A healthcare book can get this paper's only non-price IC gain across all three horizons by counting records and covering analysts. Computing that signal takes no language model.
The authors reach the same conclusion. Their ranking "learns from price trends first, then from balance sheets and from how much attention a company receives, and little from what the text says beyond persistent company-level language, with one three-month exception." I agree. The question is how much weight that exception can bear.
Inside the ranking
Fu, Zhang, Sun, Scheller and the two Roses rank US healthcare and life-science stocks monthly from 2020 to 2025. They target log excess returns over XBI at one, three and twelve months. The universe contains 4,868 companies across 13 FactSet NAICS classes. Each month, eligibility shrinks to names above $1 and in the top half by trailing dollar volume. Their market-only score comes from a random forest using 14 price-volume features, trained walk-forward from 2010.
They then screen 187 non-market columns within each fold. These include balance sheets, scientific events, analyst ratings, and news and social sentiment scored by FinBERT and a frontier-model chain. Each column is residualized on the market score and four controls: log price, dollar volume, volatility, Amihud illiquidity. Admission requires a Newey-West |t| of 4 or more; the within-fold Bonferroni value is 3.92 over 561 tests. A ridge head combines the admitted columns with the market score to produce the hybrid.
Monthly Spearman IC is reported both raw and "neutral". To obtain the neutral measure, the authors regress the score's percentile rank on the controls' percentile ranks. They then rank-correlate the residual with unadjusted returns. They explicitly describe this as an operational label rather than a standard partial correlation.
At twelve months, the hybrid raises neutral IC from 0.063 to 0.093 (+0.030, t 3.9, bootstrap [+0.014, +0.046]). Raw IC increases from 0.289 to 0.298 (+0.009, t 2.7). The raw gains at one and three months are -0.000 and -0.001. Across the screen, only 9 of 187 columns ever qualify. None is a text score. Fundamentals contribute +0.022 at twelve months; attention, meaning record recency and analyst counts, contributes +0.021.
The headline 48% lift is in neutral IC. Raw IC improves only 3%, and the authors make no claim about returns, costs or implementation.
Does timing matter for the text?
Mostly no. For each company-month, the authors embed the trailing 30 days of text using frozen GPT-2 (2019) and Qwen3-Embedding-0.6B (2025). A ridge head maps those vectors into a ranking. Text alone produces twelve-month neutral IC of +0.039 with GPT-2 and +0.034 with Qwen3.
Shuffling the vectors across months within each company preserves the company and removes the timing. Those shuffled inputs score +0.036 and +0.041. They match or beat the actual text in five of six horizon cells. Adding text to the hybrid at the learned weight of 0.1 contributes +0.003, with intervals that span zero. A plain 50/50 average loses 0.035 to 0.040.
The authors acknowledge that the shuffle is not an identity-only baseline, and they did not run one. A replicator should start there. The shuffle retains every persistent word repeated in a company's coverage.
One scientific-text lead at three months
Scientific text covers 33% of rows. On its own, it reaches three-month neutral IC of +0.040 (t 5.4), against +0.035 for its identity placebo. No interval accompanies that margin. At a weight of 0.2 in the hybrid, scientific text raises neutral IC from 0.046 to 0.056 (+0.010, t 2.8, bootstrap [+0.006, +0.015]). The placebo blend adds -0.001. The raw IC gain is only +0.003 (t 2.2).
The post hoc subgroup gives a louder result. Among the 16% of company-months with a trial, FDA or advisory record in the window, the blend adds +0.064 (t 2.9) to a hybrid baseline of -0.023. On those same rows, the identity placebo blend contributes +0.024 (t 1.0), more than a third of the gain. The raw increment is +0.011 (t 0.4). This sample contains about 49 names a month. The horizon pattern fits a catalyst-resolution story, which the authors label as an interpretation. The increment remains the evidence.
Implementation leaves room for judgment
The paper is unusually explicit about its choices. Several still require an implementer to make a call:
- Preserving rank slots. The hybrid re-orders covered names only within the slots they occupied under the market score. The all-row scientific-text three-way blend at three months does not preserve those slots. It gains +0.017 neutral and loses 0.024 raw. Without slot preservation, some neutral gain comes from moving covered names relative to uncovered ones.
- Record recency. The authors warn that this column may capture retrieval-time coverage. Removing it takes the one-month gain to 0. The three- and twelve-month gains drop from +0.017 and +0.030 to +0.009 and +0.022.
- Choosing blend weights. Leave-one-year-out weights are fitted using other evaluation years, including later years. The authors call this retrospective.
- Qwen3 pooling. The serving endpoint supplied its own pooling and normalization; neither was controlled client-side.
The regime blend carries the same caveat. Its advantage over a static two-channel blend is +0.013 (t 2.4). Against the full hybrid, the gain is only +0.004 (t 0.8).
What can the cutoff tests establish?
The rater-cutoff design has raters score identical text, comparing those that could know the outcome with those that could not. A clean-clean placebo provides the reference for that contrast. Before the cutoffs, frontier-minus-GPT-2 gaps are at most +0.004 at one and three months. They do not shrink afterward. At twelve months, both frontier gaps fall, by +0.022 and +0.031. But the clean-clean placebo falls by +0.070 (t 2.8), leaving contamination unidentified.
These tests do not certify the embedding result. Qwen3's cutoff is undisclosed, as the authors acknowledge. The prompted forecaster does show the signature: at twelve months, the revealed-minus-blind gap falls from +0.025 to +0.008 (t 1.7). The nominally significant decline is at three months, +0.012 at t 2.5. In the authors' words, "an instruction not to recall is not a control for recall."
We could not rebuild this.
We lack clinical-trial registry and FDA advisory-committee records, as well as pre-2020 social history needed to reproduce the attention counts. The paper's non-market feeds extend back to 2010, and its text panel to 2015.
Keep the hybrid and quarantine the overlay. Every non-market column covers only companies listed in June 2026. And no correction covers the entire search, from the 16 architectures through the 66-triple blend grid. The authors identify the test that would settle the issue, and I agree: the scientific-text increment must hold on 2026 data against an identity-only baseline.