REF-1333 · Emig & Peltonen, 2020 · Nature Communications · Paper · PubMed · Free full text

Does more running volume keep paying off?

In 12,233 Polar users before a marathon, more distance went with a faster modelled aerobic speed, saturating weakly; endurance plateaued, then dipped, at the highest loads. Association only.

Main caveat: Association, one season: training and performance come from the same 180 days.
limited evidenceMeasured on recreational to highly trained athletes

Emig (Université Paris-Saclay) and Peltonen (Polar Electro Oy) took Polar Flow data from runners who logged a run over the marathon distance between July 2015 and December 2018 with the same GPS watch (Polar V800).

For each runner and marathon they defined a season as the 180 days before it, took the fastest efforts over 5 km, 10 km, half marathon and marathon in that window, and fitted the authors' previously published running model, which describes performance with two indices: vm, the crossover speed associated with maximal aerobic power (MAP), and an endurance index El describing how slowly the sustainable fraction of MAP falls with duration.

The model reproduced race times with a mean error of 2.0% in seasons with three or more efforts. The training analysis used 21,605 seasons from 12,233 runners with at least 30 runs (1,616,004 runs, about 20 million km). Season distance went with a higher vm, on average linearly, with a weak saturation trend at the largest distances.

Faster runners ran their training at a lower fraction of their vm (vm fell linearly with mean relative training intensity between 50% and about 90% of vm); a runner with vm = 4 m/s typically trained at 64-84% of vm, one with vm = 5 m/s up to 66%, so both ran most runs no faster than about 5 min/km.

The endurance index rose linearly with season TRIMP, plateaued around El = 7.5 ± 2 at TRIMP ~25,000, and then showed a statistically significant drop, which the authors suggest may reflect overtraining. The authors stress that cause and effect cannot be separated: fitter runners may simply train more.

Our reading: a large, consistent association that fits the base's observational volume–speed findings (REF-197, REF-090, REF-1004) and adds a visible ceiling at very high loads; it is not a trial and gives no individual dose.

Protocol

Data source: Polar Flow web service; anonymised user IDs; subjects who completed a run over 42,195 m between 1 Jul 2015 and 31 Dec 2018 and used the Polar V800 throughout.
Seasons: all runs in the 180 days before each marathon. Racing season = the fastest runs over 5 km, 10 km, half marathon and marathon (±3% GPS tolerance; race speeds computed from the nominal distances), at least two distances, speeds required to rise with shorter distance and stay below world-record speeds. Whether an effort was an actual race could not be determined.Training season = all runs ≥1000 m (none ≥7.8 m/s), kept only with ≥30 runs (at least about weekly running, without long interruptions) and a model fit with mean race-time error below 5%.
Numbers (Table 1): 18,993 subjects available; 14,304 with a model fit including the marathon (24,858 seasons); 6749 without the marathon; training analysis 12,233 subjects, 21,605 seasons, 1,616,004 runs, 19,959,214 km, mean 12.4 km per run.
Model: relative power linear in speed; maximal sustainable power declines logarithmically with duration beyond a crossover time tc = 6 min; two fitted indices, vm (2-7 m/s bounds) and El = exp(0.1/γl) (2.1-13.0); fits by Levenberg–Marquardt, fits at the bounds excluded.
Training variables: volume = season distance; intensity = mean training speed / vm; load = TRIMP summed per run as duration × κ1 (v/vm) exp(κ2 v/vm) with κ1 0.64, κ2 1.92 for men and 0.86, 1.67 for women (heart-rate reserve replaced by v/vm). Bins of 300 km, 0.025 and 2000 TRIMP; SEM and SD by 1000 bootstrap replicates.
Funding and interests: no funding statement; Polar Electro is thanked for support in obtaining the data, which are available from Polar Electro Oy only with its permission; the second author's affiliation is Polar Electro Oy; the authors declare no competing interests. Analysis code (R) on Zenodo.

Key Results

Relation or measureResultSource
Model accuracy, seasons with ≥3 efforts (N = 12,309)mean error 2.0% between predicted and actual race timetext
Marathon predicted from shorter efforts onlyoverall accuracy better than 10% in the most likely parameter range; fastest group (< 160 min) error typically under ±2.5%text, Fig. 4
vm distributionapproximately normal, mean 4.4 m/s; 1-h utilisation peaks at about 82% of MAPFig. 2
vm vs season training distancerises on average linearly, with a weak saturation trend at the largest distancesFig. 5a (no slope or CI given)
vm vs mean relative training intensityfalls linearly between 50% and about 90% of vm: faster runners train at a lower fraction of vmFig. 5b
Typical training rangevm = 4 m/s: 64-84% of vm; vm = 5 m/s: up to 66% of vm (about 5 min/km upper pace for both)text
Endurance index vs season TRIMPlinear rise, plateau around El = 7.5 ± 2 at TRIMP ~25,000, then a statistically significant dropFig. 5c (test not named)
Endurance index vs relative training intensityexponential rise, matching TRIMP's lactate-based weightingFig. 5d
What endurance is worth (model example): at vm = 5 m/s, doubling El from 3 to 6 — 1-h utilisation from 79 to 87% of VO2max in the authors' terms — moves a modelled marathon from 3 h 27 min 38 s to 2 h 53 min 8 s with no change in VO2max or economy.
Plateau in weekly terms (our arithmetic): TRIMP ~25,000 over 180 days is about 970 a week; under the paper's formula that is roughly 8-11 h of running a week if the season averaged 65-70% of vm (men 9.4-11.2 h, women 8.4-9.8 h), assuming duration in minutes (the paper gives no unit) and one average relative speed for the whole season — an order of magnitude, not a threshold.
A number that does not add up: the abstract's "≈ 14,000 individuals with ≈ 1.6 million exercise sessions … ≈ 20 million km" joins two rows of Table 1 — 14,304 subjects belong to the model fit, while the 1,616,004 runs and 19,959,214 km belong to the 12,233 subjects of the training analysis. The text's "~22,000 training seasons" is Table 1's 21,605.

