Maintenance demand, forecast a week ahead, with an alert before the surge
A dashboard and alert service for teams that dispatch repairs. Built on 3,531,417 housing-maintenance requests to New York City's 311 line since 2022 and daily temperature. The forecast cuts weekly planning error from 30% to 21%, and for heating faults from 48% to 25%.
Executive summary
A week-ahead maintenance forecast cuts planning error from 30% to 21%, and for heating faults from 48% to 25%, on 3.5 million real repair requests.
Situation
Housing repairs teams book contractors a week out from last week’s numbers, and learn about a cold snap from the phones.
Approach
A seven-day forecast per category from request counts and the temperature forecast, with an upper range, tested over a full year on NYC housing-maintenance requests.
Result
Heating faults are where the forecast earns its keep. For steady categories a four-week average is as good, so each category runs whichever method won last quarter.
The business problem
A public or community housing provider’s repairs team books contractors and call-centre staff a week ahead from last week’s numbers. Cold snaps and storms change that overnight: housing providers have reported calls more than doubling in a hard winter week, with heating, plumbing and leak jobs leading. The team usually learns about the surge from the phones.
I built a seven-day forecast per repair category from request history and the temperature forecast, with an upper range and an alert for the categories that surge, and measured it against last week’s count.
What was built
A nightly job pulls requests by category from the maintenance system and the temperature forecast, refreshes a 7-day forecast per category with an upper range, and publishes a one-page dashboard. When a forecast or the day's running count crosses the range, it emails or posts to Teams for the operations lead with the numbers and the categories involved.
Results
Heating faults follow the thermometer
Every category, side by side
| Category | Per day | Same day last week | 4-week average | Model | Days inside range | Alerts |
|---|---|---|---|---|---|---|
| Heat and hot water | 959 | 48.4% | 47.9% | 25.1% | 97.3% | 10 |
| Unsanitary condition | 387 | 12.4% | 10.9% | 12.8% | 86.6% | 49 |
| Plumbing | 236 | 22.4% | 22.5% | 20.1% | 90.4% | 35 |
| Paint and plaster | 192 | 17.8% | 16.1% | 16.0% | 84.7% | 56 |
| Doors and windows | 151 | 17.3% | 16.1% | 17.7% | 81.4% | 68 |
| Water leak | 137 | 28.7% | 26.0% | 28.7% | 90.7% | 34 |
| General | 126 | 18.0% | 15.0% | 26.8% | 58.1% | 153 |
| Electrical | 101 | 22.3% | 19.5% | 21.2% | 84.7% | 56 |
| Flooring and stairs | 87 | 18.0% | 16.1% | 16.7% | 93.2% | 25 |
| Appliance | 73 | 19.2% | 16.0% | 18.8% | 89.6% | 38 |
Error is weighted absolute percentage error over the test year: total absolute miss divided by total requests. "Days inside range" should be about 97.5% if the upper range is well set.
Alerts: useful in three categories, noise in the rest
Alerting on every category would have raised 524 alerts in a year, too many to act on. Limiting alerts to heat and hot water, plumbing and water leaks gives 79, about 1.5 a week. The largest fell in winter cold snaps, when extra contractors need booking:
| Date | Category | Requests | Forecast | Alert threshold |
|---|---|---|---|---|
| Sun 8 Feb 2026 | Heat and hot water | 6,118 | 3,149 | 4,837 |
| Sat 24 Jan 2026 | Heat and hot water | 5,614 | 2,895 | 4,513 |
| Sun 25 Jan 2026 | Heat and hot water | 5,601 | 2,933 | 4,561 |
| Sun 1 Feb 2026 | Heat and hot water | 4,688 | 2,431 | 3,914 |
| Thu 2 Jul 2026 | Plumbing | 990 | 213 | 299 |
| Fri 30 Jan 2026 | Heat and hot water | 5,200 | 2,923 | 4,550 |
Data and method
- Requests
- NYC 311 Service Requests, agency HPD, 10 housing-maintenance categories, 2022-01-01 to 2026-11-15
- Scale
- 3,531,417 requests, aggregated to daily counts
- Weather
- Open-Meteo archive, daily mean and minimum temperature for New York
- Test
- Last 365 days (2025-11-16 to 2026-11-15); calibration of the range on the 180 days before
One gradient-boosted regressor per category predicts the count seven days ahead from lagged counts (7 to 14, 21, 28, 35 and 364 days), rolling means known at forecast time, weekday, day of year, holidays and temperature. The upper range is the 97.5th percentile of scaled errors on the calibration window. Baselines: the same weekday last week, and the average of the same weekday over the last four weeks.
Assumptions and limitations
- Observed temperature stands in for the forecast. A real 7-day forecast adds error, largest on the coldest days
- This is New York, with mandated heating. An Australian rent roll will see heating faults only in Melbourne, Canberra and Hobart winters; leaks and storm damage matter more
- Upper ranges drift when demand trends upward; they need recalibrating monthly, which the test year shows plainly
- The model forecasts requests, not jobs: duplicates and multi-request jobs are not separated in this data
What a housing repairs team would do next
- Week 1: export two years of maintenance jobs by category and date from the asset or repairs system
- Week 2: fit per category, pick model or simple rule per category on the last quarter
- Weeks 3 to 4: run the dashboard and alerts alongside current planning
- After a month: keep alerts only where they led to a booking decision