Can clinical-trial and research-paper data beat biotech stocks? (2012-2026 study)
The short answer
- One idea works against the average small biotech: buy a stock 2-5 months before the window in which the company has said trial results will come out, and sell a few days before that window. It beat an equal-weight basket of all biotech stocks in all three periods, by about 4-12 points a year (corrected 2026-10-23, see the note under section 3).
- But it does not beat the biotech ETF XBI, which held more of the bigger biotech winners. Before testing we set 'beat XBI' as part of the pass bar, so the verdict is a fail.
- Gains come from a few trades. Without its 10 best trades the strategy roughly breaks even.
- Research papers did not help us spot winning drugs. We took 2,105 past trial results and FDA decisions and looked at everything public before each one: 186,000 PubMed papers on the drug (human vs animal studies, randomized trials, top journals, outside authors), the statistics in the company's earlier results, and the trial design. None of it predicted success better than four simple facts (FDA decision or trial result, phase, cancer or not, the company's track record). An AI-reading test of the papers' actual results gave the same answer.
- Holding through the trial result is a losing bet, and so is buying after a failed trial.
1. The data
| Source | What we took from it | Point-in-time care |
|---|---|---|
| ClinicalTrials.gov | 25,462 trials matched to 1,291 listed biotech companies (2 audits of 50 matches: 100/100 correct) | Only facts fixed when a trial was first posted. Today's status, enrollment and completion dates were edited after the fact (only 4% of Phase 3 records were last updated before completion), and the registry's edit history is blocked. |
| SEC 8-K press releases | 41,000 releases read: topline results, FDA decision dates, approvals, rejection letters, clinical holds, and 14,000 statements of when results are expected | Dated by filing; trades start the next morning |
| SEC financial filings | Cash, cash burn and months of runway for every company | Usable from the day after filing |
| PubMed, Europe PMC, OpenAlex | Papers and preprints on each drug's code name, company research output, citations counted as of each past year-end | Each paper counts from 30 days after it entered PubMed |
| Massive | Daily prices of every US stock incl. delisted and acquired companies | Survivorship-free |
2. What the price moves look like (2012-2019)
| Event | Average move vs the biotech basket | Events |
|---|---|---|
| Last 60 days before a pre-announced results or FDA date (vs similar-size biotech) | +4.1% | 753 |
| Failed trial, 3 days | -33.9% | 109 |
| FDA rejection letter, 3 days | -20.5% | 74 |
| Clinical hold, 3 days | -10.8% | 38 |
| Share offering, 3 days | -2.7% | 4,459 |
| 'Positive' topline result, 3 days | +8.7% | 430 |
Stocks tend to drift up as a known results date approaches, and that is the effect the strategy tries to capture. Many 'positive' headlines are Phase 1 or partner news, so while the average positive result jumps, the typical one (median) moves only about 1%; failures are brutal (-54% for companies under $300M).
3. The strategy and how it was tested
Rules: every Friday, take US biotech stocks above $2 with at least $1M of daily trading and a $50M-$10B market value. Keep those whose company has said in an SEC filing that trial results will come in a window starting 40-100 trading days from now. Buy up to 5 a week (4% of the account each, max 25), nearest window first. Sell 3 trading days before the window starts, or the morning after any results, approval, rejection or clinical-hold filing. Costs 0.75% per side (0.25% above $2B).
An automated research loop tried 21 versions on 2012-2019 only, one change at a time, with two independent reviews checking for look-ahead and luck. The final version and two variants were written down before the 2020-2022 and 2023-2026 data were opened, once each.
| Period | Strategy | Equal-weight biotech | XBI | Strategy worst fall |
|---|---|---|---|---|
| 2012-2019 (tuning) | 15.1%/yr | 11.2% | 20.1% | -35% |
| 2020-2022 (one look) | 0.0%/yr | -12.4% | -4.3% | -55% |
| 2023-Oct 2026 (held back) | 11.9%/yr | 5.8% | 16.9% | -47% |
Year by year it beat the equal-weight basket in 9 of 15 years and XBI in 4. The best trades in the held-back period were REPL (+295%), PRAX (+286%), KOD (+226%) and NKTR (+223%); 2025 alone produced most of the gain.
4. Ideas that did not work
| Idea | Result (2012-2019) |
|---|---|
| Hold through the trial result, picking trials with better odds (randomized Phase 3, clean track record, cash runway) | Lost money; worse than random picks |
| Buy after a failed trial when the stock trades near its cash | Too few tradable cases; lost money |
| Rank stocks by research-paper and citation growth | Worse than the simple rule (see section 5) |
| Machine-learning model to choose among the candidates | No better than picking at random among the same candidates |
| Buy after a positive result once the follow-on share offering is done | Few trades, gains from a handful of names |
| Trailing stop | Cut the big winners; removed |
5. Can the papers tell a good drug from a bad one?
To test the idea that research papers show which compounds are good, we worked backwards from 2,105 trial results and FDA decisions (2012-2026): about 1,100 positive and 335 negative trial results, 430 approvals and 183 rejections, each matched to its drug (87% of releases). For each, we collected only what was public beforehand. Patterns were found on 2012-2019 and checked once on 2020-2026.
| Pattern (share of clear outcomes that were successes) | 2012-2019 | 2020-2026 | Held up? |
|---|---|---|---|
| The drug already had a positive earlier result | 58% vs 39% | 46% vs 46% | no |
| Its last results release said 'statistically significant' 3+ times | 77% | 41% | no |
| Its last results release reported p < 0.001 | 63% | 36% | no (reversed) |
| Randomized, double-blind, placebo-controlled trial | 46-49% vs 36-37% | 43-45% vs 48-50% | no (reversed) |
| The company already had 1-2 positive trial results | 52% vs 37% | 55% vs 41% | yes |
| Company worth under $300M | 37% | 33% | yes (worse) |
A prediction model built from four basic facts ranked a winner above a loser 55-64% of the time (50% = coin flip). Adding all the paper and earlier-release evidence improved that by about zero. We then had an AI model (Claude Haiku, with drug, company and date names hidden, no internet access, and every answer required to quote the source text) read the papers' and releases' actual results: endpoint met, effect size, p-value, patient numbers. On 200 results from 2012-2019 it added nothing either (+0.7 points, within noise). One limit: even with names hidden, the AI recognised the drug in 37% of cases (57% for famous drugs), so an AI cannot be fully stopped from knowing history. That is why it was only allowed to copy facts from the text, never to judge.
Our reading: the obvious evidence in papers is already in the price before the result, and a 3-day stock move measures surprise rather than how good the science is.
How much to trust this
- The edge comes from which stocks are eligible (an announced results window ahead), not from clever ranking: ordering the same candidates at random scores about the same.
- It depends on a minority of big run-ups; only about half the trades win.
- After counting every version tried in this project (377 tests), the chance that its edge over the equal-weight basket is real is low (deflated Sharpe 12%).
- The text classifier misses some results announcements (about 40% of readouts came before the guided window), so some positions still rode the binary result.
- The equal-weight basket pays about 3% a year in rebalancing costs; XBI is cost-free and leans to bigger companies, which did far better than small biotech in this period.
No redactions applied.
Published via wiis. Originals are encrypted at rest; this page reflects the redacted version. report abuse.
Discussion
No comments yet.