Skip to content
james_tavita

~/projects/general-conference/forecast

Forecasting the next conference

Every conference talk to date is a record of what the Church was saying; this page asks what it will say next. From 112 conferences since 1971 it forecasts April 2027's topic mix, the verses most likely to be cited and the phrases about to rise, and it keeps score: the same method is run on each of the last 20 conferences from the data before it, against the naive forecasts of "same as last time" and "same as always".

Supervised112 conferencesNext: April 2027Track record ~/general-conference

What stands out

  • The forecast beats “same as last conference” on topic mix 75% of the time and “same as always” 100% of the time. Mean Jensen–Shannon distance 0.379 against 0.424 and 0.565: recency-weighting the last few conferences is better than either extreme.
  • 31% of the forecast top 25 verses make the actual top 25, against 18% for simply repeating last conference's list. The verse list is the most predictable thing about a conference.
  • Rising phrases keep rising only 30% of the time, and rising words 19%. A phrase that surged last conference usually settles; the forecast says so rather than chasing it.
  • For April 2027: god’s · lord’s · savior’s leads the forecast at 63.4% of talks (last conference 63.9%); the verses most likely to be cited are Alma 7:11, Matthew 11:28, John 3:16; the phrase most on the rise is “along the covenant”.
  • This page keeps score in public. When April 2027 enters the corpus, its row fills in and the forecast rolls forward; the track record below is the record so far.
Loading the study…

Model card

Data
112 conferences from 1971 to October 2026, modern transcripts only, talks of 300 words or more. Per conference: the share of talks assigned to each of the thirty topics of the topic study, citation counts per verse, the share of talks using each tracked phrase from the phrases study, and the share of talks using each of the common words.
Forecast
Exponentially weighted means over past conferences with a half-life of 6 conferences for the topic mix and verse counts; the verse forecast is the weighted top 25. Rising phrases and words are those whose mean share over the last four conferences exceeds its share over the eight before by most, among terms above a floor of recent use.
Scoring
Each of the last 20 conferences is forecast from the conferences strictly before it. Topic mix is scored by Jensen–Shannon distance against the actual mix, with "last conference" and "long-run mean" as baselines; verses by the share of the forecast top 25 in the actual top 25, against repeating last conference's list; rising terms by the share that rose again.
Known failure modes
Topic assignment gives each talk one topic, so the mix is coarse and noisy for small topics. The verse list rewards perennials; a conference with a theme breaks it. Names of newly called leaders surge as words and are filtered as proper nouns, which can also drop a real term. The forecast knows nothing a reader of the news knows: a new president, a new programme, a world event.

Methods and data for the corpus as a whole.