City of Guelph data projectCan a simple model forecast Guelph’s green bin volume a year ahead?
The City of Guelph publishes the monthly volume of organic waste collected through its green bin program from 2012 to 2019. Our students clean the table, then forecast every month of 2019 from earlier years and check how far off each forecast is. All results here are our own calculations; the volumes are the City’s, in the units it published, which the dataset does not state.
Our forecasts of the 12 months of 2019 on the City of Guelph Total Organic Waste Volume by Month 2012_2019 file (96 monthly rows fetched 22 October 2026)| Forecast method | How it guesses a 2019 month | Average miss per month |
|---|
| Flat average | The 2018 monthly average, for every month | 98.2 |
| Same month last year | The value for the same month in 2018 | 38.5 |
| Same month, six-year mean | The average of that month from 2013 to 2018 | 51.4 |
The code below uses nothing beyond Python’s standard library, and most of it is cleaning: the year appears only on each January row, so the code carries it forward, and the month names arrive in several spellings with stray spaces, so the code trims and shortens them. Only then does it forecast. Run on 22 October 2026, it printed the three averages shown above:import json, urllib.request
URL = ("https://services5.arcgis.com/k3dd78JyG9GFoZHG/arcgis/rest/services/"
"CoG_Org_Waste_Month_Vol_2012_2019/FeatureServer/0/query"
"?where=1%3D1&outFields=*&orderByFields=ObjectId&f=json")
rows = [f["attributes"] for f in json.load(urllib.request.urlopen(URL))["features"]]
# The year is filled in only on each January row, and month names are typed
# inconsistently ("Feb ", "April", "Sept "), so clean both before anything else.
volume, year = {}, None
for r in rows:
year = r["Year"] or year
volume[(year, r["Month"].strip()[:3].title())] = r["Volume"]
MONTHS = ["Jan", "Feb", "Mar", "Apr", "May", "Jun",
"Jul", "Aug", "Sep", "Oct", "Nov", "Dec"]
print(len(rows), "rows,", len(volume), "year-month cells,",
min(y for y, m in volume), "to", max(y for y, m in volume))
def miss(guess):
return round(sum(abs(guess(m) - volume[(2019, m)]) for m in MONTHS) / 12, 1)
last_year = sum(volume[(2018, m)] for m in MONTHS) / 12
print("flat 2018 monthly average:", miss(lambda m: last_year))
print("same month in 2018:", miss(lambda m: volume[(2018, m)]))
print("same month, mean of 2013-2018:",
miss(lambda m: sum(volume[(y, m)] for y in range(2013, 2019)) / 6))
lowest = [min(MONTHS, key=lambda m: volume[(y, m)]) for y in range(2012, 2020)]
print("lowest month each year, 2012-2019:", lowest)
1. Clean like a spreadsheet
The table looks like a copied spreadsheet: a year written once, then blanks, and month labels typed sixteen different ways, such as “Aug” and “Aug ” with a trailing space. Students who skip this step get twelve months that do not line up.
2. The seasonal naive baseline
Guessing that a month will look like the same month last year is called a seasonal naive forecast, and it is the baseline every serious forecaster tries first. Here it beats both alternatives: an average miss of 38.5 against 51.4 and 98.2.
3. Why the long average loses
A six-year mean should be steadier, yet it does worse. In 10 of the 12 months the 2018 value is higher than the 2013 value, so the older years drag the mean down. Recent data wins when a series is trending, an idea students test by trying three-year and two-year means.
February and the calendar
February has the lowest total in every year from 2013 to 2019. That looks like a pattern in behaviour until a student divides by the number of days: per day, February is the lowest month only in 2014, 2015 and 2019. Most of the dip is simply a shorter month. Turning totals into per-day rates is a first taste of feature engineering, choosing what to measure so that a model learns the right thing.
A few more honest details. The program began in 2012, and January 2012 shows 74.33 against 256.67 in February, so the first months describe a launch rather than normal service, which is why the six-year mean starts at 2013. The same-month model’s biggest misses on 2019 were April, 81.5 too low, and August, 73.3 too high. The City notes that the figures do not include organic waste collected in Waterloo Region and processed at the Guelph Waste Resource Innovation Centre, so this is Guelph’s own green bin collection only. We never add the monthly rows into yearly totals; the City publishes a separate yearly file for that. The data is used under the City of Guelph’s Open Data Licence Version 2.0, which grants a licence “to use the Information including for commercial purposes”, and attribution goes to the City of Guelph under its Open Government Licence. No endorsement by the City of Guelph is implied.
