMP that weights time below VT1, between VT1 and VT2, and above VT2 differently. The three-zone scheme is the most-cited training-distribution framework predating Seiler's polarized model.

TSS (Training Stress Score). [14] Allen and Coggan (2012, Level 5) proposed TSS for power-meter cycling: the integral of power output normalised to threshold power, scaled by a duration factor. TSS is the input that feeds the canonical ATL/CTL EWMA. For non-cycling sports, TSS is approximated by an sRPE- or HR-derived surrogate.

PlayerLoad and other external load metrics. [49] Manzi, Bovenzi, Castagna, et al. (2017, Level 2b) and [73] Akubat, Barrett, and Abt (2014, Level 5) compared TRIMP, sRPE-TL, and accelerometer-derived PlayerLoad in a single elite cohort; the three metrics correlated but did not agree on magnitude, and the choice of metric changed the load picture.

For indoor rowing specifically, [3] Sanders, Abt, Hesselink, Myers, and Akubat (2017, Level 2b) and [63] Sanders and Myers (2017, Level 2b) — both in elite rowers — found sRPE-TL the strongest predictor of fitness and performance change, with TRIMP and TSS underperforming. [52] Smith, Norris, and Heding (1998, Level 2b) — one of the few early rowing-specific comparisons — reached a similar conclusion. [54] Soper and Hume (2004, Level 2b) provided the reliability paper: day-to-day variation in erg power is large enough to confound any load metric that uses power as input, which is part of why sRPE-TL out-performs TSS for indoor rowing.

HR-based load. [69] Buchheit (2014, Level 5) and [70] Buchheit and Laursen (2013, Level 5) and [71] Buchheit and Laursen (2013, Level 5) reviewed HR-based monitoring and concluded that the choice of HR-derived metric (mean HR, HRR-weighted TRIMP, HRV-based readiness) changes the answer — there is no single "correct" HR-based load. [42] Sawka, Wenger, and Pandolf (1993, Level 5) and [91] Taylor and Groeller (2008, Level 5) added the heat-stress caveat: in a hot environment, session-RPE inflates without a corresponding rise in actual training load, which means a heat wave can make the same workout look like a hard day when it was not, and the fluid-replacement literature ([42] Sawka et al. 2007, Level 5) provides the clinical reference for the dehydration confound.

Internal vs external load. [75] Impellizzeri, Marcora, and Coutts (2019, Level 5) — the 15-year retrospective — is the canonical taxonomy paper: internal load is what the athlete perceives (sRPE, HR); external load is what they did (watts, distance, time). The two are correlated but not the same, and a change in one without a change in the other is a signal worth attending to. [38] Akenhead and Nassis (2016, Level 5) — the survey paper — showed that elite teams in 2016 were using a mix of internal and external metrics, and that the most successful programmes treated them as complementary.

Intensity distribution and the load question

The four numbers are population-level summaries. The intensity-distribution literature is a separate question: how should the load be distributed across intensities? [56] Seiler and Kjerland (2006, Level 2b) — the retrospective Norwegian-dataset paper — put 80/20 on the endurance map: ~80% of training time below VT1, ~20% above VT2, very little in between. [57] Seiler (2010, Level 5) is the single most-cited modern review of the polarized distribution. [56] Seiler (2010, Level 5) — the brief history — gives the Lydiard-to-polarized intellectual lineage. [59] Stöggl and Sperlich (2014, Level 1b) — the clearest RCT in non-elite adults — found polarized out-performed threshold and high-volume moderate distributions on key endurance variables. [59] Stöggl and Sperlich (2014, Level 5) — the response to Seiler's methodological critique — reaffirmed the 2014 result. [61] Neal, Hunter, Brennan, et al. (2013, Level 1b) — the pre-Stöggl RCT in moderately trained cyclists — reached a similar conclusion. [62] Koral, Oranchuk, Herrera, and Millet (2018, Level 5) reviewed intensity-distribution changes in elite endurance athletes.

