The engine ships Stage as four baskets, and one of them is Commercial / Pending Approval — 101 revenue-generating companies blended with 23 that have filed and are waiting. They did not trade alike. Since 1 September Commercial fell −8.4% and Pending Approval −14.4%; since the FOMC, −3.0% against −9.0%. Pending Approval is statistically indistinguishable from Phase 3 (Welch t = 0.09 since 1 September) and clearly distinct from Commercial (t = −2.96 since the FOMC). 18 of 101 Commercial names held through both windows. Zero of 23 Pending Approval names did — the worst hold rate of any bucket on the stage ladder, below Phase 1/2.
A filed asset has the duration of a clinical one and none of the revenue, so a rate shock discounts it like clinical risk while the blended basket credited it with commercial defensiveness. Splitting the two raises Stage's adjusted explanatory power despite costing a degree of freedom — 3.1% to 3.4% since 1 September, 2.1% to 2.6% since the FOMC — so this is signal, not a finer cut of noise. This chart and the composition chart below use the split; the basket index and the 38-basket ranking above still carry the blended bucket, because those compound a daily index whose original construction does not reproduce here to the decimal, and one chart should not mix two methodologies.
Splitting the blend exposed a staging error rather than a missing factor level. Of the 27 names the engine stages as
Filed, four are selling product: AXSM, ORGO, AQST and CGON. Each carries cost of goods
sold and balance-sheet inventory, which a company awaiting its first approval does not — AXSM also books a $164M
quarterly sales & marketing line. They are counted as Commercial here, and the fix is queued for
company_factors.
Revenue alone was not the test. ARVN ($317M), PTGX ($282M) and VRDN ($71M) report more revenue than AQST or CGON and stay in Pending Approval, because it is collaboration and milestone revenue — no cost of goods, no inventory, no marketed product. That is lumpy deal income, not the cash flow that made Commercial defensive.
The correction does not change the conclusion, which is the point of checking: Pending Approval moves from −14.3% to −14.4% since 1 September and is unchanged at −9.0% since the FOMC, still with zero of 23 holding. A further split is not supportable anyway — four names is below the five-name floor every bucket on this page has to clear.
It looks like the second real factor — 1.9% explained, p = 0.02. It is not. Its best bucket, runway > 60M, is 81% commercial companies; every other runway bucket is 5–15% commercial. Hold stage constant and Cash Runway explains 0.0% of what is left, within clinical names (p = 0.48) and within commercial names (p = 0.97) alike. The ordering gives it away too: since the FOMC the best runway bucket is < 12M — the companies closest to running out of money.
The cause is that runway is not a defined quantity for a company with revenue. Sarepta carries $934M of cash and investments and quarterly operating cash flow of −$5.6M, −$202.7M and −$336.7M across three consecutive quarters — swings driven by unearned revenue and payables, not burn, against positive operating income. Any runway computed from that is an artefact. The after stage column exists so this cannot be read off the chart wrongly again: it demeans every name by its own stage bucket and re-runs the test. Nothing clears the null. Stage is the only factor in the set that explains the September cross-section, and it explains 3.4% of it.
- Memberships are point-in-time through 17 September, the last snapshot the engine holds, then carried forward. Any name whose factors changed in the final eleven sessions is mis-assigned.
- Residual is not a risk model. It subtracts the simple average of a name's basket returns from its own. Names sit in several overlapping baskets, so this attributes loosely.
- LQDA fell 57.2% on 30 September on 23M shares against ~1M normal, no split — the largest single move in the universe, and it post-dates the factor file.
- ROG excluded: the price feed maps it to Rogers Corporation, not Roche.
- Short interest is as of the 15 September 2026 FINRA settlement, short shares over float, from the Massive API with float from the price vendor — 402 of 411 names. Five are dropped where reported short shares exceed the vendor's float, which means the float is stale rather than the short interest wrong. An earlier version of this page called the field unreliable on the strength of a test that divided short shares by float but multiplied by shares outstanding; that test was wrong, and the field is sound. The previous values were correct too, just computed from the 31 August settlement: the median name moved 0.46 points on the update, though 24 moved more than 5 and CAMP moved 40 as its float changed.
- Stage 4 reads as “Pending Approval” here, where the engine’s stage field calls it “Filed”. Same 27 companies. The split of Commercial from Pending Approval applies to the attribution and composition charts only; the basket index, the 38-basket ranking and the correlation chart still use the engine’s blended Commercial / Pending Approval basket.
- 153 clinical names carry no dated readout at all and are treated as having nothing before year-end. That assumption is not safe. RVMD sat in the run-up cohort until its Q4 2026 combination CRC data update was found on slide 36 of the company’s own August 2026 deck — the engine catalysts table holds no such row, only a 2023 trial initiation. One deck changed RVMD from long-dated to six weeks out. The same gap may sit under the other 152.
- Time to Catalyst is rebuilt, not the engine field. Meaningful events only — Phase 2 data, Phase 3 data, PDUFA — from the engine catalysts table plus the Leerink, Guggenheim, TD Cowen and Jefferies calendars. Unlike the other ten groups it is held fixed across the window rather than point-in-time, and > 24M holds only 4 names. Preclinical holds 19.
- Correlations use 44 sessions since 1 August. Only three had the 10-year up 8bp or more, so single-day rate sensitivity is not measurable here.