Will tomorrow’s ride feel harder after lifting?

Abstract

A recovery chart is useful only when its meaning can be connected to an observable outcome. For cyclists combining endurance and strength training, that outcome need not be a universal percentage of recovered capacity. We propose starting with a narrower question: given information available today, how much effort will a familiar cycling workload require tomorrow compared with the rider's usual response? Published work supports investigating this question but does not establish that it can be predicted usefully in ordinary training. We distinguish task response from physiological mechanism, a forward forecast from a retrospective fit, and prediction from the effect of changing a programme. The proposed Good Legs research programme will test whether detailed strength execution and exposure history improve next-day prediction beyond a strong benchmark containing personal history, cycling, basic gym workload and available recovery information. Success requires prospective accuracy, informative uncertainty and useful coverage within athletes. A valid negative result would identify the limits of the proposed feature rather than justify a more complicated recovery score.

1. Start with the question a rider can recognise

A cyclist does not need an abstract number to know that some mornings feel different. What needs testing is whether a model can anticipate the difference before the next effort happens.

The proposed question is:

Given your recent riding, lifting and recovery, how much harder or easier than usual is tomorrow's familiar cycling effort likely to feel?

Three parts of the question do important work. A familiar effort anchors the forecast to a specified task. Compared with usual anchors it to the same person's history. Tomorrow establishes that the prediction must be made in advance.

Without those constraints, a model could appear successful by identifying that intervals are harder than easy riding, that riders have different habitual ratings, or that a difficult workout has already happened. None would establish the intended day-to-day forecasting value.

A useful first outcome is whole-body perceived effort during a predefined submaximal cycling segment. Measured power establishes the work actually performed. Heart rate, cadence, local sensations and successful execution can provide additional observations, but they are not interchangeable with the primary outcome.

This definition is deliberately narrower than being ready to race, generating maximal sprint power or having repaired a particular percentage of muscle tissue. A familiar submaximal effort can be useful without answering all those questions.

2. What the existing evidence supports

A cyclist-specific meta-analysis, first published online in July 2025 and appearing in a January 2026 issue, included 17 studies and 262 participants, including 60 women. It found benefits for cycling efficiency, anaerobic power and pooled cycling performance, while rating certainty as low. It did not identify a robust optimum for implementing strength training. [1]

A separate experiment in ten resistance-untrained male cyclists compared two resistance-training bouts and assessed responses before, 24 hours after and 48 hours after lifting. Soreness was lower after the second bout. Several submaximal and jumping measures were disturbed, while overall cycling time-trial performance did not change significantly. A small nonsignificant difference is not evidence of equivalence. The study supports investigating task-dependent responses, not assigning a universal recovery clock. [2]

Submaximal cycling measurement also has precedent. The Lamberts protocol was evaluated in 17 trained male cyclists and used stages controlled by predetermined heart rates. Its reported reliability belongs to that protocol and population. A shorter fixed-power Good Legs segment would be a different procedure and needs its own measurement evaluation. [3]

These findings establish a reason to investigate, not evidence that Good Legs already forecasts the response. They also set limits on the proposed novelty. Repeated-bout protection, subjective effort and submaximal testing are not new discoveries. The potential contribution is a clearly defined, prospectively evaluated combination of detailed strength information and subsequent cycling response in normal product use.

This perspective draws on targeted reading of directly relevant studies and methodological sources. It is not an exhaustive systematic review, a new meta-analysis or proof that no equivalent work exists elsewhere.

3. Four things that should not be called the same thing

The proposed Recovery Curve belongs initially to the third category. It may inform the fourth, but forecasting accurately does not prove that a suggested change helps.

For example, a model might anticipate a more effortful ride after a particular training sequence. The rider and coach may have chosen that sequence because of racing, work or existing fatigue. Even an extensive observational record does not reveal the outcome that would have occurred after every alternative schedule.

That is why the first study does not claim that moving a gym session will improve a taper. A later, targeted intervention comparison could test an acceptable decision within Good Legs. It would be a separate question, not an interpretation automatically supplied by the forecast.

4. Personal history should earn its place

The full model will not be compared with an opponent that knows nothing about strength training. Its benchmark should already contain the rider's previous responses, the intended cycling task, recent cycling, basic strength-session workload and relevant recovery observations available at the prediction cutoff.

