Can a computer guess what conference will talk about?
Probably not very well. But it’s a fun test. We took the ten subjects conference has talked about most since 1971 and asked five forecasters to guess how much of the next conference each one will come up in. Then we checked how each did on the last ten.
This is a test of the forecasters, not a prediction of what anyone will say. Talks come from inspiration, not trend lines.
Loading the forecasts...
How each one guesses
Same as last time
No math at all. Whatever share a subject got last conference, it guesses again. It sounds too simple, but it’s hard to beat, so every other forecaster has to beat it to count.
Prophet (Meta)
Draws a trend line through all 55 years, lets it bend where the trend changed, and adds a small April-or-October bump. Easy to explain. Slow to notice a sudden jump.
Chronos (Amazon)
An AI model trained on lots of other time series, from sales to weather. It has never seen conference before. It recognizes the shape of a series and guesses what usually comes next.
Timestral (Mistral)
Same idea as Chronos, newer and tiny: about 5 million settings, small enough to run on a laptop in a blink.
t0-gamma (The Forecasting Company)
Same idea again, and the newest of the three. It came out in October 2026.
We’ll grade it
After the next conference we’ll put the real numbers here and see who came closest. Then we’ll guess the one after that.
Prophet is MIT licensed, Chronos and t0-gamma are Apache 2.0, and Timestral is under Mistral’s Modified MIT license. The forecasts were worked out once, ahead of time. Nothing runs while you read this.