The sky no longer stands still for us

The trigger was a time trial. When 176 riders start 60 seconds apart, the first and the last ride about three hours apart. With frozen weather the start order means nothing: everyone rides under the same sky. The work also fixed an older flaw: every race drew a spring climate, July included, and 30% of flat stages started with wind of 5 m/s or more.

Before touching the engine, we measured what a gust is worth, with a 72 kg specialist riding alone on the flat.

The consequence is clear. On a straight road, a few extra metres per second cost minutes, and that is what makes starting early or late interesting. On an out-and-back, the wind cancels out and the effect is small. Rain will matter more once wet corners enter the engine (a separate ticket, right after this one).

The engine now accepts a timeline instead of a single weather: increasing points, the first at time 0, at most 64, each giving the wind and the state of the road at a race hour. On each tick the engine sets the conditions of the moment, and everything else reads them as before: aerodynamics, the descent speed cap, incidents, echelons and their placement draw.

The echelon AI now reads the conditions of the moment rather than those at the start. A crosswind that picks up in the second hour only splits the bunch then, and the AI only reacts then. Once the wind drops to zero, it forgets the placement drawn in the previous echelon: without that, the placement from a first gust would have been reused for the next. Tests cover each case, including a wind that picks up during a time trial (two identical riders starting two hours apart: 739.6 s against 897.2 s, what the physics gives) and resuming a saved race in the middle of a time trial.

The engine reads a day, but someone has to draw it. So that Flanders in March does not feel like Andalusia in August, we started from real data: the hourly ERA5 reanalysis from Copernicus, served by the Open-Meteo archive API. The game is not commercial, which allows this source. Only statistics go into the repository; the raw data stays outside, like the GPX traces.

Each race day of real-2026 has a location: the first point of its route where we had it (153 cases), otherwise the geocoded start town (380 cases), and 64 locations corrected by hand with a written reason for each. That makes 596 locations, which we grouped into six regions.

One instruction framed the whole calibration: the weather can stay approximate to reality, but the game must stay alive, and the weather must not be the same every time. Region and month therefore shift the averages (more wind in March on the Flemish coast, less rain in August in Spain), but the spread from one day to the next stays wide. We set variety floors: even in the calmest region, at least 5% of starts on a wet road (2% in the desert, the only truly dry climate), and a share of days at 5.5 m/s or more. Over ten seasons of the same race you will see calm days, windy days and, almost everywhere, wet days.

Once the start is drawn, the day evolves in 15-minute steps. The wind follows a random walk that returns to its mean, around a daily cycle: it picks up in the early afternoon. Direction moves more widely when the wind is light. The road follows a two-state chain whose transitions depend on local time, and rain peaks between 14:00 and 18:00 in every region. A frontal passage, which turns the wind by 45 to 90° and brings rain, stays rare (0 to 2% of days in the data): we keep a few for variety.

The player never sees the real day. They see a forecast published at the first start: for each race hour, the wind speed (median and a 10% to 90% range), its direction and the chance that the road is wet. Every team receives the same one (pillar 3), and the real timeline appears in no data contract, on no screen and in nothing the AI reads (pillar 4): a test checks it.

For a forecast to be right without knowing the future, we build it as an ensemble. The game draws a situation of the day, then the real day and 100 possible days around it, independent of one another. The gap to the situation of the day grows with lead time; it is zero at the start, since you can see the sky. For each hour the forecast gives the share of possible days where the road is wet, rounded to 5%. Since the real day is one more draw from the same law, when the forecast says 30%, the road is wet about 30% of the time. Decisions taken on the forecast are good on average, never certain.

A forecast that is too good is a false forecast. We compared the ECMWF model’s 06:00 and 12:00 UTC runs with ERA5 over 460 race days of 2026 (11,528 hours). The result: wind is forecast to within 0.9 to 1.0 m/s, with no measurable growth in the error before 12 hours. For rain, when the model clearly announces water, it rains 55% of the time; when it announces dry, 7%.

We tuned the predictable share of the forecast on these values: 0.35 for wind and its direction, 0.40 for rain. The announced wind then differs from the real wind by 0.72, 0.97 and 1.11 m/s at + 2, + 4 and + 6 hours. With a predictable share of 0.75, the forecast gave rain in 82% of the hours announced at 50% or more: far sharper than a real forecast. At 0.40 it gives 75%, still a little sharper than the real model (55%), without losing its reliability. We kept it there.