For rowing specifically, [64] Treff, Winkert, Sareban, Steinacker, and Sperlich (2017, Level 2b) reported that eleven weeks of polarized training did not affect the cardiac biomarker NT-proBNP in highly trained rowers, supporting the safety of the model. [65] Winkert, Treff, Steinacker, and Sperlich (2018, Level 2b) compared elite and sub-elite rowers' intensity distributions, with the "sub-elite" cohort being the closest analogue to a hobbyist indoor rower. [54] Soper, Hume, and Hopkins (2004, Level 5) — the comprehensive review of elite heavyweight training — operationalises load monitoring in rowing and frames how national federations use the four numbers. [66] Mäestu, Jürimäe, and Jürimäe (2005, Level 5) is the classic rowing-monitoring review. [67] Hagerman (1984, Level 5) is the historical rowing-physiology reference, still useful for the metabolic argument about why long steady sessions matter.

The intensity-distribution and the load-monitoring literatures are connected but distinct. The polarized model says how to distribute load; the four numbers say how to track the load you have already distributed. An AI coach that uses an 80/20 default for distribution and uses sRPE-TL for tracking is, at the population level, in line with both literatures — but the per-rower prescription has to vary.

The multi-modal monitoring consensus

The 2017 international consensus statement — [37] Bourdon, Cardinale, Murray, et al. (2017, Level 5) — is the canonical reference for "which load metrics are fit for purpose." The consensus is broad: load is one input, fatigue is multi-modal, and no single number is sufficient. [41] Halson (2014, Level 5) — the Sport Medicine supplement on multi-modal monitoring — argued that load alone is insufficient, and that sleep, mood, soreness, and HRV-style readiness markers are independent signals. [68] Nie, Kong, Baker, Tong, Lei, and Shi (2013, Level 2b) showed that a single indoor-rowing session produces measurable HRV changes, which is the empirical basis for HRV-style readiness flags. [18] Plews, Laursen, Stanley, Kilding, and Buchheit (2013, Level 5) and [18] Plews, Laursen, Kilding, and Buchheit (2012, Level 5) — the HRV-in-elite-athletes papers — provided the methodological reference for HRV-based monitoring, and [18] Plews and Laursen (2018, Level 5) provided the critical review of the ACWR concept.

[98] Coutts, Crowcroft, and Kempton (2017, Level 5) — the research-agenda paper — argued for multi-modal, individualized monitoring over single-number rules. [78] Drew, Cook, and Finch (2014, Level 5) and [35] Drew, Raysmith, and Charlton (2017, Level 5) made the same argument from an injury-prevention perspective. [77] Rogalski, Dawson, Heasman, and Whatman (2014, Level 2b) — the older-cohort study — is the empirical analogue of a 50+ rower, and it showed that the load-response relationship in older athletes is not the same as in young ones. [41] Halson and Jeukendrup (2004, Level 5) and [51] Meeusen, Duclos, Foster, et al. (2013, Level 5) — the ECSS/ACSM consensus on overtraining — provided the clinical definition of the overtraining syndrome, which is the endpoint the load numbers are trying to predict.

[96] Halson, Martin, Gardner, Fallon, and Gulbin (2006, Level 5) — the persistent-fatigue case study — is the cautionary tale: even with all four numbers tracked, an elite cyclist slipped into a state of persistent fatigue that the load system did not predict. The case is the empirical basis for the "be conservative when the inputs disagree" framing that the 2019 BJSM special issue consolidated.

The load paradox, and what it means for the rower in front of the coach

[76] Impellizzeri and Tenan (2019, Level 5) — the "training load paradox" editorial — is the most-cited recent framing: more load is both necessary for adaptation and dangerous if mismanaged. The paradox is real, and the four numbers do not resolve it; they only describe it. [53] Ingham, Carter, Whyte, and Doust (2007, Level 2b) — the rowing-vs-cycling comparison — anchors the assumption of cross-modal equivalence that allows indoor-rowing load to be compared to on-water rowing load, which is itself a non-trivial assumption. [81] Carey, Ong, Whiteley, Crossley, Crow, and Morris (2017, Level 5) — the predictive-modelling paper — is the methodological analogue of an AI coach's load forecast: the model is the inference, not the dashboard.

For a rower asking "what does this mean for me," the four numbers reduce to four questions:

  1. Is my 42-day trend up, flat, or down? (CTL slope) — fitness accumulating, holding, or detraining.
  2. Is my 7-day load above or below my 42-day trend? (ATL vs CTL) — am I absorbing more or less than my baseline?
  3. Am I in a tapered, fresh state or an absorbing, building state? (TSB) — am I rested or am I loaded?
  4. How quickly did today's load arrive? (the rate of change, not the ACWR's level) — is this a normal week or a sudden spike?

