How the numbers are computed

Momentum

Every meaningful change in a company’s public record is written to an event with a base weight. Momentum is the recency-weighted sum of those weights, compressed onto a 0 to 100 scale.

The weights

A trial advancing a phase is the single loudest thing a preclinical company can do, so it carries the most weight. A grant renewal is real but expected, so it carries little. These are tuned by judgment against observed outcomes, not learned from labels, and they will move as the graph accrues history.

Why the log

Without compression, a company with a hundred small awards would outrank a company that just put its first molecule into humans. The natural log flattens the top of the range so volume cannot buy a score, and multiplying by 20 puts a realistic cluster of events in the 60s and 70s rather than pinned at 100.

Worked examples

One trial registered today

63

The same trial, 180 days later, with nothing since

50

A grant, an SBIR Phase II, a trial and a Form D, all in the last two months

84

One grant, three years ago, silent since

3

Two consequences worth naming. Momentum decays on its own: a company that does nothing slides down the feed without anything bad happening to it. And a future-dated event, which sources do emit, is clamped to today rather than scoring above a same-weight event happening now.

Runway

Runway is the last disclosed raise, minus a modeled monthly burn, counted forward to today. It is the most useful number here and the least certain one, which is why it is labeled an estimate everywhere it appears.

The burn model

Burn is a function of stage and therapeutic area, and nothing else. It is deliberately coarse, because it is a model rather than a disclosure: we cannot see headcount, and a company that discloses its burn is already past the window Biodatica covers. Oncology and rare disease cost more per patient-month; neuro and immunology sit in the middle; everything else runs leaner. Here is the entire table the estimate uses.

Modeled burn per month, in US dollars.

A worked example

A Phase I oncology company filed a Form D for $26M four months ago. The model burns $978k a month at that stage in that area, which funds 26.6 months. Four of those are spent, so the estimate reads ~23 months.

The stage ladder

Stage is monotonic in evidence: it is the furthest-along thing we can point at in the public record, and never an inference beyond it. A company sits at the highest rung it has earned.

1. 1Grant-only · A federal award exists. There is a bet, but no company yet.
2. 2Formed · A Form D issuer or an SBIR firm record proves a real entity exists.
3. 3Preclinical · A granted patent plus an entity: protectable lab work is on file.
4. 4IND-enabling · Reserved for evidence of work toward the clinic. A registered study with no phase, which is what ClinicalTrials.gov reports for device and observational work, cannot push a company past this rung.
5. 5Phase I · A registered Phase I or Early Phase I study.
6. 6Phase II · A registered Phase II study.
7. 7Phase III · A registered Phase III study.
8. 8Filed/Approved · A Phase IV study, or an approval, clearance or designation on file with the FDA.

Entity resolution

There is no shared key across these six datasets. A grant is awarded to a university and a principal investigator; the spin-out files a Form D under a Delaware entity; the trial is sponsored under a slightly different name; the patent is assigned to a third spelling. Deciding that these are one company is the hard part of the product, and it is a judgment. Here is how the judgment is made.

1. 1 · Normalize. Legal suffixes stripped, case and punctuation folded, principal-investigator names reduced to surname plus first initial so “RAO, PRIYA K” and “Priya Rao” collide.
2. 2 · Block. Cheap candidate generation by trigram similarity on names and by people in common.
3. 3 · Embed. Each record’s name, abstract and field are embedded and stored in a vector index. High cosine similarity across two sources is evidence that they describe the same science.
4. 4 · Score and decide. The signals combine into one confidence through a transparent weighted sum rather than a learned model, because every merge has to be explainable to an analyst who asks why.
5. 5 · Record the evidence. Every confirmed link stores its method and its score. That record is what the dossier shows you.

What each signal is worth

The two thresholds

One more rule worth stating: academic institutions are not companies. An NIH grant to a university describes a bet, but the entity we track is the spin-out that comes later. Grants to universities, hospitals, institutes and government labs are ingested as provenance, and they never create a company on their own. They attach to one only through a shared principal investigator.

What these numbers do not know

This is the most important section on the page. Every limitation below is structural: it comes from what the federal record does and does not contain, and no amount of engineering on our side removes it.

• Money we cannot see

  State grants, foundation awards, venture debt, convertible notes, revenue, and a parent company writing checks are all invisible to the runway model. A company reading 3 months may be perfectly funded.

• Raises that never file

  Form D is required for most Regulation D offerings, but filings lag, some issuers rely on exemptions that do not produce one, and a 506(b) round may surface weeks late. An absent Form D is not evidence of an absent raise.

• Trials that never register

  Not every study gets an NCT number. Preclinical and non-US work is largely outside ClinicalTrials.gov, so a company can be further along in the clinic than its stage says.

• Burn is a model, not a disclosure

  It reads stage and therapeutic area, and nothing else. A twelve-person company and a sixty-person company at the same stage get the same modeled burn. Discount accordingly.

• Resolution errors, in both directions

  A wrong merge attaches a trial to a company that did not run it. A missed merge splits one company into two thin records. Both happen. The evidence panel on each dossier is there so you can catch us.

• Momentum measures disclosure, not quality

  It counts filing activity. A company can register a trial that fails, file a patent that never matters, and score well doing it. Momentum tells you where to look, and it makes no claim about whether the science works.

• Source lag

  NIH RePORTER and USPTO refresh weekly, so a fresh award can be up to a week old before it reaches the graph. The event carries the date the filer reported, not the date we read it, which is the right choice for scoring and does mean a score can move retroactively.

• US federal sources only

  These six datasets describe US-funded, US-registered, US-filed activity. A European company with no US grant, no US trial registration and no Form D is not in this graph at all.