When a doctor publishes a case report describing a patient who developed a serious reaction to a drug, that report is public, timestamped, and potentially material to the company that sells the drug. It is also written in dense clinical prose. The reaction may be negated, hypothetical, or a known side effect already on the label. We propose to test a specific hypothesis. Equity prices of drug makers absorb new safety information from the biomedical literature slowly, over weeks rather than days, because the information is hard to read. If that is true, a system that reliably extracts and verifies drug–reaction pairs from the literature, and flags only the first appearance of each pair, should predict abnormal returns for the affected firm and its competitors at multi-week lags.
The opportunity matters for two reasons. First, it is a clean natural experiment on information-absorption lag. The hardness of the text is the claimed cause of the delay, and the hardness can be measured. Second, it turns a weak spot in clinical NLP, mapping free-text adverse reactions onto a standard medical vocabulary, into a problem with an external, economically grounded check. If extraction quality matters, the market signal should strengthen as extraction improves. If it does not, we learn that the signal is carried by crude features.
Core Claim
The claim has three parts. First, verified, vocabulary-normalized, novel drug–reaction pairs extracted from published case reports lead abnormal returns of the manufacturer, net of market and sector moves, over horizons longer than the next day. Second, the predictive strength rises with extraction quality: a dose-response from keyword matching, to single-pass extraction, to self-verified extraction, to human-level annotation. Third, the signal spreads with a lag to firms selling substitute drugs for the same indication. A model that aggregates over neighboring firms at the same timestamp will miss this spread; one that looks at a neighbor's past should catch it.
The closest existing work splits the problem in half. CausalStock learns time-lagged causal links between stocks. It feeds them with general financial news, which an LLM scores on five generic dimensions (correlation, sentiment, importance, impact, duration). It has no notion of negation, medical vocabulary grounding, or whether a reaction is new. BioDEX builds a large benchmark that matches PubMed papers to FDA adverse event reports and tests extraction under temporal splits, but it never links extractions to markets. Our proposal differs from both. The text source is clinical. The unit of signal is a verified, novel drug–reaction pair rather than a sentiment score. The central test asks whether extraction fidelity causally matters for the market signal.
Why This Text, and Not Hospital Records
The prompt's third direction is clinical extraction from electronic health records. Medical device surveillance with electronic health records is the motivating result. It used weakly supervised extraction on one hospital's notes and found six times more revision surgeries than billing codes recorded, and that free-text "hip pain" mentions were a significant lead indicator of later revision. That is a lead–lag finding inside the clinic. However, hospital notes are private, and public research versions such as MIMIC carry shifted dates, so they cannot be aligned to market time. We therefore treat published case reports as the publicly timestamped analogue of clinical narrative. We reserve hospital-record signals for a later extension with a data partner, where the same pipeline would run on device notes. This is a deliberate narrowing, not a claim that literature equals hospital records.
Research Program
The extraction stage starts from a fine-tuned FLAN-T5 model trained on BioDEX's training split. It adds the four-step self-verification procedure of Self-Verification Improves Few-Shot Clinical Information Extraction: extract, re-prompt for omissions, ground each claim in a verbatim evidence span, then prune claims the evidence does not support. Pruning is where negated, hypothetical, and historical reactions get removed. These are exactly the framing distinctions that single-pass financial text encoders like StockNet's recurrent tweet encoder cannot make.
Each surviving reaction is mapped to a standard medical term. Because BioDEX reports reaction extraction near 34% F1, we will spend real effort here. The plan is retrieval over the vocabulary followed by a verification pass, all run on local open models so the same pipeline could later run on protected hospital notes.
The signal stage assigns each verified pair a publication timestamp and a novelty flag. The flag records whether this drug–reaction pair has appeared before in any earlier report in the corpus. Each pair also gets a seriousness label, which BioDEX already extracts at above 90% F1. Drugs are linked to marketing firms, and firms are linked to each other through shared therapeutic indications.
The market stage measures abnormal returns over windows from 1 to 60 trading days after publication. It uses two models. The first is a transparent lagged panel regression with a distributed-lag structure. The second is a lag-aware graph model. That model follows MASTER's idea of encoding each stock's history first and then attending across stocks, so a substitute firm's response can attend to a target firm's past signal. This contrasts with the same-timestamp graph attention in MAN-SF.
We treat the regression as the primary scientific instrument. The graph model earns its place only if it beats the regression out of sample.
Baselines
The naive reference point predicts zero abnormal return, equivalently a market-plus-sector factor model. Any signal must produce a positive out-of-sample rank correlation with returns, measured with the information coefficient that MASTER reports, and the correlation must clear zero across time-blocked bootstrap intervals.