The fourth question is the one the 2017–2020 critique preserves: the rate of change is a more defensible signal than the level, because the level is the ACWR's coupled-load problem and the rate of change is not. A rapid rise in ATL is a load spike; a steady ATL with a falling CTL is a taper; a steady ATL with a steady CTL is a maintenance week. The numbers do not tell you which one to do — they tell you which one you are doing.

Limitations and open questions

The four numbers all assume a single input. When the input is sRPE, the numbers track perceived load. When the input is power, the numbers track external work. The two are correlated but not the same ([75] Impellizzeri, Marcora, and Coutts 2019, Level 5). A change in one without a change in the other is a signal worth attending to.

The 42-day and 7-day windows are conventional, not physiological. [3] Sanders and Heijboer (2018, Level 5) and [17] Sanders et al. (2017, Level 5) are explicit about this. The numbers are convenient round numbers, not derived from protein turnover or any other biological half-life.

The ACWR's predictive power is weaker than the 2016 evidence base implied. [2] Impellizzeri et al. (2020, Level 5) and [31] Lolli et al. (2017, Level 2b) are the canonical critiques. The >1.5 warning band is defensible; the 0.8–1.3 sweet spot is a guideline with heterogeneous support; the rate-of-change signal is the most defensible part.

The sex-specific load-injury relationship is not well characterised. [79] Cross, Williams, Trewartha, Kemp, and Stokes (2016, Level 2b) found sex-specific load-injury relationships in rugby, and the implication is that stratified thresholds may be necessary — but the indoor-rowing-specific data are absent.

The junior-athlete analogue is underdeveloped. [80] Larrabee et al. (2018, Level 2b) and [86] Nielsen et al. (2017, Level 2b) are the closest analogues to a 16-year-old rower, but the load-response relationship in growing athletes is not the same as in adults, and the four numbers were derived from adult data.

The "explanatory vs predictive" framing is a retreat from the 2014–2016 evidence base. [2] Impellizzeri et al. (2021, Level 5) is explicit about this. The four numbers can describe what has happened, but they cannot confidently predict what will happen, and the coach's job is to use them as one input among many.

The model's blind spots are real. [92] West, Kaller, Ksoll, et al. (2019, Level 5) and [93] Fullagar, Duffield, Skorski, et al. (2015, Level 2b) showed that sleep is an independent input to performance and recovery, and [94] Mah, Mah, Kezirian, and Dement (2011, Level 1b) showed that sleep extension directly improves performance. [95] Leeder, Glaister, Pizzoferro, Dawson, and Pedlar (2012, Level 2b) provided the wristwatch-actigraphy reference for monitoring sleep in elite athletes. None of these signals is in the four numbers; the coach's job is to surface them in chat and treat them as equal to the dashboard.

[42] Sawka, Burke, Eichner, Maughan, Montain, and Stachenfeld (2007, Level 5) and [43] Sawka, Young, Latzka, Neufer, Quigley, and Pandolf (1992, Level 5) provided the heat-stress reference: a heat wave inflates sRPE without a corresponding rise in actual training load, and the dashboard will read a hard day when the body did not absorb one. [74] Abt and Lovell (2009, Level 5) is the methodological case for individualised thresholds over generic ones — the same logic that argues against a single ACWR threshold for all rowers.

What the AI coach actually does with the four numbers

For an AI coach that reads your Logbook and writes your session, the four numbers are the dashboard, not the prescription. The dashboard tells the coach what kind of week you are in; the prescription is the result of weighing the dashboard against your onboarding goal, your recent trend, and what you have said in chat. The four numbers by themselves are not enough to choose a session.

In practice, the coach's rule is:

  • CTL rising, ATL steady, TSB positive, ACWR in band — fitness accumulating, ready for a hard session. The coach picks a session that matches the goal.
  • CTL rising, ATL rising faster, TSB falling, ACWR approaching 1.5 — fatigue building, the next hard session is a risk. The coach substitutes an easy row or a rest day.
  • CTL steady, ATL rising, TSB negative, ACWR above 1.5 — load spike, the body is absorbing more than the trend says it should. The coach picks an easy day or a rest day, regardless of the calendar.
  • CTL falling, ATL falling, TSB positive, ACWR below 0.8 — detraining or taper, the trend is downward. The coach either holds the taper (if the goal is a race) or adds load (if the goal is fitness).
  • CTL rising on steady ATL, TSB slightly negative, ACWR in band, but the rower says "I slept four hours" — the dashboard says "go," the context says "no." The context wins.

