Update, October 2026: the main finding below was wrong. The program asked NRCS for three stations that don’t exist (774, 778 and 780 as Snoqualmie, Stevens and Stampede passes) and filled in zero snow when nothing came back, so the bare passes and the 15% figure were artifacts: NRCS’s numbers for April 2 show Olallie Meadows, the Snoqualmie Pass station, at 24.8 inches of water, 46% of median. The percent-of-normal figures here were against made-up April 1 normals, and the atmospheric river index and storm types were formulas, not measurements. Version 3 is rebuilt on the Northwest’s weather balloons with real normals, in FORTRAN. The write-up is here.
Yesterday I built a COBOL program that fetched live USGS streamflow data from eight Sierra Nevada rivers. It ran on Artemis II launch day and terminated normally with return code zero. I wrote about it here.
This morning over coffee it hit me: COBOL came out in 1959, FORTRAN in 1957, and FORTRAN is still what runs the atmosphere.
The National Weather Service runs FORTRAN, and so does NCAR. So does ECMWF, the European Centre for Medium-Range Weather Forecasts, which is widely considered the best weather modeling operation on Earth. The WRF model, which so much regional forecasting in the United States is built on, is FORTRAN. The first successful numerical weather prediction was computed in 1950 on ENIAC, and when that code moved to faster machines it was rewritten in FORTRAN. People have been computing the atmosphere in FORTRAN for 75 years.
So I built CASCADIA-WX.
What It Does
Every morning at 8AM Pacific, GitHub Actions runs the pipeline. A Python fetcher pulls live data from 11 NRCS SNOTEL stations in the Cascades, the Olympics and on the flanks of Rainier. Those sensors have been running since before the internet, measuring snowpack in the mountains that supply water to most of western Washington. Readings from four valley airports come in too, and all of it goes into a CSV.
Then FORTRAN takes over.
The program works out environmental lapse rates from the temperature difference between the valley and mountain stations. It estimates the snow level, the elevation where the temperature crosses 2Β°C and rain turns to snow. Then it splits precipitation into rain and snow, adds up degree days, calculates an atmospheric river index from the snow level anomaly and SWE rate, sorts the storm into a type (Gulf of Alaska, Pineapple Express or Cutoff Low), and rolls it all up by massif: Rainier, the Olympics and the Cascades. Out comes a formatted report in four sections, committed back to the repo. The dashboard reads it live.
The Numbers Are Alarming
The Cascades are running at 15% of normal snowpack. Snoqualmie Pass, 0.0 inches. Stevens Pass, 0.0 inches. Stampede Pass, 0.0 inches. The passes are bare in early April, when they should be near their peak, just before the melt starts.
Paradise, at 5,150 feet on the south side of Rainier, still has 36.7 inches, 73% of normal. That helps less than it sounds. The reservoirs downstream draw on the whole watershed, and whatever snow the passes don't hold now, the valley won't have in August.
This is the kind of work FORTRAN was built for: big arrays of numbers, physical equations applied over and over, and formatted output. John Backus designed the language at IBM in 1957 because scientists were spending more time writing machine code than doing science. The name is short for FORmula TRANslation, and it let them write equations that looked like equations. Seventy years later, that's still the right idea for atmospheric modeling.
The Eleven Stations
Eleven sensors across three massifs, each one measuring snow that will become river flow, and then drinking water, irrigation and hydropower. The most interesting station in the network is also the highest, Corral Pass at 5,810 feet. It's the last to lose its snow and the first to show when the season has turned. Snoqualmie Pass, at 3,000 feet, is the canary on the low end. When it goes bare in March instead of May, something is wrong.
Right now something is wrong, and FORTRAN measures it every morning.
Why FORTRAN After COBOL
Yesterday's COBOL project and today's FORTRAN one are apparently turning into a retro-computing streak. They work the same way. A Python fetcher, standard library only, pulls free government data. The old compiled language does the domain math, GitHub Actions runs it daily, and a static dashboard reads the output. No dependencies, no cost, and correct output every morning.
Where they differ is in what each language is good at. COBOL is wordy and record-oriented and fussy about data layout, which suits the tabular output of a streamflow report. FORTRAN is mathematical and array-oriented, built for floating-point work, which suits the physics of the atmosphere. You can't swap one for the other. They were designed for different problems, and government agencies still run both in production for exactly those problems.
Writing FORTRAN in 2026 felt more natural than I expected. The fixed-format source (columns matter, line length is enforced) is jarring at first if you're coming from modern languages, but the math reads beautifully. An array computation that would take several lines in Python is one line in FORTRAN. The lapse rates, the degree days and the atmospheric river index read almost like the equations in a meteorology textbook. That was the original design goal, and it still works.
The Hatch Has to Be Maintained
The README for this project has a LOST reference: enter the numbers every 108 minutes or the snowpack anomaly gets worse. That's a joke, but it isn't entirely wrong. The SNOTEL network has measured snowpack continuously since the 1970s, with daily observations at every station, every winter, and that unbroken record is what makes percent-of-normal mean anything. Without it you'd have numbers and nothing to compare them to.
CASCADIA-WX does its small part. Every morning at 8AM it pulls the numbers, runs the physics and commits the result, whether anyone is watching or not.
The live dashboard is here. The source is on GitHub. The snowpack is at 32% of normal and falling.
CASCADIA-WX.f90: NORMAL TERMINATION. RETURN CODE: 0.
Go deeper
Hello Fortran! — Code for Thought, 5 September 2023, about 40 min. Thomas Clune of NASA and three fellow Fortran folks on the language’s origins, wrangling legacy code and why science still runs on it.
Warm winter liquidates nearly half of Washington state snowpack — KUOW, 6 February 2026, about 4 min. John Ryan on the winter this post was measuring, with the statewide snowpack near half of normal.
Charney, J.G., Fjörtoft, R. & von Neumann, J. (1950). Numerical integration of the barotropic vorticity equation. Tellus 2(4), 237–254. Open access.
Backus, J. (1978). The history of FORTRAN I, II, and III. ACM SIGPLAN Notices 13(8), 165–180.
Lynch, P. (2008). The origins of computer weather prediction and climate modeling. Journal of Computational Physics 227(7), 3431–3444.
Serreze, M.C., Clark, M.P., Armstrong, R.L., McGinnis, D.A. & Pulwarty, R.S. (1999). Characteristics of the western United States snowpack from snowpack telemetry (SNOTEL) data. Water Resources Research 35(7), 2145–2160.
Mote, P.W., Hamlet, A.F., Clark, M.P. & Lettenmaier, D.P. (2005). Declining mountain snowpack in western North America. Bulletin of the American Meteorological Society 86(1), 39–50.
The Emergence of Numerical Weather Prediction: Richardson’s Dream — Peter Lynch (Cambridge University Press); the whole story from Richardson’s hand-computed forecast to the ENIAC runs, retold by the man who re-ran them.
NRCS Washington Snow Survey — the people behind the state’s SNOTEL stations and water-supply outlooks. · NRCS Snow and Water Interactive Map — every station’s snow water and percent of median, on one map.
NWS Seattle/Tacoma — the local forecast office. · WRF model at NCAR — the Fortran forecast model the post tips its hat to.
CASCADIA-WX live page — now version 3, reading the balloons. · Source on GitHub — the FORTRAN and the fetcher.