The forecast is not updated during the race. An hourly “race radio” weather report, with its age, is an idea for later.

This is where the weather gets its meaning. Before a time trial where slots are a choice (a prologue, the opening time trial of a stage race, a one-day race), the screen shows what each slot should cost or gain according to the forecast, with its range, the published start of the other teams’ favourites and the order the staff proposes. The director can keep the proposal, go back to the public order, or build their own. Their order is validated as before: an exact permutation of their starters.

When the order is imposed (a stage that follows the reverse of the general classification, a team time trial), nothing changes: the luck of the weather falls on whoever it falls on, as in real life, and the briefing shows it. The calculation lives in the engine package, not in the screen, which only displays it.

In road races, the effects go through mechanisms that already existed. A crosswind above 3 m/s makes echelons. Rain lowers the descent speed cap by 15%, doubles crashes and multiplies punctures by one and a half. The AI changes in only one place: it reads the conditions of the moment.

The briefing adds a staff reading: the exposed sections of the route where an echelon is possible, with a chance expressed as “7 days out of 10” (the share of the forecast’s 100 days where the crosswind exceeds the threshold, for a head of the race at 42 km/h) rather than an adjective. Neighbouring sections are merged at 5 km, with a minimum length of 3 km: without that, a windy day counted up to 34 sections, against at most 9 now. A wet road is announced for the finale only if it lasts to the finish: a morning shower does not appear. In a time trial, a lone rider makes no echelon, and the reading is absent.

No engine constant changed during the recalibration: the measurements confirmed the settings, or explained their gaps. Here is what they showed, in the order they came in.

Our long flat time trials are simulated 3 to 5% too slow. We hoped the real wind of those days would explain the gap. We pulled the weather of 15 real time trials, at the start time and at three points of their route, from the 0.8 m/s of the UAE Tour (which picks up to 4.3 m/s in three hours) to the 8.9 m/s of the Tour Down Under prologue, with 7.5 m/s due west at Valencia. Then we replayed the bench.

Over 600 time-trial days of real-2026, we measured the weather gap between a team’s best and worst slot, for the reference rider. Then we compared it with the same measurement on the real weather (ERA5) of the same places and dates, over the years extracted.

In a grand tour, riders eliminated outside the time limit average 18.9 with the old weather, 18.3 with an evolving day and 18.2 with the climate. Mountain breaks of more than ten minutes go from 95 to 104, then 96. The five career seasons of the reference run pass. Nearly one race in two sees its weather change, without the averages moving.

One gap remains, and the weather is not its cause: in the mountains and on cobbles, favourites win 85% and 87% of the lab’s races, above the 80% target. With the old weather it was 83% and 85%. The evolving day adds 0.6 to 2.6 points, within the noise. The gap comes from the content’s hierarchy, where a clear number one often wins. We raised the ceiling to 85% for those two terrains.

• Rain costs nothing on the flat. It only slows descents and raises crashes. Wet corners come with the next ticket, and a slot’s cost ignores them until then.
• No local showers. At a given hour the weather is the same over the whole course. No gusts, heat or fog.
• The forecast does not move during the race, and each stage of a stage race draws its own day, with no memory of the day before.
• The AI reacts to wind once it blows. It does not position itself before an announced exposed section, and has no wet-descent caution order.
• A slot’s cost ignores the value of starting late. Seeing the sky before the others has an information value that neither the staff’s rule nor the screen counts.
• The echelon reading assumes a head of the race at 42 km/h over the whole route, whatever the relief. A time trial’s wet road only looks as far as the last start.
• The real wind of a time trial comes from a single 0.25° ERA5 cell at the midpoint. The wind of a valley, a pass or a seafront is not in it, and light rain is too frequent. To finish the extraction in a day, 213 cells cover 10 years (2016 to 2025) and the rest only 5 (2021 to 2025): the statistics of the northern regions, the most affected, have wider intervals.
• Riders eliminated outside the time limit on the first stage remain more numerous than aimed for (median of 15 to 17 against a target of 0 to 6), before and after the weather: a drift that comes from elsewhere, to be picked up with the time limits.
• Old races do not load. A race saved before version 10 is refused, with no migration.