Your training data can predict a race time
Medium confidence
With enough training logged, models can estimate your race times within about 2 to 5 percent. That accuracy holds for well-trained runners. It drops for newer runners and for very long ultras, and the evidence so far comes mostly from specific groups and distances.
Why it works
Race times come down to three things the body does. VO2max is how much oxygen your body can use. Lactate threshold is the pace where easy effort tips into hard. Running economy is how much energy running costs at your pace. Your training data reflects all three. Pace, heart rate, volume, and intensity let a model estimate them and predict from there.
The evidence
-
The authors built a simple rule of thumb to estimate a runner's lactate threshold (the pace where fatigue-causing lactate starts building up fast) without lab blood tests. The rule sets the threshold at 60% of a runner's endurance running speed reserve. Tested on data from recreational runners (803 volunteers, with a 48-runner validation set after outlier removal), the heuristic estimated the threshold within an acceptable error range for about 87% of runners, close to the roughly 99% ceiling an ideal estimator could reach. They conclude it is about as reliable as the Dmax lab protocol while being far easier and cheaper to use.
-
Using raw training and race data from more than 25,000 recreational marathon runners, the authors showed that critical speed (the fastest pace a runner can hold in a physiological steady state) can be estimated straight from everyday GPS data. The best model, built from 400 m, 800 m, and 5,000 m efforts, predicted marathon finish time with about 7.67% error. Runners raced the marathon at roughly 84.8% of critical speed on average, with faster runners holding about 93% and slower runners about 78.9%. Starting a marathon above about 87.6% of critical speed was linked to bigger slowdowns later in the race.
-
The authors analyzed wearable-tracker data from about 14,000 runners across 1.6 million sessions and roughly 20 million kilometers. They built a model with two personal numbers, an aerobic power index and an endurance index (how much sustainable power fades as a run gets longer). The model predicted race times with a mean error of about 2%, and it could also pull out physiological markers such as the lactate threshold from ordinary training data.
-
Weekly distance predicts running success only in part and masks large differences in cumulative stress: the same 10 km run produces about 14% more foot strikes and about 6% greater accumulated peak vertical ground reaction force when run fatigued rather than fresh. Pace alone misleads for a similar reason, because identical paces impose different internal loads across runners and across days depending on recovery and daily stress. As a practical alternative that needs no extra hardware, the authors favor duration multiplied by session RPE. Wearables now capture richer external metrics like cadence, tibial shock, ground contact time, and leg stiffness, but their validity for predicting injury stays uncertain; ground reaction force accounts for only 20-30% of peak tibial bone force, with muscle forces the largest contributor, and the acute-to-chronic workload ratio remains a contested frame for injury risk.
-
Ten runners wore three accelerometers (on the foot, the shin, and the lower back) and ran on a treadmill at 13.5 km/h until they could not continue. A model built from all three sensors together predicted time to exhaustion best, explaining about 79% of the variation (R-squared 0.792) using 21 movement indicators. Any single sensor did worse: the foot explained about 63%, the lower back about 57%, and the shin about 56%.
-
In 19 recreational runners tested indoors and outdoors, the authors checked whether heart rate and heart rate variability could stand in for lactate threshold, a common performance marker usually found with blood tests. The links were only moderate: the standard deviation of heartbeat intervals tracked lactate-threshold heart rate at r=0.66, and a heart-rate-variability measure tracked lactate-threshold speed at r=0.70. The conclusion was that these heart-rate measures give useful but imperfect estimates rather than a full replacement for lab testing.
-
The authors built a simple universal model of running performance based on relative metabolic power, using just two personal parameters: a maximum sustainable power and how fast that power fades over longer distances. The model reproduced male and female record times from 800 m up to the marathon with mean errors often well under 1%. It also lets a runner's short-race results predict their longer-race potential.
-
The authors tested whether machine learning could estimate a runner's rating of perceived exertion (a 6-to-20 self-reported fatigue scale) from wearable motion sensors during outdoor running. Across 29 runners, a subject-independent model predicted perceived exertion with an average error of about 1.8 points (root-mean-square error, roughly 12%). A single sensor on the wrist worked best, and combining sensors from several body locations added only small gains.
-
The authors measured body composition, training, strength, and heart-rate recovery in 84 male amateur marathon runners (average age 41) from the 2013 Madrid Marathon. Their best prediction model combined body fat percentage, heart-rate recovery after a step test, and half-marathon time, and it correlated strongly with actual marathon time (r=0.77). Swapping in 10 km time instead of half-marathon time lowered the correlation to r=0.73.
-
Using the HRV4Training smartphone app, the authors collected morning heart rate and heart-rate variability (measured through the phone camera) plus self-reported training from 797 users over 3 weeks to 5 months. They found a strong link between these heart measures and reported training load that held across both sexes and all age groups, with heart-rate variability shifting more than resting heart rate in response to training. The study showed that phone-based training monitoring is workable at scale outside a lab.
-
This classic modeling paper combined published values for three physiological factors (maximal oxygen uptake, lactate threshold, and running economy) to estimate how fast a marathon could theoretically be run. Many combinations produced predicted times faster than the 1991 world record of 2:06:50, and the fastest predicted time was 1:57:58 for a hypothetical runner with a VO2max of 84 ml/kg/min, a lactate threshold at 85% of VO2max, and exceptional running economy. Joyner concluded that a sub-2-hour marathon was physiologically possible, decades before it was approached in practice.
Why we call confidence medium
Several modeling approaches (critical speed, TRIMP-based, and machine learning) have shown they can predict. But most studies cover only specific groups or distances, which keeps the confidence here at medium. Joyner 1991 laid the physiological framework. Recent work from Emig and Peltonen (2020) validates the big-data approach.
Where it applies
Healthy adults
Last reviewed May 25, 2026. See how we score.