Ville de Gatineau data projectWhich of Gatineau’s traffic counts cannot be right? An anomaly detector in Python
The Ville de Gatineau publishes thousands of 24-hour traffic counts on Données Québec, with French field names such as DEBIT_TOTAL24H and PCENT_V_LOURD. Before training any model on data like this, a careful student hunts for the rows that are impossible or suspicious. Every count of flags in this section is our own result, not the City’s.
Our anomaly checks on the Ville de Gatineau Débits de circulation file (7,384 rows from 2016 to 2023, downloaded from Données Québec on 22 October 2026)| Check | What it looks for | Rows it flags |
|---|
| Validity rule | A heavy-vehicle share (PCENT_V_LOURD) above 100 per cent, which no real count can have | 25 |
| Validity rule | A 24-hour total of zero | 644 |
| Missing label | No sector recorded | 1,916 |
| City-wide fence | Totals above the upper quartile plus one and a half times the interquartile range (14,100) | 624 |
| Paired check | Two rows for the same corner, approach and year where one is over three times the other | 16 of 1,864 pairs |
This is the program, in standard-library Python. It downloads the GeoJSON file from Données Québec, applies three validity rules, builds one fence for the whole city from the quartiles, then compares each row with its partner row at the same corner. On 22 October 2026 it printed every number in the table:import json, urllib.request
from collections import defaultdict
URL = ("https://www.donneesquebec.ca/recherche/dataset/f1e7f552-278d-47d6-89c9-6aa9d3293bd1/"
"resource/388e4ff5-e1f1-4c81-a734-79b907fcf236/download/comptage.json")
raw = urllib.request.urlopen(URL).read().decode("utf-8-sig")
rows = [f["properties"] for f in json.loads(raw)["features"]]
print(len(rows), "counts from", min(r["ANNEE"] for r in rows), "to", max(r["ANNEE"] for r in rows))
# 1. rules a count must obey
print("heavy-vehicle share above 100%:", sum((r["PCENT_V_LOURD"] or 0) > 100 for r in rows))
print("24-hour total of zero:", sum(r["DEBIT_TOTAL24H"] == 0 for r in rows))
print("no sector recorded:", sum(not r["SECTEUR"] for r in rows))
# 2. one fence for the whole city: quartiles of the positive totals
flows = sorted(r["DEBIT_TOTAL24H"] for r in rows if r["DEBIT_TOTAL24H"] > 0)
q1, q3 = flows[len(flows) // 4], flows[3 * len(flows) // 4]
fence = q3 + 1.5 * (q3 - q1)
print("city-wide fence", fence, "flags", sum(f > fence for f in flows), "rows")
# 3. compare each count with its partner: same corner, same approach, same year
pairs = defaultdict(list)
for r in rows:
pairs[(r["NOMRUE"], r["APPROCHE"], r["ANNEE"])].append(r["DEBIT_TOTAL24H"])
twos = [v for v in pairs.values() if len(v) == 2 and min(v) > 0]
odd = [v for v in twos if max(v) > 3 * min(v)]
print(len(twos), "two-way pairs;", sum(v[0] == v[1] for v in twos), "identical;",
len(odd), "where one side is over three times the other")
print("most lopsided pair:", max(odd, key=lambda v: max(v) / min(v)))
1. Rules before statistics
Some values are wrong by definition. A share of heavy vehicles cannot pass 100 per cent, yet 25 rows do, the highest at 324.6, all from 2017 to 2019. No statistics are needed to flag those; a single comparison in code does it.
2. A global fence misfires
A fence built from the whole city flags 624 rows, and most of them are simply busy roads. Gatineau mixes quiet residential corners with arterials, so “far from the typical count” mostly means “a main road”, not “an error”.
3. Compare like with like
Comparing each count with its partner at the same corner is far sharper. Of 1,864 pairs, 1,762 match exactly and only 16 differ by more than three times. Those 16 are worth a human look; some will be genuine, such as one-way streets.
Two rows a student should question
At Aylmer / Vanier in 2018, one approach records 127,400 vehicles in 24 hours beside a partner row of 12,700, and 127,400 is the largest value in the whole file, three times the next largest (41,400). At Fournier & Lac Leamy in 2019, a pair reads 243 against 24,300, exactly a factor of one hundred. Both look like slips in entering digits, but the file alone cannot prove it. A flag is a question for the data owner, not a verdict, and a good analyst writes it up that way.
Two patterns are worth noting honestly. In 1,762 of the 1,864 pairs the two directions carry exactly the same total, which suggests the file sometimes repeats one figure on both rows rather than counting each direction separately; we pose that as a question rather than state it. And the sector field is empty on 1,916 rows, including every 2023 row, so any analysis by sector (Aylmer, Buckingham, Gatineau, Hull, Masson-Angers) has to drop the most recent year. The file is published by the Ville de Gatineau on Données Québec under the licence “Attribution (CC-BY 4.0)”; Données Québec explains that this licence lets others “distribuer, remixer, arranger et adapter votre œuvre, même à des fins commerciales”, and the Creative Commons deed says use is allowed “for any purpose, even commercially.” Source: Ville de Gatineau, Débits de circulation, Données Québec. The City has not endorsed this page or our teaching.
