Waterloo data projectCan a model guess tomorrow’s traffic on the Laurel Trail? A first machine learning project
The City of Waterloo publishes hourly counts from counters on its trails. Our students download one counter’s history, train three very simple models on 2018 and test them on 2019. Every number in this section is our own calculation, not the City’s.
Our results on City of Waterloo Trail Counter Data, counter LOC_ID 4, Laurel Trail at University Ave. W. (44,114 hourly rows fetched 22 October 2026; trained on 356 complete days in 2018, tested on 360 complete days in 2019)| Model (trained on 2018) | How it guesses a 2019 day | Average miss on 2019 (people per day) |
|---|
| Baseline | The same number every day: the 2018 daily average, about 171 | 90 |
| Weekday model | The 2018 average for that day of the week | 90 |
| Month model | The 2018 average for that month, from about 59 a day in January to 324 in July | 65 |
Here is the whole project in standard-library Python. It pages through the City’s public query service 2,000 rows at a time, adds the hourly counts up by date, keeps only dates with all 24 hours, and scores each model by its mean absolute error. It printed the three numbers in the table when we ran it on 22 October 2026:import json, urllib.request, datetime
URL = ("https://services.arcgis.com/ZpeBVw5o1kjit7LT/arcgis/rest/services/"
"City_of_Waterloo_Trail_Counter_Data/FeatureServer/0/query"
"?where=LOC_ID%3D4&outFields=DATE,TOTAL_COUNT&orderByFields=OBJECTID"
"&resultRecordCount=2000&f=json&resultOffset=")
daily, hours, offset = {}, {}, 0
while True:
page = json.load(urllib.request.urlopen(URL + str(offset)))
rows = page["features"]
if not rows:
break
for r in rows:
a = r["attributes"]
day = datetime.datetime.fromtimestamp(a["DATE"] / 1000, datetime.timezone.utc).date()
daily[day] = daily.get(day, 0) + (a["TOTAL_COUNT"] or 0)
hours[day] = hours.get(day, 0) + 1
offset += len(rows)
full = [d for d in daily if hours[d] == 24]
train = [(d, daily[d]) for d in full if d.year == 2018]
test = [(d, daily[d]) for d in full if d.year == 2019]
def fit(key):
groups = {}
for d, n in train:
groups.setdefault(key(d), []).append(n)
return {k: sum(v) / len(v) for k, v in groups.items()}
def error(key):
model = fit(key)
return round(sum(abs(model[key(d)] - n) for d, n in test) / len(test))
print(offset, "hourly rows,", len(train), "train days,", len(test), "test days")
print("same guess every day:", error(lambda d: 0))
print("guess by weekday:", error(lambda d: d.weekday()))
print("guess by month:", error(lambda d: d.month))
1. Beat the baseline
A model is only interesting if it beats a dumb guess. Guessing the yearly average every day misses by 90 people a day on 2019. Knowing the weekday does not help at all here, which surprises most students; knowing the month cuts the miss to 65.
2. Split by time
Training on 2018 and testing on 2019 is honest in a way a random split is not: the model never sees the future it is graded on. Students then try training on 2017 to see whether a longer history helps.
3. Audit the labels
The same counter can carry two spellings. In the City’s file, LOC_ID 2 appears as both “Iron Horse Tail at John St. W.” and “Iron Horse Trail at John St. W.”, and LOC_ID 10 as both “Spur Line Trail at Roger Street” and “Spurline Trail South of Roger St.”. Group by ID, never by name.
What the model cannot know
The 2019 days averaged about 133 people against 171 in 2018, so every 2018 model guesses high. Is that a real change in walking and cycling, a counter moved or partly blocked, or something else? The data alone cannot say. The City’s own note warns that construction meant traffic “has been diverted around many EcoCounter installations resulting in a perceived drop in traffic”, without naming this counter. And 15 April 2018, 14 September 2019 and 15 September 2019 record zero people across all 24 hours, which looks more like a counter gap than an empty trail. A student who flags those days before trusting the error figure is doing real data science.
Two more honest details. The item description says “This data is updated daily”, yet the newest rows we could fetch are dated 30 November 2021, so this is a historical exercise, not a live feed. The whole table held 534,489 rows across all counters on the day we queried it; we used one counter so a laptop finishes in seconds. The data is used under the Waterloo Open Data User Licence, which lets anyone “use the Information commercially”. Contains information provided by the City of Waterloo under licence (licence, dataset). The City has not endorsed this page or our teaching.
