Model cards
Every model behind a TELOS number: what it sees, how it was tested, how it did against the obvious alternative, and what it can't do. Tests are "walk-forward": a season is always predicted using only the seasons before it. Including the one that didn't make the cut.
Race model LIVE · PRE-RACE PICK
Picks the winner of every Grand Prix from three things known before the start: qualifying position, the team's recent form and the driver's recent form. Their weights are fitted to history (a conditional logit); podium and points odds come from simulating the race thousands of times. It makes the official pick on the dashboard, /accuracy, the Instagram card and the simulator's starting point.
| Test 2019-2026 | Winner called | Log-loss ↓ |
|---|---|---|
| TELOS race model | 58.7% | 1.189 |
| "The pole-sitter wins" | 55.7% | 1.445 (grid-slot odds) |
Limits: no weather, strategy, upgrades or track-specific car strength. In a season where the pole-sitter nearly always wins, it is no better than picking the pole.
Race model v2 DID NOT PASS · NOT USED
We tried to beat the live model with more information: the gap to pole in qualifying, grid penalties, team reliability, how much the grid matters at each circuit, and the Driver Rating's driver and car strengths. To stop ourselves cherry-picking, the choice of model for each season was made using only earlier seasons, then scored on the season it had never seen.
| 2019–2026 (167 races) | Winner called | Log-loss ↓ |
|---|---|---|
| Richer model (selected blind each season) | 59.9% | 1.2313 |
| Live race model | 56.3% | 1.1917 |
It called more winners but was over-confident when it was wrong (2024 in particular), so its probabilities were less trustworthy. The difference is within noise (95% interval -0.0257 to 0.1044), so the simpler model stays. The next attempt adds practice long-run pace from the official timing archive.
Driver Rating v3 LIVE · /RATING
Separates the driver from the car. Every race and qualifying order is explained as driver skill (allowed to change each season) plus the car's level that season plus how the car suited that track. It is a Bayesian rank-ordered logit (the approach of van Kesteren & Bergkamp, 2023), fitted with MCMC so every rating carries its uncertainty. Races count half as much as qualifying; in a race a skill gap weighs 0.791× and a car gap 0.546× their qualifying effect.
| Teammate duels 2023–2026 (1271) | Called right | Log-loss ↓ |
|---|---|---|
| Driver Rating v3 | 67% | 0.604 |
| Previous teammate-only model | 67.9% | 0.6053 |
Settings chosen on 2019–2022 only. On the old model's home ground — calling teammate duels — it is level; what it adds is a car rating and comparisons between drivers who never shared a garage. Limits: an upgrade given to one car only is credited to that driver; drivers with one career teammate are measured almost entirely against that teammate.
Simulator effects /SIMULATOR · /TITLE-RACE
How much each lever in the simulator is worth, measured on history rather than chosen: starting P1 instead of P2 multiplies a driver's win odds by 2.7×; the car and driver effects come from the Driver Rating above. The championship simulator and the Title Race odds use the same "future race" model (driver skill, car, team form, driver form).
Limits: the chaos dial is a randomness control, not a weather or safety-car model; there are no track-specific effects yet.
Live pre-race record, scored after every race: /accuracy. Questions about a number: contact.