The problem
An internet or mobile provider sells larger plans to households that are about to outgrow theirs. The usual trigger is a household that has already gone over its allowance, so the offer arrives after the bill shock. Providers that choose offers from recent usage, and send them before the limit, report better take-up than blanket campaigns, with the gains coming from offering the right size of plan rather than the biggest.
I measured how well last month’s metered traffic identifies the households that will exceed their allowance next month, so the offer can go out in time, against offering to everyone who was over last month.
- Team
- Head of base marketing, pricing, customer value management, CRM
- Scale
- Hundreds of thousands to millions of fixed broadband households
- Systems
- Usage records from the network, billing and plan catalogue, CRM and campaign tools
- Measure
- Share of next month’s heavy users offered in time; hit rate of the offer list
What I built
A monthly offer list from each household’s metered traffic: the month’s pace to a 30-day total, the last week against the first, the busiest days, the evening share and the Wi-Fi and upload shares. I compared a gradient-boosted score on those signals with the plain ranking by last month’s pace.
Results
I tested on the FCC’s Measuring Broadband America raw data usage records for January and February 2023: hourly traffic counted on each volunteer household’s test box, 4,025 households with at least 20 days of records in both months. The line was a pace of 1 TB a month; 4.8% of households were above it in February. The score was tested out of fold across five repeated splits.
Of the households above the line in January, 70% were above it again in February, and they made up 70% of February’s heavy users. The other 30% were under the line the month before, most of them close to it.
| Ranking | AUC, all households | AUC, households under the line | Share reached, top 5% of those under the line |
|---|---|---|---|
| Last month’s pace | 0.975 | 0.941 | 68% |
| Usage score, twelve signals | 0.972 | 0.932 | 66% |
The usage score did not beat the plain pace ranking on any cut, with an AUC of 0.972 against 0.975. Heavy use is steady from month to month, so a pace line set a little under the allowance and widened to the top tenth is the offer list I would run. The score earns its place once offer responses are recorded and the question becomes who accepts, not who is heavy.
Data and method
- Data
- FCC Measuring Broadband America fixed raw data, January and February 2023, data usage records (curr_datausage): wired and Wi-Fi bytes sent and received by household devices, hourly. The FCC publishes the programme data for public use without restrictions.
- Target
- February traffic on a pace of 1 TB or more for a 30-day month
- Population
- Households with at least 20 days of records in both months; hourly counter readings above 450 GB removed as counter faults
- Methods
- Gradient-boosted classifier on twelve January usage signals; five-fold out-of-fold predictions repeated over five seeds
- Comparisons
- Households above the 1 TB pace in January; the ranking by January pace
Limitations
- The panel is volunteer US households, and usage is counted on the test box, which can miss devices connected elsewhere.
- Only two months of broad usage records are in the 2023 release, so the test is one month ahead and seasonal swings are not covered.
- The target is heavy use, not acceptance of an offer. A live campaign records responses and refits on them.
- Plan allowances differ by provider; the line is set to the provider’s own allowance in production.
How I would take it live
- Pull three months of household usage and plan allowances from billing
- Set the line under the allowance and size the list to the campaign
- Run one cycle with a held-back group and record responses
- Refit on accepted offers once a few cycles are recorded