Data, sources, and methods

How the routes were reconstructed

Confidence varies by stage. Start and finish locations, and average speed, come from reliable historical sources and are treated as correct. Documented intermediate town names, available for most stages, further narrowed down the road actually taken. For roughly 25 stages we could not find a roadbook, mostly team and individual time trials, prologues and split half-stages, several of them run on closed circuits. For these only the start and finish are certain and the route is a plausible reconstruction rather than a confirmed path; none are missing outright.

Start times are less certain than routes. About three-quarters of stages before 2012 take their start time from an official roadbook; the rest are estimated from each stage’s era and length. Stages in the 1950s to 1970s were often very long and could start as early as 7am, while later editions were increasingly timed to finish in the evening once TV broadcasting made that desirable, a shift that also pushed stage distances down compared to the earliest Tours.

For 1950 to 2011 stages, location and movement of the peloton were estimated using a simple gradient-dependent speed model, calibrated on the stage start times and average speeds. From 2012 onwards the rider recordings already provide position over time directly.

How stage weather was reconstructed

For each stage, we have the peloton’s longitude, latitude and elevation at every point in time, typically every few seconds, throughout the stage. The weather data we use (see below) consists of temperature, humidity, wind and radiation information on a 0.1° × 0.1° grid at hourly time resolution. To approximate the weather experienced by the riders, we resample the track to 30-second timestamps and find the model grid cell that contains the peloton’s position for each of these timestamps. We then take the weather information at the two model hours before and after the peloton’s actual passing time, linearly interpolate all weather values between these two times and extract the values at the peloton’s passing time.

Because of the resolution of the model grid, mountains and valleys cannot be represented accurately in the model data — in other words, the grid cell in the model might have a different elevation than the road the peloton was actually on inside that cell. Since temperature depends strongly on elevation, we correct the extracted temperature values to match the peloton’s actual elevation, using a lapse rate of -6.5°C per 1'000 vertical meters. The same correction is applied to dew point temperature.

To compute wet-bulb globe temperature from the weather data, we use the Liljegren method, as implemented in the Python package Thermofeel from the European Centre for Medium-Range Weather Forecasts. In order to reflect riders’ experience, WBGT was computed with relative wind, derived by adding ambient U- and V-wind components from ERA5-Land and headwind in the direction of the peloton’s travel generated by their speed. We also compute WBGT with ambient wind only, reflecting conditions for spectators.

Weather data and sources

Who made this

Le Tour de Temps was created and maintained by Iris de Vries and Nino Zumstein. Iris de Vries is a climate scientist at the University of Graz, Austria. Nino Zumstein is a software-developer in Zürich, Switzerland.

Limitations

Historical routes are reconstructions and contain errors in the exact roads taken, as well as the speed and therefore the exact timing of the location of the peloton. Weather variables stem from reanalysis datasets and as such are observation-informed model estimates which can differ from the true historical weather, and are subject to observational as well as model uncertainties and errors. Translating model grid cell average parameters to specific point locations of the peloton requires interpolation and assumptions, which introduces uncertainties as well. The radiation and ambient wind variables used for wet-bulb globe temperature calculations were, as opposed to temperature and dew point temperature, not adjusted to the peloton location using sub-grid-scale features, and are thus a representative area average rather than a highly specific reflection of the conditions at single point locations.