Is our Elon tweet-count model actually getting better? Not that we can prove.
We log every forecast our model makes for Elon Musk tweet-count markets and score it against the real count after the market closes. This page audits that record honestly, including the parts that don't flatter us — 211 forecasts across 90 markets, March–July 2026, all made 24–72 hours before close.
The short version: the raw numbers show accuracy improving, but that entire gain rests on one anomalous month (June). Remove June and the trend vanishes; the most recent month (July) reverted right back to where we started. And the model's stated 90% prediction band is over-confident — the true count landed inside it only 67% of the time. This is an accuracy audit, not a claim of any edge.
90% band actually covers
67%
per forecast (77% per market) — nominal is 90%
Forecasts audited
211
across 90 Elon markets, 24–72h horizon
Durable improvement?
No
trend collapses to ~0 once June is removed
June's error dips to 11.9% and coverage hits 96%, then July reverts to 41.5% / 50%. Every month except June sits below the 90% coverage line.
What the record actually shows
Across all 211 forecasts, the model's stated 90% interval contained the true count only 67% of the time (per market, 77%). Both are materially short of 90%: when the model says 90%, treat it as roughly 70%.
June's improvement is real, not noise — its 11.9% error versus the March–May baseline of 39.5% is a large, significant gap. But it's one month. The correlation between a market's close date and its error is r = −0.22, a genuine-looking improvement trend; drop June alone and it collapses to r = −0.02, statistically indistinguishable from zero. No durable trend survives removing one month.
We checked whether June was just an easy-markets illusion, and it isn't. Larger markets are easier to forecast (error vs size, r = −0.38), but later markets aren't larger (close date vs size, r = −0.08) and the forecast horizon is flat over time. So difficulty drift works against an inflated-skill story rather than manufacturing one — but ruling out that confound doesn't turn a one-month spike into a trend.
Two honest limits on all of this: every forecast here is Elon-only, and every one is from the 24–72h window — never the final-16h window that would actually matter for a position. The sample is small (90 markets, 122 days) and 27 model versions are entangled with calendar time, so read every trend as suggestive, not established.
Each dot is one prediction; y is its absolute % error (clipped at 100%). The good run is a June cluster of low-error green dots, not a steady march — and the red misses make the over-confident band visible.
Month by month
| Month (2026) | Forecasts | Markets | Mean |error %| | Coverage |
|---|---|---|---|---|
| Mar | 46 | 20 | 46.8% | 48% |
| Apr | 47 | 20 | 36.1% | 53% |
| May | 50 | 22 | 36% | 70% |
| Jun (anomaly) | 56 | 23 | 11.9% | 96% |
| Jul (partial) | 12 | 5 | 41.5% | 50% |
This page is an accuracy audit, not a betting claim. It measures point and interval accuracy for Elon Musk tweet-count forecasts made 24–72 hours before close, and nothing else — no profit, ROI, or tradeable edge; nothing about the final-16-hour window; nothing about any other person. The sample is small and model versions are entangled with calendar time, so every trend here is suggestive, not established. June was anomalous and is not representative. Numbers are computed from our logged forecast record; past accuracy does not predict future accuracy.