# Consistency analysis and lap-time variation

> Race pace is won by consistency. Quantify it with lap-time statistics, sector reports and per-corner metrics, separate it from fuel and tyre effects, and recognise the data signatures of fatigue.

Area: Driver Analysis · Starting level: L2 Practitioner · Updated 2026-11-16
Source page: https://race-engineer-academy.vercel.app/lesson/driver-consistency-analysis

## L1 · Foundation

**Fast once vs fast every lap.** A driver can beat the competition in qualifying but lose over a race because of larger lap-time degradation and scatter. Consistency is therefore a strategic input: who starts, who does the night stint, who is put on old tyres.

**Simple tools:**
- **Average and standard deviation** of lap time over clean laps. Segers' endurance example: driver Blue averaged 2:07.8 with σ = 0.8 s; Red 2:06.1 with σ = 0.6 s.
- **Sector report with highlighting**: the best sector time per segment highlighted, and all times within a tolerance (Segers used 2 %) shaded. More shading = more consistent. A 2 % window is generous (on a 94 s lap, 2 % is 1.9 s); for pros a few tenths per sector is the real target.
- **Run chart of lap times** with in/out laps, safety-car and traffic laps marked.

**Typical clean-air consistency** (σ of lap time over a stint, after fuel/tyre trends are removed): top professionals ~0.15–0.3 s; good amateurs 0.4–0.8 s. Multi-class traffic in endurance races inflates raw σ massively – always clean the data first.

## L2 · Practitioner

**Separate the three sources of lap-time variation (Segers §16.2):**

1. **Fuel load** – lap time improves as fuel burns. Lap simulation or experience gives a sensitivity, e.g., ~0.014 s/kg for Segers' touring car (~103 s lap); 0.02–0.04 s/kg is common for GT and single-seaters on 4–5 km tracks.
2. **Tyre degradation** – grip falls (after an initial warm-up/peak phase), and balance often drifts.
3. **Driver** – skill, consistency, fatigue, adaptation to changing grip and balance.

**Fuel-corrected lap time:**
$$t_{corr}(n) = t(n) - k_{fuel}\, m_{fuel}(n)$$
Plot $t_{corr}$ vs lap: the trend is tyre degradation (plus track evolution); the scatter around the trend is the driver (plus traffic). Fit a line or low-order polynomial and compute the residual σ – that is a fair consistency number.

**Per-corner consistency.** Lap-time σ hides *where* inconsistency comes from. Compute per lap: brake point at the key corners, minimum speed, throttle pickup distance, curvature peak. Their σ shows which corners are unstable. Segers shows average curvature per lap for two endurance drivers: the driver with more scatter and higher average curvature was also slower – line consistency and lap-time consistency go together.

**Fatigue signatures:** missed or slow gearshifts, changes in throttle-blip quality, earlier brake points, increasing corrections, more track-limit excursions – typically appearing late in long stints or in heat. Compare the first and last 10 laps of a stint per metric.

## L3 · Advanced

**Robust statistics for race data.** Lap-time distributions are skewed (traffic and mistakes only add time). Use:
- **median** and **interquartile range** instead of mean and σ,
- or **trimmed** statistics (exclude laps > 101–102 % of a rolling median),
- or a percentage-of-laps metric (share of laps within 0.5 % of the stint's fuel-corrected best).

**Modelling a stint.** The lap-time model is

$$t(n) = t_0 + k_{fuel} m_{fuel}(n) + d\,n_{tyre} + \varepsilon_n$$

but within a single stint on one tyre set, fuel mass ($m_0 - b\,n$) and tyre age ($n$) both change linearly with lap number. The regressors are perfectly collinear: the data only determine the net slope $d - b\,k_{fuel}$, and any $(k_{fuel}, d)$ pair with the same net slope fits equally well. A free fit returns arbitrary, unstable coefficients. So **fix $k_{fuel}$ beforehand** – from lap simulation (Segers takes it from a lap-simulation weight sweep), a fuel-only test, or cross-stint data that breaks the collinearity (e.g., a short-fuel run on old tyres) – then regress $t - k_{fuel} m_{fuel}$ on tyre age. Add a warm-up term for the first 2–3 laps or a quadratic term for late-stint drop-off only once $k_{fuel}$ is fixed. $t_0$ is then the driver's pace on fresh tyres and empty tank, $d$ the degradation rate and $\sigma_\varepsilon$ the consistency. Comparing drivers on $t_0$, $d$ and $\sigma_\varepsilon$ is fairer than comparing average lap times from stints run in different conditions.

**Degradation is partly driver-induced.** Segers' Dubai example: three drivers on the same car and compound; driver 2 had the highest average understeer angle (0.97° vs 0.81–0.85°) and was 1.4 s slower per lap – either suffering more understeer or causing it through style (e.g., overdriving entries). Run charts of balance metrics per driver help separate "car changed" from "driver drove it differently".