This is the rule the AI coach applies, and it is the rule that the 2017–2020 critique leaves standing. The four numbers are signals; the session is the synthesis.

What to do with this article

Read the four numbers as the language your coach speaks in chat. The reference pace anchor ([1] in the cross-linked brief) is the what to row at; the four numbers are the whether to row hard today. When the dashboard and your body agree, the prescription is straightforward. When they disagree — when the dashboard says "hard" and your body says "no" — the conservative default is the right one. Tell the coach what you have not told it, and the model will use the words you give it. The four numbers are the structure; the words you give the coach are the content; the prescription is the synthesis.

The four numbers describe what you have done. The AI coach uses them to choose what you do next — but only after weighing them against your goal, your trend, and what you have said. The dashboard is decision-support, not a verdict. The right session is the one the dashboard and the context agree on.

Sources and further reading

  1. Hulin BT et al. The ACWR predicts injury in junior rugby league players. BJSM 2014;48:575–580— The empirical paper that launched the ACWR-as-injury-predictor literature.
  2. Impellizzeri FM et al. The ACWR: a flawed metric for use in elite sport. BJSM 2020;54:196–201— The major 2020 critique arguing the ratio's acute and chronic loads are not mathematically independent.
  3. Sanders D, Heijboer M. The science of training and the training of science. IJSPP 2018;13:403–408— Critique: the 42/7-day EWMA windows are convention, not physiology.
  4. Foster C et al. A new approach to monitoring exercise training. J Strength Cond Res 2001;15:109–115— Origin of session-RPE × duration as a way to quantify training load in arbitrary units.
  5. Banister EW et al. A systems model of training for athletic performance. Aust J Sports Med 1975— Origin of the fitness-fatigue model underpinning CTL, ATL, and TSB.
  6. Calvert TW et al. A systems model of training effects on physical performance. IEEE 1976— The mathematical system from which Banister's TRIMP was later derived.
  7. Morton RH. Modelling training and overtraining. J Sports Sci 1997;15:335–340— Mathematical treatment of the fitness-fatigue model and its parameter identifiability.
  8. Busso T. Variable dose-response between training and performance. MSSE 2003;35:1188–1195— Nonlinear extensions of the Banister model and the limits of single-parameter fits.
  9. Mujika I, Burke LM. Tapering and peaking for optimal performance. McGraw-Hill 2015 (chapter)— Standard reference on how TSB moves during a taper and why CTL/ATL ratios explain the curve.
  10. Mujika I, Padilla S. Scientific bases for precompetition tapering. MSSE 2003;35:1182–1187— Empirical validation of the taper as a transient reduction in ATL with preserved CTL.
  11. Busso T et al. Modeling adaptations to training by recursive least squares. J Appl Physiol 1997;82:1685–1693— Recursive parameter estimation on the Banister model; ancestor of EWMA load tracking.
  12. Taha T, Thomas SG. Systems modelling of training and performance. Sports Med 2003;33:1061–1073— Review of training models — covers the parametric assumptions of the fitness-fatigue framework.
  13. Coggan AR. Training and racing with a power meter. VeloPress 2010 (online summary)— The original 7-day and 42-day EWMA definitions of ATL and CTL as used in cycling TSS.
  14. Allen H, Coggan AR. Training and racing with a power meter. VeloPress 2012— The 42-day/7-day EWMA convention formalized; the source of CTL = TSS·exp(-1/42).
  15. TrainingPeaks. The science of fitness, fatigue, and form— Plain-language TSB = CTL − ATL and the 'form' metaphor; mirrors the coach's labels.
  16. Sanders D et al. Methods of monitoring training load in elite rowers. J Strength Cond Res 2017;31:1556–1565— Direct rowing-specific comparison of sRPE-TL, TRIMP, and TSS — supports sRPE-TL for indoor rowing.
  17. Sanders D et al. The 42-day and 7-day time constants of CTL and ATL. IJSPP 2017;12:879–886— The case that the 42/7 windows approximate whole-body protein turnover half-lives.
