Predicting session attendance from voting results

Data from a 2025 unconference: 259 attendees, about 9 sessions running in parallel at any given time, and 13 sessions whose hosts could tell us roughly how many people showed up.

Voting was a three-way choice per session — Interested, Maybe, or Skip. Around 40% of attendees voted (99–107 votes cast per session).

Every attendance figure is a host’s recollection, not a count. Treat them as ±25% at best. Sessions are numbered by interest share; titles are withheld.

The data

# Interested Maybe Skip Total votes Interest share RSVPs Attended
Session 1 57 26 24 107 53% 52 ~40–60
Session 2 43 26 35 104 41% 18 ~12
Session 3 40 19 46 105 38% 35 ~35
Session 4 36 30 37 103 35% 45 ~30–40
Session 5 28 20 54 102 27% 13 ~15
Session 6 27 31 43 101 27% 12 ~15–20
Session 7 27 29 47 103 26% 13 ~10
Session 8 24 29 50 103 23% 7 ~5–8
Session 9 22 15 65 102 22% 12 ~18
Session 10 19 34 51 104 18% 16 ~15
Session 11 16 35 51 102 16% 8 ~25
Session 12 12 13 76 101 12% 15 ~20
Session 13 9 17 73 99 9% 8 ~6

Interest share = Interested ÷ total votes on that session. Average across the 13: 27%.

What the data says

  • About 69% of Interested votes turn into bodies in the room. That is the formula in one sentence; the rest is bookkeeping for events of a different size.
  • “Maybe” votes carry no signal. Adding them to the interest share at any weight made predictions worse. Only Interested counts.
  • The link is real but noisy. Sessions with near-identical interest shares drew crowds differing by a factor of two or more, in both directions.
  • RSVPs predict far better than votes (R² 0.63 against 0.20 for the formula below, both on log attendance) — but they arrive after you have already had to pick a room, and once you know them the vote adds nothing.

The formula

It predicts a range, not a number. Given the noise in the underlying data, a single number would be false precision.

            s          N
  M  =  k · ──   ×    ───          the midpoint
            s̄          P

  range  =  M ÷ g  …  M × g        g = exp(z · σ)

                  ┌─────────────────────┐
        σ  =  sqrt│  0.31  +   1 − s    │      z = 0.67 for a 50% range
                  │            ───────  │      z = 1.28 for an 80% range
                  └             s · V   ┘
symbol meaning 2025 value
N total attendees at the event 259
P sessions running in parallel 9
s Interested ÷ total votes, for your session varies
s̄ average s across all sessions at the event 0.27
V number of people who voted on your session ~103
k share of attendees in any session at a given moment 0.66
0.31 irreducible variance — what the vote cannot know —

Does the number of voters matter?

Yes, but only for the width of the range, never for the midpoint.

s is a share, not a count, so if 200 people vote instead of 100 and the same proportion are Interested, the prediction is unchanged. That is the desired behaviour: turnout should not change the answer, only the confidence in it.

Where turnout does enter is the (1 − s) / (s · V) term — the sampling error of estimating a share from V voters. Few voters, wide range:

Voters on the session Midpoint 80% range width
10 19 7–51 ×7.1
25 19 8–44 ×5.3
50 19 9–42 ×4.7
100 19 9–40 ×4.5
259 19 9–40 ×4.3

(holding interest share at 27%)

Two honest caveats on this:

  1. This part is derived, not measured. In 2025 every session got 99–107 votes, so there is no variation in turnout to fit against. The sampling term comes from binomial theory, not from evidence.
  2. It assumes voters are representative of attendees. With 40% turnout that is a real assumption. If the people who bother to vote are systematically keener session-goers, the formula over-predicts across the board — and k is silently absorbing some of that.

At 2025 turnout, sampling error is only ~10% of the total error. Polling more people sharpens the estimate a little; the floor is the 0.31 intrinsic term, which no amount of voting can reduce.

Look-up table

For a 2025-sized event (N=259, P=9, s̄=0.27, ~103 voters per session).

Interest share Likely (50%) Plan for (80%)
5% 2–6 1–9
10% 5–11 3–16
15% 7–16 5–23
20% 10–21 7–30
25% 12–26 8–38
30% 14–31 10–45
35% 17–37 12–52
40% 19–42 14–59
45% 22–47 15–66
50% 24–52 17–74
60% 29–62 21–88

Use the 50% range to decide whether the session is worth running, and the top of the 80% range to pick a room. Under-booking a room hurts more than over-booking.

How well does it work?

# Interest share Likely (50%) Plan for (80%) Actual In 80%?
Session 1 53% 26–55 18–78 ~40–60 yes
Session 2 41% 20–43 14–61 ~12 no
Session 3 38% 18–40 13–56 ~35 yes
Session 4 35% 17–37 12–52 ~30–40 yes
Session 5 27% 13–29 9–41 ~15 yes
Session 6 27% 13–28 9–40 ~15–20 yes
Session 7 26% 13–28 9–39 ~10 yes
Session 8 23% 11–25 8–35 ~5–8 yes
Session 9 22% 10–23 7–33 ~18 yes
Session 10 18% 9–19 6–28 ~15 yes
Session 11 16% 7–17 5–24 ~25 no
Session 12 12% 6–13 4–19 ~20 no
Session 13 9% 4–10 3–15 ~6 yes

Calibration — how often the range caught the session. Where the host gave a range rather than a figure, it counts as caught if the two ranges overlap at all:

Nominal Actual
50% 62% (8/13)
80% 77% (10/13)

Close enough to nominal that the ranges can be taken at face value — with the caveat that they were calibrated on the same 13 sessions they are scored against.

The 3 misses fall on both sides: 1 drew well under the predicted range and 2 well over. There is no sign that the formula is biased high or low overall — it is simply imprecise.

Caveats worth stating plainly

  1. 13 sessions. The vote–attendance correlation is not statistically significant at this sample size (p ≈ 0.07 on log attendance; on ranks it is 0.19). This is the best available guess, not an established relationship.
  2. Attendance was recalled, not counted, and several figures are ranges.
  3. Self-selected sample — the hosts who replied. Their sessions averaged 20 attendees against a fair-share expectation of 29.
  4. P is treated as a constant 9. In reality it varies by slot, and a session competing against 5 others should draw more than one competing against 12.
  5. Time slot, room, and what a session runs against are all ignored — probably the largest missing variable.

What would sharpen this next year

  • Count attendance. A host entering a number at the end of each session turns ±25% recollections into data. By far the highest-value change.
  • Record the timetable, so each session’s real competition is known instead of a flat P = 9.
  • Vary or record turnout deliberately — with 99–107 votes on every session, the effect of the number of voters is currently theory, not measurement.
  • Give voters a limited budget (say, N stars) instead of unlimited Interested flags. “Interested” is cheap, which is very likely why Maybe votes turned out to carry no signal at all.
Next Overview