To test whether or not political pundits understood the political moment they were in - and that those moments were turning points in the campaign, I ran an analysis of language used by 71 political newsletters and podcast transcripts, gathered by me between September 1 and November 14, 2026. The analysis used seven “turning points” I chose, based on post-election armchair analysis by pundits:
Conclusion: Most of the identified dates did not show a significant shift in tone - except for the launch of Ramaswamy’s ad among left-wing pundits and the Haley Stevens endorsement among centrists. When viewing the data as a continuum, it is hard to see that pundits had any kind of consistent foresight, and that when viewing events that have occurred, there seems to be almost no connection between “major” events and how pundits discussed the likely election outcome.
For every day, how different was the language in the 5 days starting that day from the 5 days before it? Higher means a sharper break. Numbered lines mark your seven dates.
Each cell shows how unusual the vocabulary shift was, as a percentile of all other dates: 90 means only 10% of dates changed more. Click a row for the detail.
Five-day average, with every source weighted equally so that one prolific source or a long podcast cannot move the line. Rates are per 1,000 words.
The terms that rose and fell most against the previous week, with each week's tone levels.
Uses the Jensen-Shannon divergence between the 5-day-before and 5-day-after distributions of two-word terms, limited to terms a group used at least 25 times over the period so that rare phrases do not create false movement. The percentile compares each date with every other date that has a full 5 days on both sides (61 dates). The weekday-matched percentile compares only with dates that fall on the same day of the week, because newsletters and podcasts publish on different days.
Each document gets a score per 1,000 words. Each source's before and after rates are compared, and every source counts equally. Whether the shift is bigger than chance comes from shuffling documents between the two windows within each source 1,500 times. With 147 tests, some will look significant by luck, so q adjusts for that (Benjamini-Hochberg); q below 0.10 is the standard to trust.
If the sources publishing in the two windows differ, vocabulary can move without anyone changing how they write. "Source mix changed" appears when the mix differs more than on 85% of other dates.
Word-list tone scores miss sarcasm and negation. Several right-leaning newsletters only began arriving in early September and nearly all went quiet from Sept 11 to 16 after an ingestion failure, so the right group's September results are unreliable. This version adds recovered Fastmail columns (Bouie, Bruni, French, Wallace-Wells), transcripts from Bongino, TrueAnon and The Dispatch Podcast, full article text for The Federalist and City Journal, and The Morning Dispatch and Dispatch Energy from email (44 Morning Dispatch emails for about 54 weekdays). Just the News is left out because it has no articles from Oct 18 to Oct 31, and Commentary Magazine could not be fetched in full. The Young Turks is not collected, The Dispatch Podcast stops on Nov 6, and one small right-leaning source has no items from Oct 18 to Nov 7. Dates close together (Sept 8 and 10; Oct 28, Nov 1 and Nov 3) share days, so their results are not independent.