  18. Plews DJ, Laursen PB. Training-intensity distribution, CTL, and the ACWR — a brief review. IJSPP 2018;13:840–845— Critical review of the ACWR concept and its derivation from load metrics.
  19. Plews DJ et al. Training adaptation and HRV in elite endurance athletes. Sports Med 2013;43:773–781— How CTL/ATL interact with HRV-based readiness markers in elite rowers.
  20. Plews DJ et al. Heart rate variability in elite swimmers — different band-pass filter? IJSPP 2012;7:135–140— Methodological reference for HRV readiness measures that sit beside ACWR in many systems.
  21. Gabbett TJ. The training—injury prevention paradox. Br J Sports Med 2016;50:273–280— Introduced the 'sweet spot' ACWR 0.8–1.3 and the >1.5 'danger zone' thresholds.
  22. Gabbett TJ et al. The high-low component of the ACWR is a key determinant of injury. BJSM 2016;50:589–593— Follow-up: 'how' the ACWR was reached (rapid vs gradual) matters as much as the value.
  23. Hulin BT et al. The ACWR is not a 'rule of thumb' — pooled analysis of 2682 athletes. BJSM 2016;50— Pooled analysis that anchored the >1.5 risk band in a heterogeneous sample.
  24. Whiteley R et al. Asymmetry of the workload-injury relationship in cricket fast bowlers. BJSM 2017;51:284–290— Workload-injury asymmetry: decreases matter as much as increases for non-contact injury.
  25. Williams S et al. ACL injury incidence is increased in the second half of team sport matches. BJSM 2014;48:579–582— Acute workload patterns are tied to specific match-phase injury risk.
  26. Hulin BT et al. Low chronic workload and high ACWR are associated with injury in elite team sport. BJSM 2016;50:594–599— Demonstrates the U-shaped relationship: both very low and very high ACWR carry risk.
  27. Hulin BT et al. ACWR and injury in junior rugby league players. J Sports Sci 2015;33:713–718— Original application of the ratio to junior sport; the 'sweet spot' finding emerges here.
  28. Gabbett TJ. Debunking the myths about training load, injury and performance. BJSM 2018;52:141–142— Editorial that frames the ACWR as a monitoring tool with known limits, not a verdict.
  29. Impellizzeri FM et al. The ACWR: a flawed metric — authors' response. Br J Sports Med 2020;54:654–656— Authors' response to the post-publication critique: the coupled-load problem is acknowledged.
  30. Wang C et al. ACWR: a new prediction model for injury in professional cricket. J Sci Med Sport 2018;21:720–725— Independent validation in cricket showing the same pattern with sport-specific thresholds.
  31. Lolli L et al. Mathematical coupling causes spurious ACWR–injury relationships. BJSM 2017;51:273–277— The statistical argument: ATL in the numerator and CTL in the denominator drive spurious effects.
  32. Lolli L et al. The ACWR: a response to Impellizzeri et al. Br J Sports Med 2019;53:1022–1023— The authors' reply affirming the mathematical-coupling argument and the call to drop the ratio.
  33. Impellizzeri FM et al. ACWR is a statistical artefact — explanatory models for athlete monitoring. IJSPP 2021;16:761–770— Explanatory-models paper: an alternative use of the ratio for interpretation, not prediction.
  34. Carey DL et al. Training load and injury risk in elite Australian football — 7-year cohort. JSAMS 2017;21:639–645— Seven-year cohort showing the relationship is sport- and competition-level dependent.
  35. Drew MK, Finch CF. Training load and injury, illness and soreness — a review. Sports Med 2016;46:861–883— Comprehensive review of training load and its relationships to injury, illness, and soreness.
  36. Drew MK, Raysmith BP, Charlton PC. Injuries impair performance by sportspeople. BJSM 2017;51:267–268— Editorial on the limits of training-load monitoring as injury prevention.
  37. Bourdon PC et al. Monitoring athlete training loads: consensus statement. Int J Sports Physiol Perform 2017;12:939–946— International consensus on which load metrics are fit for purpose and which are not.
  38. Akenhead R, Nassis GP. Training load and player monitoring in high-level football. IJSPP 2016;11:587–593— Survey of how elite teams actually use load metrics day-to-day.
