Can AI predict your cycling performance?

AI can turn your rides into forecasts: whether your FTP will rise next block, the chance you’ll nail tomorrow’s VO2 set, or how likely a PR is on your local climb. The promise is real, but so are the limits. Here’s what AI can do with power data, how the models work, and how to make the predictions reliable enough to guide your training.

What AI can (and can’t) predict from your power data

• Short-term readiness: Models combine recent load, sleep/HRV (if you log it), resting HR, and prior session outcomes to estimate the probability you can complete intervals at target watts today or tomorrow.
• FTP and your power–duration curve: Given consistent training and clean data, forecasts over 2–8 weeks can often land within a few percent of actual changes. AI excels at spotting trends in time-in-zone and intensity density that precede gains.
• Breakthrough likelihood: The model can estimate your odds of setting a 5-minute or 20-minute PR in the next 2–4 weeks based on how similar blocks affected you in the past.
• Event-day pacing: With your critical power and W′ estimates, AI can simulate sustainable watts over climbs and flats and flag where you’re likely to overpace.
• What it can’t do: Diagnose illness, guarantee a race result, or replace judgment. It can flag anomalies (e.g., rising HR for a given power), but it won’t explain why without your context.

How the models actually work

Under the hood, most cycling performance models are supervised learning systems trained on your historical rides plus outcomes (tests, PRs, session success/failure). Common approaches include gradient-boosted trees, random forests, Bayesian hierarchical models, and sequence models that handle time-series data.

Features extracted from your rides

• Load and structure: Time in training zones, high-intensity density, weekly ramp rate, and monotony.
• Response signals: HR–power decoupling, cadence patterns, recovery of HR between bouts, and session RPE.
• Capability curve: Critical power (CP), W′, and the full power–duration curve (from sprints to long efforts), including how it shifts week to week.
• Context: Indoor vs. outdoor, temperature, altitude, terrain variability, and sleep/HRV if available.

Targets can be today’s interval completion probability, next-week 5-minute power, or 8-week FTP. Good systems use walk-forward cross-validation (training only on data available up to that date) and per-athlete modeling, because your response patterns are not identical to the group average.

AI doesn’t “discover secrets.” It scales careful analysis, reduces noise, and quantifies uncertainty so you can plan with clearer odds.

Make your data predictive: a rider’s checklist

• Standardize efforts: Anchor each mesocycle with at least one hard effort in three ranges: 3–5 minutes, 12–20 minutes, and a sprint. This stabilizes your power–duration curve without formal test days.
• Device hygiene: Zero-offset your meter, verify crank length, keep firmware updated, and avoid changing devices mid-block. Note indoor trainer power offsets if you use more than one device.
• Reduce artifacts: Filter obvious dropouts, set a coasting threshold (e.g., <10 W), and disable excessive ERG smoothing when you care about kinetics.
• Log context: Add session RPE, sleep quality, illness, and carbohydrate intake per hour. HR and cadence add predictive power; notes explain outliers.
• Program for contrast: Include clearly defined Z2 volume and distinct high-intensity days. Homogeneous “gray zone” weeks make learning harder and blur signal.
• Protect privacy: Back up data, anonymize exports if sharing, and keep control of who can access raw files and personal notes.

Using predictions to plan your next block

Let’s say you want to raise 20-minute power from 280 W to 295 W in eight weeks. An AI assistant can simulate scenarios based on your history and similar athletes, returning a predicted change and confidence interval. Use that to choose a plan, not to chase a single number.

• Set guardrails: Keep weekly ramp <8% of training load, schedule a deload every 3–4 weeks, fuel 60–90 g carbohydrate per hour on key days, and aim for 7–9 hours of sleep.
• Decide with uncertainty: If two plans’ confidence bands overlap, pick the one with fewer constraints and better recovery fit to your life.
• Close the loop: Mark interval success/failure, RPE, and any illness. The model improves fastest when it sees outcomes.

Red flags and limitations

• Garbage in, garbage out: Infrequent maximal efforts, miscalibrated meters, or heavy ERG smoothing will mislead any model.
• Non-stationary life: Injury, heat waves, travel, or bike changes shift relationships. Retrain or reset baselines after big disruptions.
• Population bias: Models trained mostly on one demographic may misestimate others. Push for per-athlete calibration and transparent metrics.
• Indoor vs. outdoor gaps: Your outdoor 5-minute power may exceed indoor by 3–8%. Tag ride type so predictions are context-aware.
• Black-box temptation: A higher predicted FTP isn’t a win if the path requires unsustainable recovery or logistics.

Bottom line

AI can’t ride the intervals for you, but it can turn your watts into better decisions. Clean up your data, log the basics, and use predictions to choose plans with the best odds for your context. Treat the numbers as probabilities, not promises, and you’ll get most of the upside with minimal downside.