← Back to the feed
predictions analyzed
Snapshot as of —
Does the mood get darker the further out the prediction?
Sentiment (VADER) by predicted era. Excludes joke far-future posts (year 3000+, ~4% of the dataset) as a separate tonal category — see the note below.
Negative Neutral Positive
Average sentiment drifts steadily downward with distance: near-future predictions (2026–2049) average −0.026, mid-future (2050–2099) −0.037, far-future (2100–3000) −0.064. The joke-year posts (predicting year 3000+, e.g. "the year is 9989") buck the trend entirely at −0.028 — once a prediction stops being a plausible extrapolation and becomes pure absurdism, the doom fades.
Do gloomier predictions get more engagement? Checked it directly: sentiment and likes are essentially uncorrelated (Spearman ). A prediction's mood has no measurable bearing on how many people like it — doom doesn't sell any better than hope does here.
What the predictions are about
Discovered via topic modeling (TF-IDF + NMF over the post text) — not hand-picked keywords. A post only gets a label when one topic clearly dominates it; near-duplicate clusters (four separate government-related topics: Trump, Biden, White House, US politics) were merged by hand after reading their top terms. "Other" is still the honest majority — most posts don't cluster into any single clean topic.
Top 5 predicted years
Which specific year gets predicted most often — not just which decade.
Top 5 people named
Named via spaCy NER — unsurprisingly, almost entirely real-world politicians rather than fictional or invented figures. Name variants are merged (a bare "Trump" counts toward "Donald Trump"); distinctly-named relatives like "Barron Trump" are kept separate.
Top 5 organizations named
One NER mistake removed by hand: "Trump" was tagged as an organization in some sentences — clearly a person, folded out rather than left in as noise.
Top 5 places named
Aliases are merged into one canonical name ("US", "America", "the United States" → United States).
Top 5 most-liked predictions