Your watch predicts a 3:43 marathon, a website calculating from your shorter race says 3:55, and a coaching spreadsheet gives 4:06. Three numbers, twenty-three minutes apart. It is tempting to choose the most favourable, or average them and treat the middle as the truth. Yet these are not three votes in the same election. They may answer different questions—or use almost exactly the same information.
An uncertainty range is useful when it shows what could change your decision. That first requires separating input data, the model’s limitations and possible race-day conditions. You then gain more than a wider set of numbers: you gain an understandable explanation of why the expected result remains uncertain.
How many genuinely different data sources lie behind three predictions?
Write down what each estimate is based on. One app may calculate from a previous 10K, another from the best efforts of recent weeks, and a third may use threshold speed. Different screens and brand names do not guarantee different information sources. If all respond to the same single fast run, their agreement is less convincing than it initially appears.
Suppose all three services received the same half-marathon file. Two apply a similar distance-ratio formula, while the third derives a fitness indicator from it. If all three suggest roughly four hours, this does not mean that three independent methods have proved a four-hour marathon. The same starting error may remain in all of them.
Think in terms of families of data. Race results, repeated training observations and standardised laboratory measurements can approach the question from different directions. Even together they cannot guarantee a result, but they offer a more meaningful comparison. The aim is not to open as many apps as possible, but to recognise which new information genuinely adds knowledge.
Point estimate, error range and scenario
A point estimate is a single number, such as 3:55, assigned by a particular model to the input provided. Displaying the number does not require the program to explain the distribution of its errors. Substantial uncertainty can therefore remain behind a very precise-looking display. An interface that draws no range is not evidence that there is no error.
A statistical prediction interval is calculated using specified methods and assumptions. It is not the same as a confidence interval for the studied group’s mean. In a large sample, we may know the group mean relatively precisely while a new individual’s outcome can vary over a much wider range. These two uncertainties cannot be interchanged under the same label.
Your own decision-making range, by contrast, may be a constructed set of scenarios: favourable conditions, the expected situation and a cautious alternative. This is a perfectly usable planning tool if you call it that. You need not claim statistical certainty for it. The lowest and highest of three different predictions do not, on their own, form a ninety-five per cent interval.
A model’s error is not a personal time penalty
A reported average model error belongs to the sample studied and the definition of error. Absolute differences in minutes, percentage differences and squared error are not the same. If one study reports a percentage and another a mean squared difference, the smaller-looking number is not necessarily better. First understand what was averaged, and how.
The study examining critical speed estimated from training data also retained appreciable error in its marathon predictions. This illustrates how a strong association and an accurate individual prediction are separate questions. A database containing many runners can improve understanding of the relationship without establishing which way your next race will deviate.
For illustration, eight per cent of a 240-minute estimate is 19.2 minutes. It is easy to misinterpret this as “I will allow nineteen extra minutes, then”. But the calculation only illustrates the size of the proportion. It does not show that your individual error will be exactly this large, always point towards a slower result, or make an additional nineteen minutes a safe race target.
Enlarge imageWhy is the best fit not enough?
A formula can follow the data used to create it very closely. This does not establish that it works equally well for new runners or another race. If a model is repeatedly adjusted to the same small sample, it may partly learn that sample’s random peculiarities too. This risk is commonly called overfitting.
The research examining marathon prediction during preparation followed sixteen runners at several measurement points. Such research can describe the development of a particular programme in detail. When applying its results, however, it remains important to ask whether they would be similar at another performance level, within another training system or in a differently composed group.
As a user, you do not need to become a statistician. Four questions already help: who was the model built from, what data does it request, how was it checked on new cases, and what error is reported? If these questions have no answers, give the displayed time less weight as evidence. Professional graphics do not replace a documented method.
Data uncertainty starts at the input
Errors in your own input may compound the calculator’s error. Distance or time can be entered incorrectly by hand. Automatic pausing can shorten moving time, while GPS can distort the distance covered. If the model treats these as accurate, it can turn uncertain data into a seemingly precise conclusion.
