Novartis had three phase 3 readouts in six trading sessions. On 1 September remibrutinib met its primary endpoint in relapsing multiple sclerosis and the stock rose 6%. On 4 September pelacarsen missed its primary endpoint in a cardiovascular outcomes trial, which is the largest and most expensive study design a drug company ever runs, and the stock fell 1.9%. On 8 September del-desiran missed its primary endpoint in myotonic dystrophy and the stock fell 13.9%, gapping down before the market opened.
That ordering makes no sense on the face of it. The cheap failure cost seven times more than the expensive one. I wanted to know whether that was strange or ordinary, so I went and got every trial readout I could find.
Good news is already in the price
A drug company's share price is a bet on the future, and before a trial reports that bet already contains everyone's guess about whether the drug will work. The guess is usually optimistic, because nobody spends hundreds of millions on a final-stage trial they expect to fail.
So when the drug works, very little happens. The market had already assumed it would, and learning that you were right is worth almost nothing. When the drug fails, that is genuinely new, and the price falls to whatever the company is worth without it.
Across 1,394 clean events, meeting the primary endpoint moved the stock a median of 0.7% against the sector. Missing it cost a median of 17.0%, and the worst quarter of failures lost more than 61% in a single session. About one success in four went down anyway.
The later the trial, the worse the fall
Filter the panel below by phase and the pattern sharpens. A phase 2 failure costs a median of 14.4%. A phase 3 failure costs 18.5%. Success pays roughly nothing at either stage, so the phase does not change what you gain, only what you stand to lose. That is what it looks like when the market is confident and occasionally wrong.
A third of failures cost nothing at all
The failure side is not one hump. There is a heavy cluster of collapses, and a second cluster sitting almost exactly on zero, where about 34% of failures move the price less than five percent in either direction.
Those are the drugs nobody was counting on. Either the market had already written the programme off, or the asset was too small to matter against the rest of the pipeline. A failure only hurts if somebody believed, which is the same reason a large diversified company can absorb a miss that would end a single-asset biotech.
The averages lie
The mean move for a successful readout is +4.7% while the median is +0.7%, because a handful of enormous days drag the average up. Amarin's REDUCE-IT in 2018 was worth +314%. Madrigal's MAESTRO-NASH in 2022 was +270%. ProKidney in 2025 was +513%. I checked all three against the surrounding sessions, and the new price levels hold for days afterwards, so these are real repricings rather than the stock split artifacts that usually produce numbers that size.
Most successes do nothing and a few are transformative. The average describes neither.
Where the data came from
Companies file topline trial results with the SEC as a material event, so the full text search covers them. I swept every 8-K and 6-K from 2015 onwards for the language a readout uses, which produced 21,189 candidate filings. A filing that contains the words is not the same as a filing that announces them, so each one is classified from the headline of its press release rather than from a keyword match, which leaves 2,376 genuine readouts across 546 companies.
Prices come from daily bars, and every move is measured against the biotech sector on the same sessions, so a stock that fell because the whole market fell does not count as a reaction to its trial.
What I had to fix before I believed any of it
Almost every bug I found in this pipeline failed the same way. Each one turned a real signal into a silent skip or a wrong label, nothing ever crashed, and every total still looked plausible afterwards. That is the failure mode to fear in this kind of work, because there is nothing to notice.
The worst was a word list containing lease without word boundaries, which
matched the word RELEASE inside FOR IMMEDIATE RELEASE and quietly discarded
most genuine press releases. The most consequential was a ticker pattern that
only matched a single symbol, so every dual-listed company was dropped for
having no ticker at all. That one silently excluded AstraZeneca, Sanofi,
Novartis, GSK and Takeda, which are among the largest trial sponsors in the
world.
A third read the words "meet primary" inside "does not meet primary endpoint" as a success, because the pattern never looked left at the negation sitting in front of it. Fixing that alone moved the median failure from 13.2% to 17.0%, since the events it had been mislabelling as failures were mostly ones that barely moved.
The one that best illustrates the problem is FibroGen. Its pamrevlumab failure in idiopathic pulmonary fibrosis priced at minus 83%, one of the largest single day losses in the whole study. The price API returns a plain 404 for FGEN because the company was renamed, and before I resolved renames through the company's SEC identity rather than its ticker, that observation did not exist. The largest failures are exactly the ones most likely to vanish from a naive dataset, which biases the answer in the most flattering possible direction.
What is still wrong with it
Three biases remain, and they do not point the same way, so correcting one without the others would make the answer worse rather than better.
Companies that fail badly stop trading, and 13.8% of readouts have no usable price for that reason. Those missing events are concentrated in the tail this piece is about, so the real left tail is heavier than what you see here.
Failures are announced more quietly than successes. A win gets its own filing while a loss is often folded into the next quarterly report, which my classifier correctly discards as an earnings release. The announcement record itself is therefore biased toward good news, on top of the survivorship gap.
Companies sell shares into good news. Around 11% of positive readouts come with an equity offering filed the same day or the next, against 4% of failures, so dilution lands on the upside roughly twice as often as on the downside. Those events are excluded here, which cleans the sample and simultaneously makes it less representative.
I can also put a number on how often the classifier is simply wrong. If a trial that supposedly met its endpoint falls more than 40%, or one that supposedly missed rises more than 25%, something is probably mislabelled. That happens in 1.8% of labelled events. It was 2.4% before I fixed the pattern that read "Does not meet primary endpoint" as a success, because it matched the words "meet primary" without ever looking left at the negation in front of them.
The honest summary is that 1,394 of 2,376 readouts survive both a price check and a clean window, and the 41% discarded are not a random sample of the rest. Every number above is a description of what happened on the day of a readout. None of it is a claim that the readout caused it, and none of it is advice about a trade.