The strongest structured competitor is the count of matched FDA adverse event reports per drug, dated by report receipt. It needs no text processing and comes from the very database BioDEX links to. It gets the same lag structure and the same regression or graph model. Beating it means the literature-derived novel-pair signal adds predictive information when both enter the same model.
The strongest text competitor is CausalStock's LLM scoring of the five generic dimensions, applied to the same case-report abstracts and fed into the same downstream models. For the related market task the file reports 63.42% accuracy on the ACL18 news benchmark for CausalStock. No reported result was found on Lacuna for any method predicting pharmaceutical returns from biomedical literature, so we must establish that number ourselves.
For extraction, the bar is BioDEX's reported 59.1% F1 for fine-tuned FLAN-T5-Large, against 53.1% for few-shot GPT-4 and 72.0% for human clinicians (BioDEX). Self-verification gains elsewhere, such as 0.419 to 0.530 F1 on clinical trial arm extraction (Self-Verification), suggest headroom, but these are not guarantees for this task.
All text baselines share the same corpus, the same timestamps, the same model-selection budget on a validation period, and the same final held-out years.
The key ablations isolate the proposal's ingredients. We remove verification but keep extraction. We remove novelty filtering and count all pairs. We remove vocabulary normalization, which breaks novelty detection. We replace lagged cross-firm edges with same-timestamp edges.
Why not just…?
Why not just count structured FDA adverse event reports? This is the most serious threat, and it may well capture most of the signal. Spontaneous reports are underreported and delayed, as BioDEX notes, and structured codes miss most events in the hospital setting (Medical device surveillance). Neither fact proves that literature adds market information. The nested comparison in the baselines settles it. If novel literature pairs add nothing once report counts are in the model, the extraction story fails for markets, and we will report that.
Why not just run generic sentiment or LLM impact scoring on the abstracts? Generic polarity treats "no evidence of hepatotoxicity" and "fatal hepatotoxicity" as similarly negative. Topic-blind sentiment was already near chance on financial message boards, at 50.36% against 56.43% for topic-aware modeling (TSLDA). Still, an LLM may implicitly sense novelty and seriousness. The CausalStock-style baseline on identical text tests this directly.
Why not just use news? If markets react to news coverage of a case report, the literature signal could be a slow proxy for news. We will split events by whether any major news coverage of the drug–reaction pair appeared within the window. The delayed-absorption hypothesis predicts that the lagged effect concentrates in the uncovered events. If the effect appears only after coverage, the honest conclusion is that the lag lives in journalism, not in reading difficulty.
Why not just a linear lag regression instead of a graph model? It may suffice, and it is our primary analysis. The graph model is retained only if it improves out-of-sample prediction of the substitute-firm response.
Proposed Experiments and Decisive Checks
We build the event corpus from BioDEX's abstracts and full texts, using its temporal train, validation, and test splits. This keeps every extraction model from seeing future reports. Daily prices for US-listed drug makers, plus a drug-to-firm and indication map, will be assembled by us, since the file contains no such resource.
Success on extraction means beating 59.1% overall F1 on BioDEX and materially improving reaction F1 above roughly 34%. Success on prediction is measured by information coefficient and rank information coefficient on abnormal returns at each horizon, cumulative abnormal returns around novel serious events, and the Sharpe ratio of a long-short substitute-versus-target portfolio after transaction costs.
The decisive check is the dose-response. We evaluate the market signal produced by keyword matching, single-pass extraction, verified extraction, and, on a subsample, human-annotated pairs. The claim survives only if predictive strength rises with extraction F1, and if the novel-pair signal beats the matched structured-report counts.
A second decisive check is the lag profile itself. Mass at horizons beyond one week, concentrated in hard-to-read, uncovered reports, supports information-absorption delay. Mass on day one supports fast pricing and kills the delay hypothesis. Either outcome is publishable, and only the first justifies the graph model.
References
- CausalStock: Deep End-to-end Causal Discovery for News-driven Multi-stock Movement Prediction
- BioDEX: Large-Scale Biomedical Adverse Drug Event Extraction for Real-World Pharmacovigilance
- Medical device surveillance with electronic health records
- Self-verification improves few-shot clinical information extraction
- Stock Movement Prediction from Tweets and Historical Prices
- MASTER: Market-Guided Stock Transformer for Stock Price Forecasting
- Deep Attentive Learning for Stock Movement Prediction From Social Media Text and Company Correlations
- Topic Modeling based Sentiment Analysis on Social Media for Stock Market Prediction