**Using consistency in strategy.** A consistent driver allows tighter fuel and tyre planning (smaller safety margins), more accurate pit-window predictions and confident undercut calls. In Pro-Am formats, minimum drive-time rules force the slower driver to drive; strategy often places them in the stint with the least traffic or most predictable conditions.

## L4 · Expert

**Pooling stints: a mixed-effects model.** Across a weekend, fit all clean laps of all drivers together:

$$t_{ij} - k_{fuel} m_{fuel,ij} = \mu + \delta_{driver(i)} + d_{c}\,n_{tyre,ij} + u_{stint(i)} + \varepsilon_{ij}$$

Driver offsets $\delta$ and compound degradation $d_c$ are fixed effects; a random intercept $u$ per stint (or session) absorbs track evolution, temperature and wind. Driver differences are then estimated *within* comparable conditions instead of being confounded with them.

**Identifiability is a design problem, not a statistics problem.** Any two effects that always move together cannot be separated by any fit:
- fuel and tyre age within one stint (fix $k_{fuel}$, or run a short-fuel stint on used tyres);
- driver and track state, if driver A always does the first stint on a green track and B the evening stint (swap the order at least once);
- driver and tyre set, if the slower driver always gets the older tyres.
Plan the running so the effects you need to compare are not aligned.

**How big is a real difference?** With residual σ = 0.4 s and 15 clean laps each, the standard error of a difference in mean pace is ≈ $0.4\sqrt{2/15} \approx 0.15$ s, so the 95 % interval is ≈ ±0.3 s – a 0.2 s "gap" between drivers is not established. Lap residuals are also autocorrelated (traffic episodes, tyre-temperature phases), which shrinks the effective sample size further. Use a **block bootstrap** (resample runs of consecutive laps, refit, repeat ~1000 times) to put honest confidence intervals on $\delta$, $d$ and σ.

**Limits of the model.** Linear degradation is a simplification: real tyres show warm-up, a plateau, then a thermal or wear cliff, and the shape depends on the driver's management. Traffic residuals are skewed, so use robust regression (Huber loss) or trimming before fitting. Always plot residuals against lap, tyre age and track temperature – structure left in the residuals means a missing term, not driver inconsistency.

## Key formulas

**Fuel-corrected lap time**

$$t_{corr}(n) = t(n) - k_{fuel}\, m_{fuel}(n)$$

- k_{fuel}: lap-time sensitivity to fuel mass (s/kg)
- m_{fuel}(n): fuel mass on lap n (kg)

**Stint lap-time model (k_fuel fixed beforehand)**

$$t(n) - k_{fuel}\, m_{fuel}(n) = t_0 + d\, n_{tyre} + \varepsilon_n$$

- t_0: base pace (s)
- k_{fuel}: fuel sensitivity from lap sim or a dedicated test (s/kg) – not fitted on the same stint
- d: degradation rate (s/lap)
- \varepsilon_n: residual (driver consistency + traffic)

**Standard deviation of lap times**

$$\sigma = \sqrt{\frac{1}{N-1}\sum_{i=1}^{N}\left(t_i - \bar{t}\right)^2}$$


## Trackside takeaways

- Clean the data (in/out, SC, traffic laps) before computing consistency.
- Fuel-correct lap times; the trend is degradation, the scatter is the driver.
- Use per-corner σ (brake point, min speed, pickup) to find where inconsistency lives.
- Use median/IQR or trimmed statistics – lap-time distributions are skewed.
- Fatigue shows as gearshift errors, earlier braking and more corrections late in stints.

## By car category

### GT3

Pro-Am and Bronze drive-time rules make per-driver consistency a strategy input; the Bronze driver’s σ often matters more than the Pro’s best lap.

### LMP2

Multi-class traffic dominates raw scatter; use sector-based metrics in clean sectors and trimmed statistics.

### Hypercar (LMH/LMDh)

With energy-limited stints, consistency in energy use per lap is as important as lap time; deviations can force lift-and-coast later.

### FIA F2

Feature race tyre management: consistent degradation management lets the team time the mandatory stop precisely.

### IndyCar

Fuel-saving targets per lap demand consistent lift patterns; data shows who can hit a fuel number while keeping lap time.

### One-make Cup

Sprint races are short; consistency under pressure (defending, attacking) is visible in brake-point scatter during battles.

## References

- Segers, Analysis Techniques for Racecar Data Acquisition, 2nd ed. (SAE 2014): §16.2 Lap Time Variation During a Race (§16.2.1–16.2.3)
- Segers, Analysis Techniques for Racecar Data Acquisition, 2nd ed. (SAE 2014): §14.8 Driver Consistency over Multiple Laps
- Segers, Analysis Techniques for Racecar Data Acquisition, 2nd ed. (SAE 2014): §14.7.2 How to Determine the Correct Driving Line (average curvature per lap and lap-time σ)