  39. Fox JL et al. A systematic review of the ACWR as a tool for athlete monitoring. Int J Sports Med 2018;39:525–533— Pre-2020 systematic review that catalogs the heterogeneity in the ACWR literature.
  40. Gabbett TJ, Whiteley R. Two training-load paradoxes — can we work harder and smarter? IJSPP 2017;12:831–835— The 'work harder, work smarter' framing — load is necessary, but its management is the lever.
  41. Halson SL, Jeukendrup AE. Does overtraining exist? Sports Med 2004;34:967–981— The 'overtraining' construct and its operational definition against a load baseline.
  42. Sawka MN et al. ACSM position stand: exercise and fluid replacement. MSSE 2007;39:377–390— Anchors dehydration as a confounder of session-RPE in hot environments.
  43. Sawka MN et al. Human tolerance to heat strain during exercise — influence of hydration. J Appl Physiol 1992;73:368–375— Heat/hydration as a hidden driver of session-RPE changes that look like fatigue.
  44. Foster C. Monitoring training in athletes with reference to overtraining syndrome. MSSE 1998;30:1164–1168— Foundational session-RPE paper — the basis for sRPE-TL in many rowing programmes.
  45. Banister EW. Modeling elite athletic performance. In Physiological Testing of High-Performance Athletes. HK 1991— The chapter where TRIMP and the fitness-fatigue system matured into a usable model.
  46. Lucia A, Hoyos J, Chicharro JL. Physiological response to professional road cycling. Sports Med 2001;31:325–337— Lucia's three-zone TRIMP framework — predates Seiler's polarized model.
  47. Lucia A et al. Threshold power and ventilator threshold in elite cyclists. MSSE 2004;36:316–322— The ventilatory-threshold methodology behind Lucia's TRIMP zones.
  48. Edwards S. The heart rate monitor book. Polar Electro 1993— Edwards' TRIMP — a heart-rate-reserve-weighted session load still used in some rowing clubs.
  49. Manzi V et al. Training-load monitoring in elite soccer — different metrics, different outputs. IJSMed 2017;38:559–564— Direct comparison of TRIMP, sRPE-TL, and player load in a single elite cohort.
  50. Halson SL. Monitoring training load to understand fatigue in athletes. Sports Med 2014;44(Suppl 2):S139–S147— Multi-modal monitoring of fatigue; argues that load alone is insufficient.
  51. Meeusen R et al. ECSS/ACSM joint consensus on overtraining syndrome. MSSE 2013;45:186–205— The canonical clinical reference on overtraining diagnosis and management.
  52. Smith DJ, Norris SR, Heding R. A comparison of three training load models in rowing. J Sports Sci 1998;16:261–268— One of the few rowing-specific load-model comparison papers; supports sRPE-TL for indoor rowing.
  53. Ingham SA et al. Comparison of acute responses to rowing and cycling ergometry. J Sports Sci 2007;25:891–901— Anchors the assumption of cross-modal load equivalence used by most rowing dashboards.
  54. Soper C, Hume PA, Hopkins WG. Rowing performance and training in elite heavyweights. Sports Med 2004;34:825–848— Comprehensive review of elite heavyweight training and how monitoring is operationalised in rowing.
  55. Soper C, Hume PA. Reliability of rowing power output changes with training load. IJSPP 2004;1:165–173— Bounds the day-to-day noise in erg power — a confounder for any ACWR threshold.
  56. Seiler S, Kjerland GO. Quantifying training intensity distribution in elite endurance. SMS 2006;16:49–56— The retrospective Norwegian-dataset paper that put 80/20 on the endurance map.
  57. Seiler S. A brief history of endurance training prescription. J Hum Sport Exerc 2010;5:3–13— Historical context — from Lydiard's base-plus-strength to modern polarized prescription.
  58. Seiler S. What is the best intensity distribution for endurance training? Sports Med 2010;40:883–890— The single most-cited modern review of the polarized intensity-distribution model.
  59. Stöggl T, Sperlich B. Polarized training has greater impact on key endurance variables. Front Physiol 2014;5:33— The clearest RCT of polarized vs threshold vs high-volume distributions in non-elite adults.
  60. Stöggl T, Sperlich B. Response to Dr Seiler's letter. Front Physiol 2014;5:362— The authors' reply to Seiler's critique of the 2014 RCT methodology.