The detailed candidate can then add information such as actual exercises, completed sets, load conventions, session timing and recent exposure history. This tests whether those details add value beyond ordinary monitoring.

Individualisation is also a hypothesis. A different curve for every rider can fit historical noise. We should compare individualised and population approaches that have access to comparable information. A simple shared model with a personal baseline may be more reliable early in use than a separate complex model for each person.

Recent exposure to an exercise could prove useful. It might reflect familiarity, adaptation or a correlated change in training. A predictive association would not, by itself, identify a new biological state called strength tolerance. The first obligation is to show reproducible forecasting value, not to name every fitted parameter as physiology.

5. An attractive historical curve is not enough

A 2025 study of fitness-fatigue modelling in two short-track speed-skating datasets reported parameter-identifiability difficulties and no reliable predictive improvement from adding the fatigue component in its analyses. Its findings should not be generalised to every training model, but they illustrate why physiological plausibility and a good historical fit are insufficient tests of a forecast. [4]

The Good Legs comparison should therefore preserve the original prediction, its model version and the information actually available when it was issued. A workout uploaded later can improve a later forecast; it cannot improve the past one. The same applies to subsequent sleep, corrected thresholds and measurements collected during the target ride.

A second obligation is to test changes within a rider. Strong pooled correlation can be produced by recognising that people differ in their usual response. The companion methods paper demonstrates that failure mode with synthetic data. A benchmark that already knows the rider's usual response makes the real question harder and more relevant.

6. What the chart should communicate

A research-backed curve needs an identified variable, a reference task, a time horizon and a definition of uncertainty. The initial quantity is a predicted effort response, reported relative to an explicitly defined personal reference. It is not a percentage of recovered muscle.

The chart should distinguish previous observations from predictions that were actually issued in advance. It should identify missing evidence and uncertainty rather than connect every day with an equally confident line. Changes in the planned week should create updated forecasts, not rewrite earlier predictions.

A forecast of the median response is also different from a prediction interval for one future observation. Confidence in an average model advantage is different again. A shaded region should state which of these quantities it represents.

A multi-day display creates an additional obligation. Accurate next-day predictions do not establish accurate seven-day predictions. Intervals that work separately at each horizon do not automatically cover the whole future trajectory at their stated level. The companion methods paper provides a transparent probability example of that distinction.

The safest initial product claim would therefore identify the tested task, horizon and population. More ambitious claims should follow new evidence rather than a more elaborate screen.

7. Tapering is a destination, not an inherited claim

A taper is an attractive eventual application because riders care about the timing of their response near important events. The first study could examine a prespecified taper subgroup, provided sufficient observations exist, without changing its primary question.

However, normal perceived effort during a submaximal segment does not establish peak race performance. The event may demand sprinting, repeated high-intensity work or endurance after several hours. A model developed during regular training might also behave differently when training is reduced.

The next defensible step would be to evaluate forecasting during taper periods and against suitable event-relevant observations. Only after that should an intervention test ask whether using the curve to alter taper decisions improves outcomes. This sequence protects the distinction between monitoring, prediction and optimisation.

8. What would count as progress before participant data?

The programme can already define outcomes, audit measurement origins, test forecast timing, challenge statistical methods with synthetic data and publish the intended evaluation. These are substantive research activities, but they are evidence about the method, not about athlete recovery.

Progress also includes stating how the feature could fail. Detailed strength data might not improve a strong benchmark. Useful prediction might be confined to unfamiliar lifting or returns after a break. The outcome could vary too little, or too noisily, for the proposed display. Those findings would narrow or stop the feature rather than justify changing the endpoint after seeing a disappointing result.

Reporting guidance such as TRIPOD+AI offers useful principles for describing model development and evaluation. Its clinical setting means the framework must be adapted rather than presented as certification of a sports product. [5]

Conclusion

The first Good Legs Recovery Curve should answer a modest question well: whether a familiar amount of cycling will feel different from usual tomorrow. Its potential value comes from connecting actual strength training with a measurable subsequent response and demonstrating that the extra information improves a genuine advance forecast.

A meaningful chart does not need to claim a universal recovery state. It needs a definition, a fair test and a willingness to remain simple when the evidence requires it.

Declarations

This working paper is prepared for a commercial product, Good Legs. The company has an interest in the proposed feature's success; that interest does not establish scientific validity. No Good Legs participant results, institutional endorsements or independent review are claimed.