Consider a simple fictional case. The program sees a 60-minute run and 12 kilometres, so it uses a pace of 5:00 per kilometre. If the actual distance was only 11.7 kilometres, the average becomes about 5:08. One such workout should not support sweeping conclusions, but the example shows how systematic optimism can enter the starting data.
The date of the data matters just as much. Last year’s personal best may be perfectly measured yet describe your present condition poorly. Add a brief assessment of how current and comparable each input is. You need not discard every run that occurred outside a laboratory; you need to understand what each is suitable for.
How should you handle the three different numbers?
Do not average them first. Beside the example’s 3:43, 3:55 and 4:06, record the data source and last update. Then identify which actually uses current information of appropriate quality. An estimate derived from outdated, incomplete input does not become more valuable because it lies near the middle of the three numbers.
Second, look for a common explanation for the difference. One model may require more long-distance history, while another emphasises short, fast efforts. They may also handle terrain or inactive time differently. If you can identify the reason, you have a specific question for a discussion with your coach. That is more useful than staging a competition between program brands.
Third, rank the quality of the data rather than the size of the predictions. A more recent official race in appropriate conditions may provide a stronger reference than an old, manually edited file. This is not an automatic weighting formula. The point is to identify reliability rather than unknowingly give more weight to your desired number.
Three conditional alternatives rather than just numbers
Describe the favourable scenario as more than a faster time. Include conditions such as preparation remaining consistent, current information supporting the target, and the race environment not differing substantially from what was planned. For the more cautious alternative, name the uncertain condition just as specifically. The range will then still make sense in the week before the start.
Another useful distinction is what can and cannot be known beforehand. The course structure and official aid-station arrangements can be checked. A last-minute gust of wind or how the day feels is less fixed. Clarify the former and retain flexibility for the latter. A good plan does not pretend that both are certain.
The cautious alternative need not be the same race merely a few minutes slower. It may mean finishing without a time goal, choosing another event or reconsidering entry. The calculator does not decide this. The decision following from its number must fit your actual background and the role running plays in your life.
Enlarge imageWhen is recalculating worthwhile, and when is it not?
Identify new information that could genuinely change the assessment. This might be a comparable assessment, a new race or a review of several weeks’ training. An unusual easy run often does not qualify on its own. If you ask the future to vote again every day, you are following short-term fluctuations on the display more than your preparation.
When new information arrives, do not look only at whether the estimate has become faster. Uncertainty may also decrease or increase. A stable result might support the previous range without moving its centre. Conversely, a period with repeated interruptions may increase the number of open questions even if an app still displays an unchanged target time.
A decision diary can work instead of constant recalculation. Three columns are enough: what new information arrived, which assumption it affects, and what decision changes as a result. If the last column is always empty, another appearance of precision is unnecessary. If something changes, you will be able to trace the reason later.
A pre-race decision with several predictions
The empirical study of recreational race times also cautions that a model built from shorter distances is not equally useful at every distance. Matching the data to the purpose is therefore fundamental to a marathon decision. Having many numbers does not automatically improve the assessment if none illuminates uncertainty about longer exertion.
Before the start, keep a short sheet with you or in mind: what supports each estimate, which circumstances have changed the situation, and what will prompt a more cautious goal. Do not let a newly flashing watch prediction launch your race. Previously considered conditions are easier to apply consistently than choosing a new model in excitement in the start area.
The value of an uncertainty range ultimately lies in more than its width. It is useful when it separates known facts from assumptions and leaves room for new information. The three marathon predictions then cease to be three competing authorities. They become tools whose conditions you understand, alongside which you can continue making conscious decisions.
Sources
- Esteve-Lanao and colleagues: marathon prediction at several points during preparation
- Smyth and Muniz-Pumares: calculating critical speed from raw training data
- Vickers and Vertosick: empirical analysis of recreational runners’ race times
How Futószív articles are prepared: our editorial guidelines.