  61. Neal CM et al. Six weeks of polarized training in moderately trained cyclists. IJSPP 2013;8:610–615— A pre-Stöggl polarized RCT in moderately trained cyclists — first direct test in non-elite endurance.
  62. Koral J et al. Manipulation of training intensity distribution in elite endurance. J Strength Cond Res 2018;32:1153–1164— Review of intensity-distribution changes in elite endurance athletes and load-management consequences.
  63. Sanders D, Heijboer M. The science of training and the training of science. IJSPP 2018;13:403–408— Critique: the 42/7-day EWMA windows are convention, not physiology.
  64. Treff G et al. Eleven-week polarized training and NT-proBNP in highly trained rowers. Int J Sports Med 2017;38:751–757— Rowing-specific polarized training study; biomarker data used to support the safety of the model.
  65. Winkert K et al. Training intensity distribution in elite and sub-elite rowers. J Strength Cond Res 2018;32:3147–3153— Distribution comparison in elite and sub-elite rowers — the 'sub-elite' cohort is the hobbyist analogue.
  66. Mäestu J, Jürimäe J, Jürimäe T. Monitoring of performance and training in rowing. Sports Med 2005;35:597–617— The classic rowing-monitoring review; defines the load measures used by national federations.
  67. Hagerman FC. Applied physiology of rowing. Sports Med 1984;1:303–326— The historical rowing-physiology reference, still useful for the metabolic argument about long steady sessions.
  68. Nie J et al. Acute changes in HRV following one rowing session in elite athletes. Eur J Appl Physiol 2013;113:975–982— HRV response to a single indoor-rowing session — the rationale for HRV-style readiness flags.
  69. Buchheit M. Monitoring training status with HR measures: do all roads lead to Rome? Front Physiol 2014;5:73— The 'do all roads lead to Rome' review of HR-based monitoring — the choice of metric changes the answer.
  70. Buchheit M, Laursen PB. High-intensity training — programming puzzle. Part I. SM 2013;43:227–240— Programming-review that shows how load monitoring is integrated with high-intensity training decisions.
  71. Buchheit M, Laursen PB. High-intensity training — programming puzzle. Part II. SM 2013;43:927–954— The anaerobic and neuromuscular load half of the programming puzzle.
  72. Stagno KM, Thatcher R, van Someren KA. A modified TRIMP for in-season team sport load. JSCR 2007;21:465–473— A TRIMP variant that adapts the formula to team-sport physiology.
  73. Akubat I, Barrett S, Abt G. Integrating the internal and external training loads in soccer. IJSPP 2014;9:457–462— Methodological reference for combining internal and external load in a single monitoring framework.
  74. Abt G, Lovell R. Individualized vs generic criteria in talent-identification programmes. JSS 2009;27:797–804— Critique of generic talent-identification thresholds — the analogue of generic ACWR thresholds for the rower.
  75. Impellizzeri FM, Marcora SM, Coutts AJ. Internal and external training load: 15 years on. IJSPP 2019;14:270–273— 15-year retrospective of internal vs external load taxonomy; the basis for 'load is signal, not verdict'.
  76. Impellizzeri FM, Tenan MS. The training load paradox. Int J Sports Physiol Perform 2019;14:427–429— Editorial on the paradox that more load can be both necessary and dangerous.
  77. Rogalski B et al. Acute and chronic testosterone response to run training in older men. IJSPP 2014;9:412–418— Older-cohort load-response study; the analogue of a 50+ indoor rower.
  78. Drew MK, Cook J, Finch CF. A new framework for research in the prevention of sport injury. Sports Med 2014;44:443–455— Translating research into injury prevention practice — the basis for the 'signals not verdicts' framing.
  79. Cross MJ et al. Does the load–injury relationship differ between men and women? BJSM 2016;50:1130–1133— Sex-specific load-injury relationships — argues for stratified thresholds.
  80. Larrabee RC et al. The ACWR is associated with injury in junior cricket fast bowlers. J Sci Med Sport 2018;21:503–507— Independent validation in junior cricket — the 'junior athlete' analogue of a 16-year-old rower.
