Changelog

We publish statistics people may act on. A number is only trustworthy if you can see what happened the last time it was wrong. Nobody else in this category publishes a changelog. That is the point.

We had grown it to 1,986 using an SEC-filing crawler. Measured recall showed the crawler sees only 38% of large-cap presenters vs 57% of small-caps — a poster is material to a $200M biotech (8-K filed) and immaterial to Merck (no filing). Adding those events re-weighted the study toward micro/small and moved the nano median for sampling reasons, not real ones. Full correction.

The published figure was computed on a catalyst panel containing 202 duplicate rows. Deduplicated, the median is +0.43% (n=1,661). The conclusion is unchanged — still within half a percent of zero — but the number was wrong.

It resolved a conference name to that conference's next occurrence, ignoring the year written next to it — so a 2026 filing saying “presented at ESMO 2025” produced an upcoming ESMO 2026 event. Five phantoms. The rule now: a stated year always wins, and a past year can never produce a future event.

The crawler wrote raw press-release text into CSV without escaping quotes. One " inside a snippet destroyed the file from that row to end-of-file. It now writes QUOTE_ALL and re-reads its own output, failing the build if the row count doesn't round-trip.

We pulled a row labelled “ANE” calling it a parsing artifact. It was a garbled ENA — the AACR-NCI-EORTC conference, which is real. Restored the 23 genuine October events; marked the rest unknown.

Every presentation had been bucketed by the company's market cap today. A company that was mid-cap in 2018 and has since collapsed was filed under nano — so the nano bucket was enriched with companies that later fell apart. Now computed as of the event date.

Our Conference Overlay claimed nano/micro presenters ran up +4.88%. Our own 1,425-event study found the nano median is negative. Refuted and retired.

“68% of readouts do nothing” would have travelled. The bucket it came from spans −14.9% to +5.0% — a 20-point range is not “nothing.” We published the honest cut instead: 56.5% land within ±5%.

A reviewer said our 1-day return column was mis-scaled and the 25th percentile was −521%. It is already in percent; the real figure is −5.19%. Acting on it would have corrupted our most valuable dataset. A corrections page that only shows the times we were wrong is marketing.