Fan segmentation and broadcast conversion
This market watches more hockey than any other in the United States and fills its building less than almost any club in the league. The model clusters a 40,000-profile sample of the club’s resolved fan file on six behavioral features, reads the segment names off the resulting centroids, then scores the broadcast-first fans on their probability of buying a ticket. The output is a ranked target list and a sized campaign.
Segmentation
Number of segments (k)
6
k-means is re-fit from scratch at each k. Names are derived from the centroids, so segments appear, merge, and split as k moves.
Focus segment
Campaign
Reach of addressable pool
45%
Share of the addressable pool the campaign touches, worked highest score first.
Response multiplier
1.00×
Scales every modeled conversion probability. 1.00× is the fitted baseline; a strong offer or a winning season pushes it up.
Incremental per-cap margin
$27
Concessions, merchandise, and parking margin on each new admission — the same figure the pricing model uses.
Cost per reached fan
$6.00
Loaded media, offer, and fulfilment cost for each fan the campaign actually touches.
Conversion model
Log-odds contributions at the observed range of each driver. Intercept -2.85 — a base conversion of 5.5%.
Telecasts watched
+2.40 × (watched/82 − ½)
Distance from arena
−0.012 / mile, capped at 150
Digital engagement
+1.10 × (index/100 − ½)
Prior single-game purchase
+0.85
Household income index
+0.60 × (index − 1)
Tenure with the club
+0.35 × (years/20, capped)
Incremental annual revenue · broadcast to building
$1,047,223
from a campaign reaching 45% of the 47,326 broadcast-first fans the model can identify, at $78 net ticket yield plus $27 per-cap
47,326
Addressable broadcast-only fans
3,596
Expected converts
+243
Incremental attendance per game
4.3×
Margin return on $127,780
Where the fan base sits
Each point is a fan profile, plotted by telecasts watched against games attended, colored by discovered segment. The mass in the lower right is the target: fans who watch most of the schedule and never buy a ticket. The frame is downsampled to roughly 600 points, drawn proportionally from each segment; points are jittered slightly and the vertical axis is square-root scaled so the low-attendance band does not collapse onto the baseline.
The segments k-means found at k = 6
Segment profiles in original units, weighted back to the full resolved file. Distance is a median because a few out-of-market fans drag the mean past anything a member looks like. Names are assigned by rule from the profile, so a segment that stops looking like a broadcast loyalist stops being called one.
| Segment | Fans | Games | Telecasts | Miles (med) | Annual spend | Identified value | Addressable | Avg P(convert) |
|---|---|---|---|---|---|---|---|---|
| Core attenders | 2,731 | 31.2 | 61 | 17 | $6,340 | $17.3M | — | — |
| Occasional attenders | 22,784 | 4.2 | 48 | 23 | $525 | $12.0M | — | — |
| Low-signal followers | 15,586 | 0.9 | 13 | 28 | $111 | $1.7M | 946 | 5.1% |
| Digital-native casuals | 18,857 | 0.8 | 22 | 17 | $106 | $2.0M | 8,722 | 5.6% |
| Regional & diaspora | 17,828 | 0.8 | 48 | 165 | $155 | $2.8M | 13,430 | 3.0% |
| Broadcast loyalists | 25,214 | 0.4 | 67 | 24 | $62 | $1.6M | 24,228 | 15.5% |
Identified value is annual spend the resolved file can attribute to a named fan. Group sales, resale buyers, and walk-ups who never identify themselves sit outside it, which is why this total is smaller than the gate the pricing model forecasts.
Conversion funnel
The pool is worked in score order, so the first fans reached are the ones the model likes most.
Addressable
attend ≤ 1 game, watch ≥ 20 telecasts
47,326
Reached
45% of the pool, highest score first
21,297
Converted
16.9% response among reached
3,596
Incremental admissions
2.8 games per convert
9,974
Spread evenly across the 41 home dates that is +243 per game, 9% of the 2,649 seats that sit empty on an average night, and 1.28 points of arena fill.
Where to spend the reach
Segments ranked by incremental revenue per fan reached, at the current settings. This is the ordering a campaign should be built in, and the break-even line says where to stop.