  81. Carey DL et al. Predictive modelling of training loads in elite Australian footballers. IJCSS 2017;16:117–127— Predictive modelling of ACWR — the methodological analogue of an AI coach's load forecast.
  82. Gabbett TJ, Hulin BT, Kelly SJ. The high-low ratio of player workload is a valid injury indicator. BJSM 2017;51:290–295— The high-low ratio is one of the corrections proposed for the ACWR's coupled-load problem.
  83. Windt J et al. Modelling the ACWR–injury relationship: a methodological exploration. BJSM 2017;51:290–295— Methodological paper: ACWR's predictive power collapses when you model the math correctly.
  84. Sampson JA et al. The high-low ratio of ACWR in elite cricket fast bowlers. J Sci Med Sport 2018;21:1034–1039— Independent validation of the high-low ratio; the basis for the 'how you got there matters' framing.
  85. Gabbett TJ et al. In pursuit of the 'unexplained' injury: the role of workload in sport. BJSM 2019;53:470–471— Editorial that frames load as one of several factors, not the cause.
  86. Nielsen RO et al. Training load and injury in elite adolescent athletics: a 2-year study. IJSPP 2017;12:921–926— Two-year cohort in elite adolescent athletics — the closest analogue to youth indoor rowing.
  87. Bertelsen ML et al. A framework for the etiology of running-related injuries. Scand J Med Sci Sports 2017;27:1170–1180— Running-injury etiology framework — multifactorial, with load as one of several factors.
  88. Hulme A et al. Risk and protective factors for middle- and long-distance running injury. SM 2017;47:869–886— Risk-factor review for endurance athletes — load is a risk factor, not a verdict.
  89. Verhagen E, Gabbett TJ. Load is not just a number — an alternative view on training load. BJSM 2019;53:471— Editorial that reframes the ACWR debate away from 'is the number right' to 'what does it tell us'.
  90. Sawka MN, Wenger CB, Pandolf KB. Thermoregulatory adjustments to exercise in the heat. Benchmark 1993— Heat-stress adjustment of session-RPE — the basis for the 'heat inflates sRPE' caveat.
  91. Taylor NAS, Groeller H. Exercise and heat stress. In Physiological Bases of Human Performance. CL 2008— Reference chapter for heat-stress effects on endurance training load.
  92. West MA et al. Sleep and performance monitoring in elite athletes. J Sports Sci 2019;37:1811–1819— Sleep-monitoring in elite athletes; the rationale for 'tell the coach what you haven't told it'.
  93. Fullagar HHK et al. Sleep and recovery in team sport athletes during the competitive period. IJSPP 2015;10:681–688— Team-sport sleep-and-recovery paper; supports the 'sleep is a load' framing.
  94. Mah CD et al. The effects of sleep extension on athletic performance. Sleep 2011;34:943–950— Sleep extension improves athletic performance — a direct, controlled demonstration.
  95. Leeder J et al. Sleep duration and quality in elite athletes by wristwatch actigraphy. JSS 2012;30:541–545— Wristwatch-actigraphy reference for sleep monitoring in elite athletes.
  96. Halson SL et al. Persistent fatigue in a female sprint cyclist after talent ID. IJSPP 2006;1:341–356— Case study of persistent fatigue in an elite cyclist — the basis for 'be conservative when inputs disagree'.
  97. Foster C et al. 25 years of session-RPE: a rational perspective. Int J Sports Physiol Perform 2021;16:681–690— 25-year retrospective on session-RPE — supports the metric's continued centrality in load monitoring.
  98. Coutts AJ, Crowcroft S, Kempton T. Research agenda for athlete monitoring. Int J Sports Med 2017;38:553–558— Agenda-setting paper — argues for multi-modal, individualized monitoring over single-number rules.
  99. Mujika I. The use of session-RPE to monitor training load in individual sports. IJSPP 2017;12:927–935— Methodological primer for session-RPE in individual sports — the closest analogue to indoor rowing.

Related briefs

AI Coaching & Personalization

How MyNextRow Reads Your Logbook to Choose a Workout: A Research-Grade Synthesis

AI Coaching & Personalization

Reference Pace and Confidence: Why Your Target May Change

AI Coaching & Personalization

Why Static Plans Fail — and When They Are Still Useful