Coding and AI classes in Fürth

import csv

days = []
with open("fuerth_1947_1955.csv", encoding="utf-8") as f:
    for row in csv.DictReader(f):
        if row["rain_mm"] != "-999":
            days.append((row["date"], float(row["rain_mm"]), float(row["pressure_hpa"])))

pairs = []
for today, tomorrow in zip(days, days[1:]):
    pairs.append((today[0], today[1] >= 1.0, today[2], tomorrow[1] >= 1.0))

test = [p for p in pairs if p[0] >= "1953"]
wet = sum(1 for p in test if p[3])
print("Test days:", len(test), " wet tomorrow:", wet)

def score(name, guess):
    right = sum(1 for p in test if guess(p) == p[3])
    missed = sum(1 for p in test if p[3] and not guess(p))
    false_alarm = sum(1 for p in test if guess(p) and not p[3])
    print(name, "accuracy", round(100 * right / len(test), 1), "% missed wet:", missed, "false alarms:", false_alarm)

score("Always dry:", lambda p: False)
score("Same as today:", lambda p: p[1])

Read that twice. A program that answers "dry" every single day is right 73.6% of the time, because in Fürth most days are dry. It beats the program that copies today's weather. And it never warned anyone about a single one of the 216 wet days. That is why "accuracy" alone is a dangerous number, in weather and in every AI system built to spot something rare.

So we count the two kinds of mistake separately: a wet day the model missed, and a false alarm on a dry day. Then we let the program learn something. Falling air pressure often comes before rain, so Version 2 tries every pressure threshold from 960 to 999.5 hPa on the training years and keeps the one that does equally well on wet and dry days.

train = [p for p in pairs if p[0] < "1953"]

def balanced(rows, t):
    wet = [p for p in rows if p[3]]
    dry = [p for p in rows if not p[3]]
    caught = sum(1 for p in wet if p[2] < t) / len(wet)
    calm = sum(1 for p in dry if p[2] >= t) / len(dry)
    return (caught + calm) / 2

candidates = [960 + x / 2 for x in range(80)]
t = max(candidates, key=lambda c: balanced(train, c))
print("Training days:", len(train), " learned threshold:", t, "hPa")
score("Pressure rule:", lambda p: p[2] < t)

The learned rule catches 164 of the 216 wet days, which neither earlier model came close to, yet its accuracy falls to 56.6% because it raises 303 false alarms. Which model is better? There is no answer until someone decides which mistake costs more. For a farmer cutting hay, a missed shower is expensive. For a school trip, a false alarm only means an umbrella nobody needed.

falling = {}
for yesterday, today in zip(days, days[1:]):
    falling[today[0]] = today[2] < yesterday[2]

score("Low and falling:", lambda p: p[2] < t and falling.get(p[0], False))

Data: Deutscher Wetterdienst, Climate Data Center, daily climate data, station 01522 Fürth/Bayern (A), 1 June 1947 to 31 March 1955, licence CC BY 4.0, downloaded 3 November 2026. A wet day is our definition: 1.0 mm of rain or more. All models, thresholds and counts are our own.

Our count from the same DWD file. 1949 had the fewest wet days of the seven full years and 1952 the most, so a model trained on some years and tested on others meets a different climate each time, one more reason to test on separate years.

1. Ages 6 to 8, no screen. For two weeks, guess each evening whether tomorrow will be wet, and keep a tally of right guesses, missed rain and false alarms. Then try "always dry" on the same days.
2. Ages 9 to 12, Python. Run Version 1 and explain to a parent why "always dry" wins on accuracy. Then change the wet-day line from 1.0 mm to 5.0 mm and predict, before running, what happens to the score.
3. Teenagers. Write Version 2 yourself before reading ours, then add a column for the share of wet days caught. Which model would you pick for a hay farmer, and which for a school trip?
4. Adults. Replace the two-way score with a cost: a missed wet day costs five, a false alarm costs one. Find the threshold with the lowest total cost on the training years and check it on the test years.