Mechanisms

Measured: associations across runners within a season: more distance with a higher vm; lower relative training speed in faster runners; higher endurance index with more TRIMP up to a plateau, then lower.
Authors' interpretation: they hypothesise that longer training distance improved running economy; faster runners training at lower relative intensity is consistent with low-intensity training improving high-intensity performance; the drop at the highest TRIMP may be due to overtraining, which suggests an optimal training load per season. They also give the reverse explanation: fitter runners with a higher vm may log more kilometres.
Inference: the drop at the highest loads may also be selection — runners who log the very highest loads may differ in age, injury or recording habits — and was not tested against any of these; the figure does not show how many runners sit beyond the plateau.
Inference: "lower relative intensity" here partly reflects how the average is computed: a fast runner's easy pace is a smaller fraction of a faster vm, and stops or intervals with the timer running pull the average down.

Practical Application

Does more running volume keep paying off? In 12,233 Polar users before a marathon, more distance went with a faster modelled aerobic speed, saturating weakly; endurance plateaued, then dipped, at the highest loads. Association only.
Volume and speed go together across runners, with a ceiling: the more a runner ran in the 180 days before a marathon, the higher their modelled aerobic speed, and endurance rose with load until a plateau near TRIMP ~25,000 per season (B, observational, recreational to well-trained marathoners). The plateau is roughly 8-11 h of running a week at typical easy paces by our arithmetic; it is a population average, not a personal limit.
Faster runners ran easier relative to their ceiling: a 4 m/s runner trained at about 64-84% of vm, a 5 m/s runner up to about 66%. Not from this paper; the base's own practice: keep easy work comfortably below the first threshold and keep the middle out of the easy days (SYNTH-51).
Do not prescribe from this: the direction of cause is open — fitter runners may simply train more — and the indices are model fits to self-recorded best efforts. Not from this paper; the base's own practice: for an individual, judge added volume by their own response (resting signs, interval progression, how the week is absorbed) rather than by a population curve (SYNTH-29).
Cyclists: no comparable riding dataset is in this entry; transfer of the plateau to cycling hours is inference, and running minutes cost more tissue than riding minutes (REF-235).

Limitations

Association, one season: training and performance come from the same 180 days; vm could not be measured at the start and end of a season, so no change was tracked (the authors' limitation).
Model-derived outcomes: vm and El are fits of the authors' own model (tc fixed at 6 min) to each runner's fastest GPS efforts, which may not have been races; no laboratory VO2max, threshold or economy.
Only distance and duration per run: no heart rate; average speed includes stops when the timer ran; TRIMP from speed, not heart rate; terrain, weather and surface ignored.
Reported as figures: the saturation and the drop are shown, not quantified with slopes, CIs or counts; "statistically significant" for the drop names no test.
Who: Polar V800 owners who ran a marathon distance between 2015 and 2018; sex split, age and nationality are not reported; selection towards motivated, device-owning runners.
Reporting: the abstract's ≈14,000 runners and ≈1.6 million sessions come from different rows of Table 1 (14,304 subjects in the model fit; 1,616,004 runs from 12,233 subjects in the training analysis).
Interests: the second author is affiliated with Polar Electro Oy, which supplied the data and controls access to them; no funding statement; the authors declare no competing interests.
How this entry was checked and changed
Why this level

a very large observational dataset (n = 12233 runners in the training analysis), which the rubric would place at B+, held one step lower: training and performance come from the same 180-day window, so the relations are associations whose direction the authors call open; the performance indices are parameters of the authors' own model fitted to self-recorded best efforts that could not be verified as races, not laboratory measures (a model-based criterion costs a grade under the base's rule); the volume relations are given as figures without effect sizes or confidence intervals; and the second author's printed affiliation is the platform's maker, Polar Electro Oy. Read from the full text (open access, CC BY 4.0)

Full text (checked 2026-10-28)
Training season: # Subjects 12,233, # Activities 1,616,004, Total distance 19,959,214 km, 21,605 seasons; model fit with marathon: 14,304 subjects, Table 1 (full text)
For all racing seasons with three and more races (N = 12,309), the mean error between model prediction and actual race time was only 2.0% (full text)
As the total training distance increases, vm increases on average linearly, with a weak saturation trend at largest dtrain (full text)
A linear decrease of vm with the mean relative training intensity between 50% and about 90% of vm; a runner with vm = 4 m s−1 typically trains between 64 and 84% of vm, a runner with vm = 5 m s−1 typically up to 66% (full text)
An initial linear increase of El with TRIMP, a plateau around El = 7.5 ± 2 for TRIMP ~25,000, and a statistically significant final drop which may be due to over-training (full text)
Doubling endurance from El = 3 to El = 6 (one-hour utilization from 79 to 87% of VO2max) improves a vm = 5 m s−1 runner's marathon from 3 h 27 min 38 s to 2 h 53 min 8 s (full text)
TRIMP per run with κ1 = 0.64, κ2 = 1.92 for male subjects, and κ1 = 0.86, κ2 = 1.67 for female subjects; season = the 180 days before the marathon (full text)
Affiliation 2: Polar Electro Oy; the support by Polar Electro in obtaining the exercise data is acknowledged; the authors declare no competing interests (full text)