Japanese Grand Prix 2026

FP2 long-run called 26% of race pace ranks correctly; HAM was the exception

• ›5 of 19 drivers finished within 3 positions of their FP2 long-run rank
• ›HAM: FP2 ranked P18, race result P2 (-16 positions)
• ›Best prediction match: HAD — FP2 P9, actual P10

FP2 long-run data is the strongest single predictor of race pace (academic correlation ρ ≈ 0.67 vs race, strongest of any practice session). At 26% of the field within 3 positions, the model's practice-session intelligence is working as designed.

HAM's divergence from their FP2 signal is worth investigating: a meaningful gap between Friday pace and Sunday result suggests either a setup change between FP and race, a significant fuel load difference in long runs, or conditions that changed between sessions.

The model predicted SAI P20; they finished P8 — this is the model's known bias, not a surprise

• ›Season average delta: better than predicted by 3.0 positions across 22 rounds
• ›This race: predicted P20, actual P8 (Δ -12.0)
• ›Bias direction matched: the model consistently over-rates SAI's race-day performance at this level

SAI's rolling form shows a persistent better-than-predicted trend. The FP long-run data the model uses captures raw pace, but it doesn't fully capture how SAI converts qualifying position to race result under pressure — a trait that shows up reliably at race end.

The model has been correcting for this bias since it was identified. Until SAI's race performance converges with their qualifying pace, their predicted positions will be treated as an upper bound rather than a central estimate.

The model predicted PER P19; they finished P12 — this is the model's known bias, not a surprise

• ›Season average delta: better than predicted by 3.1 positions across 22 rounds
• ›This race: predicted P19, actual P12 (Δ -7.0)
• ›Bias direction matched: the model consistently over-rates PER's race-day performance at this level

PER's rolling form shows a persistent better-than-predicted trend. The FP long-run data the model uses captures raw pace, but it doesn't fully capture how PER converts qualifying position to race result under pressure — a trait that shows up reliably at race end.

The model has been correcting for this bias since it was identified. Until PER's race performance converges with their qualifying pace, their predicted positions will be treated as an upper bound rather than a central estimate.

The model predicted RUS P3; they finished P9 — this is the model's known bias, not a surprise

• ›Season average delta: worse than predicted by 2.4 positions across 22 rounds
• ›This race: predicted P3, actual P9 (Δ +6.0)
• ›Bias direction matched: the model consistently over-rates RUS's race-day performance at this level

RUS's rolling form shows a persistent worse-than-predicted trend. The FP long-run data the model uses captures raw pace, but it doesn't fully capture how RUS converts qualifying position to race result under pressure — a trait that shows up reliably at race end.

The model has been correcting for this bias since it was identified. Until RUS's race performance converges with their qualifying pace, their predicted positions will be treated as an upper bound rather than a central estimate.

ANT was quickest through the spoon complex, the biggest sector gain of any driver in the comparison group — over 1.52s ahead of the field median.

ANT improved consistently as the session progressed, benefiting from an evolving track surface — their best effort came as conditions peaked.