Broadcast loyalists
19,896 reached · 17.2% response · $1.0M
$50.35
Digital-native casuals
998 reached · 12.8% response · $33K
$33.29
Regional & diaspora
344 reached · 12.0% response · $11K
$30.59
Low-signal followers
60 reached · 11.4% response · $2K
$28.47
Work broadcast loyalists first: $50.35 of incremental revenue per fan reached against $6.00 of cost. At $6.00 per reached fan and $27 of per-cap margin, the marginal fan stops covering her own cost at 62% reach, which is where the campaign should be sized rather than at whatever budget is available.
Sensitivity · incremental revenue
Reach against response multiplier. Response is the assumption that deserves the least trust, so it gets its own axis.
| Reach | 0.70× response | 1.00× response | 1.40× response | Fill pts @ 1.00× |
|---|---|---|---|---|
| 15% | $396K | $565K | $792K | 0.69 |
| 30% | $596K | $852K | $1,192K | 1.04 |
| 45% | $733K | $1,047K | $1,466K | 1.28 |
| 60% | $820K | $1,172K | $1,640K | 1.43 |
| 75% | $867K | $1,239K | $1,735K | 1.51 |
Reading the model
- 01
51,212 fans in the resolved file watch 20 or more telecasts a season and attend at most one game — 49.7% of the identified fan base. At k = 6 the model isolates 4 broadcast-first segments holding 47,326 of them.
- 02
Reaching 45% of that pool at a 1.00× response converts 3,596 fans into buyers and 9,974 additional admissions — 243 more people per game, or 1.28 points of arena fill, moving the building from 86.1% to 87.4%. That is 9% of the 2,649 seats currently empty on an average night.
- 03
The money is concentrated in broadcast loyalists: $50.35 of incremental revenue per reached fan, 96% of the total on 93% of the reach. The program costs $127,780 and returns $549,343 in contribution margin, 4.3×.
- 04
Response is steeply ranked, not uniform: the top fifth of the addressable pool converts at 22.5% against 1.8% for the bottom fifth. At $6.00 per reached fan the marginal fan stops paying for herself at 62% reach, so reach beyond that buys non-responders.
Method
The frame is 40,000 fan profiles carrying six features: games attended, annual spend, telecasts watched out of 82, a digital engagement index, distance from the arena, and tenure. Spend and distance are log-transformed, then every column is standardized, and k-means runs with k-means++ seeding and fifteen Lloyd iterations from a fixed seed. Segment names are applied afterward by a rule over the segment profile in original units: a group that attends fewer than 2 games and watches 42 or more telecasts is called a broadcast loyalist because of what it looks like, not because a label was reserved for it. When a k splits one base label into two, the second is qualified on whichever profile dimension separates it most from the first. Moving k re-fits the model rather than re-cutting a fixed answer, which is why segments merge and split as the slider moves.
Conversion is a logistic score of the probability a fan buys at least one ticket next season if the campaign reaches her, with coefficients shown in the sidebar. Every fan in a broadcast-first segment who watches at least 20 telecasts and attends no more than 1 game enters the addressable pool; the pool is then worked in score order, so reach and response compose sensibly and the marginal fan is always the next-best one rather than an average one. Revenue is admissions × ($78 net ticket yield + per-cap margin); a convert is assumed to attend 1.8 games plus a propensity-scaled increment. Program return is computed on contribution margin, not gross revenue — 36% on the ticket line, consistent with the renewal and lifetime-value models — because crediting a full ticket price to a piece of media flatters the campaign.
What production would need: the response multiplier is the weakest link and belongs to a holdout test, not a slider. Real cluster counts get chosen by silhouette and stability across resamples rather than by eye, and clusters get re-fit each season because behavior drifts. The incremental attendance here is spread evenly across the 41 home dates, which the pricing model shows is wrong — the room is in weak midweek inventory, so a real campaign would target dates as well as people. And the addressable pool is deliberately limited to fans the club can already identify and contact; the telecast audience is several times larger than the resolved file, so the honest reading of these numbers is a floor set by how much of the market has been captured into a record.
The fan file is simulated and seeded, so the figures are stable across visits. Bounded features are drawn by reflecting at the bound rather than clipping to it, so the simulator does not leave a spike of fans sitting on a boundary for the clustering to find. It is drawn as a stratified sample of the suite’s ~103,000-fan book — every season-ticket account, mini-plan holder, and playoff-cohort buyer, plus weighted samples of single-game and digital-only fans — and each profile carries a weight so counts and revenue project back to the full file. Market inputs (19,070 capacity, 16,421 average attendance, 41 home dates) are public. Not affiliated with any professional team.