The American Birkebeiner is North America’s biggest cross-country ski race, run every February through the woods of northern Wisconsin. I’ve pulled the results back to the first race in 1973, the older years read from scanned results booklets, and used them to look at who finishes, how fast, and why some years are so much faster than others.

Finishers

The Birkie grew from 33 finishers in 1973 to nearly 5,000 ten years later, and most years since 2010 have had between 5,000 and 6,000. Before 2008, classic skiers raced the same course as the skaters, and most results before 2002 don’t record who skied which. The Kortelopet, the shorter race run the same weekend, has grown alongside it.

Bar chart of Birkie finishers each year from 1973 to 2026, split into skate, classic and open track where known. The field grows from 33 in 1973 to nearly 5,000 in 1983, and to between 5,000 and 6,000 most years since 2010; 2000 and 2017 were cancelled.
Birkie finishers each year. Every finisher in the results, the elite waves included. Grey years don’t record technique, and dashed outlines are finishers on pages the scans lost, counted from the places either side.
Two bar charts of finishers each year: the Kortelopet from 1973, growing to a peak of 2,824 in 2020, and the Prince Haakon from 2010, peaking at 647 in 2020.
Kortelopet and Prince Haakon finishers each year. Each panel has its own scale. No Prince Haakon results were found before 2010.

Finish times

Average times differ by more than an hour between fast and slow years. The course has changed too: the first races ran from Hayward to Telemark, and since 1993 the race has run from Cable to Hayward. In the modern record, 2014 was the slowest year by far and 2026 the fastest.

Line chart of mean Birkie finish time from 1973 to 2026, for skate and classic together and, where known, separately. Modern skate averages mostly sit between about 3:30 and 4:20, with a spike to about 4:48 in 2014 and a low of about 3:20 in 2026.
Average finish time each year. Every finisher, open track left out. Hollow circles are races cut short or run on another course, and years cut much shorter aren’t shown.

The full spread of finish times shows the same thing in more detail. Each curve is a kernel density estimate of finish times. Read its height as “how many skiers came in around this time.” The by year views compare every race since 2009, leaving out years the race was cancelled or run in a modified format. The by wave views split the 2026 race by start wave.

Overlapping density curves of Birkie skate finish times, one per year from 2009 on. Most peak between 3:00 and 4:00; 2014 sits far to the right and 2026 furthest left.
Skate, by year. Snow and weather move the whole field at once. 2014 stands out as a slow year, with its peak pushed past 4:30, and 2026 was the fastest in the record.
Density curves of Birkie classic finish times, one per year.
Classic, by year. The same view for the classic race.
Density curves of skate finish times for each start wave, stepping from just past 2:00 for the first wave out to 5:00 and beyond.
Skate, by start wave. Skiers start in waves seeded by past results, so each wave’s curve sits a little later than the one before it. The curves also widen wave by wave, probably because the later waves mix in skiers without a recent qualifying time.
Density curves of classic finish times for each start wave.
Classic, by start wave. The same view for the classic race.

Weather and race speed

Comparing raw times across years mixes the weather with who showed up. So I model each finish time as the skier’s own ability plus that day’s conditions, their age and their experience, fit on the same skiers from one race to the next. That gives every race day a speed of its own, which I set against the weather along the course: snowfall and snowpack from SNODAS, NOAA’s daily snow model, and temperatures from the Hayward airport. The Kortelopet and the Pre-Birkie, raced on the same trail in other weather, add more race days.

Two things hold up. New snow in the three days before the race slows it, about 0.6% per millimeter of water (roughly a centimeter of fresh snow). Each day above freezing since the last real snowfall speeds it up, about 1.1% a day. Race-day temperature adds nothing reliable. Predicting each race day from all the others, the model misses by 6.7% on average, against 8.3% with no weather at all.

Four scatter plots, one point per race day. New snow before the race against how much slower than average it was, rising about 0.6% per mm of water with 2014 far to the upper right. Days above freezing since the last snowfall, falling about 1.1% a day with 2026 lowest. Predicted against actual speed for unseen race days. Unexplained speed against race-day temperature, with no clear pattern.
Weather and race speed. Each point is one race day of the Birkie, Kortelopet or Pre-Birkie, skate or classic. Top: speed against new snow and against thaw days, each with the other effect taken out. Bottom left: each race day predicted from the other race weekends. Bottom right: what’s left over, against race-day temperature.

Does snow at home help?

For each skier, I counted the skiable days at home (4 or more inches of snow within 10 km, from SNODAS) between November 1 and the race, and compared that with the town’s normal winter. Because the race-day conditions already take out everything the whole field shared, the effect comes only from skiers whose winter at home was snowier or leaner than usual, compared with skiers from elsewhere in the same race.

It helps, but not much. 30 more skiable days than normal is worth about 0.4% (0.2–0.6%), well under a minute for a three-hour skier. That holds for skate and classic, for skiers who never moved, and for every definition of a skiable day. The following winter’s snow, which can’t have helped, shows no gain. The effect is biggest for slower skiers and in towns that normally get little snow.

Left, a line chart of how much faster skiers were than in a normal winter, rising from 0.3% slower with 30+ fewer skiable days at home to 0.7% faster with 30+ more, above a chart of skiable days each winter in Cable, Duluth, Minneapolis, Madison and Chicago. Right, a forest plot of the effect of 30 more skiable days checked across subsets and definitions, almost all between 0.2% and 0.6%, with the placebo near zero.
Snow at home and Birkie times. Top left: speed against how much snowier than normal a skier’s winter at home was. Bottom left: skiable days each winter in a few skier towns. Right: the effect of 30 more skiable days, checked across subsets and definitions, with the shaded band showing the main estimate’s 95% range.

How it’s made

Results since 1999 come from the official results pages and PDFs, scraped with Python (requests and Selenium). Earlier years are read from scanned results booklets and magazines with Tesseract OCR and a parser for each year’s layout, then checked against the places, times and champions list. Skiers are linked from year to year by name, age, gender and hometown, never by finish time. Snow data is NOAA’s SNODAS from NSIDC, and airport weather comes from the Iowa Environmental Mesonet. Plots are made with matplotlib, and the finish-time curves with SciPy’s gaussian_kde. The code and the yearly data are on GitHub.

Plots generated from birkie_data @ f3f7cb4.

Look up a skier

Type a name to see one skier’s races since 2002, charted with the same models as above: their Birkie times with each year’s conditions taken out, every race on one scale, and how much of each field they finished ahead of. Results are matched from year to year by name, age and hometown, so two people with the same name can occasionally get mixed up.