Research

What the evidence can and cannot support

One question, one event store, and a record of what the store failed to support. Every number below is in-sample and carries its artefact path.

package v0.5.0dataset 2026-10-31375,378 eventsno out-of-sample claim

The question

Can publicly observable evidence give reliable early warning of meaningful change in medical-technology development?

source: README.md; research/MISSION.md

Method

The method is fixed before the result is known.

Three commitments make a number quotable, and a fourth makes it honest: the limitation it inherits.

Frozen artefacts

Experiments are immutable inputs. A rerun that changes a number produces a new artefact; the old one stays and is cited by path, never edited.

experiments/reverse_rd_v1/ · experiments/reverse_rd_v2/

Named baselines

Every model is scored against a stated base rate — for the market-to-R&D continuation target, a per-pair base rate of 0.7913 — never against an unnamed default.

docs/METHODS.md · docs/SCHEMA.md

Holm-corrected block bootstrap

Direction-aligned abnormal returns are bootstrapped in blocks; Holm correction runs over the full category × horizon family, so no horizon can be quoted without its correction.

medradar/stats.py · docs/METHODS.md

The limitation the method cannot remove

The snapshot has no independent public time: published_at = available_at = event_date on 375,378 / 375,378 rows, and timestamp_quality = 'date'. Return studies therefore rest on a same-day-public-time assumption and are not point-in-time results.

source: docs/DATA.md → 时间戳质量; README.md → Limitations 1–2

Result 01 · event study

Only the 60-day horizon survives.

Direction-aligned abnormal returns over the simulated event study, block bootstrap, Holm-corrected over the full category × horizon family.

Abnormal returns by horizon — in-sample simulation
HorizonnMean abnormal returnt95 % CI
+1 d122+0.15 %0.57[−0.33 %, +0.65 %]
+5 d109−0.37 %−0.92[−1.17 %, +0.42 %]
+20 d89+1.14 %1.17[−0.65 %, +3.13 %]
+60 d61+5.65 %2.45[+1.35 %, +10.32 %]

Read this with the table

In-sample simulation, long-only, 38 simulated trades over about eleven months. No independent public time in the snapshot, so this is not a point-in-time result and not investment advice.

source: docs/metrics.json → trading.event_study; README table

metrics.json (machine-readable)

Event-study figure: mean direction-aligned abnormal return by horizon at +1, +5, +20 and +60 days with 95 percent confidence intervals; only the +60-day interval excludes zero. In-sample simulation, package v0.5.0, dataset snapshot 2026-10-31.
Committed snapshot — dataset 2026-10-31, package v0.5.0. Plot of the table above; it is not a live series and not a forecast.

source: docs/assets/figures/event-study-en.svg; docs/metrics.json → trading.event_study

Result 02 · negative results

The negative results are the deliverable.

Market-to-R&D lane, 12-month horizon. Frozen artefacts: experiments/reverse_rd_v1/ and experiments/reverse_rd_v2/.

Market-to-R&D lane — what was tested and what it measured
TestedMeasurementVerdict
Does any added source produce a detectable ranking gain? Marginal PR-AUC on continuation: financial −0.0013, market −0.0015, publication −0.0001, all sources −0.0068 No detectable gain
Does the model beat a per-pair base rate? AP 0.7935 (model) vs 0.7913 (base rate), Δ +0.0022, 95 % CI [−0.032, +0.037] Barely, and not beyond the interval
Does more data improve ordering? ECE 0.0811 → 0.0628, FPR 0.2753 → 0.2472 Calibration only, not ordering
Is the entry target stable across windows? Differences sit inside the 0.014–0.026 window standard deviation Unstable, so not claimed
Patents USPTO Open Data Portal requires an API key Not evaluated
Procurement Snapshot is not point-in-time usable Excluded

source: README.md → The negative results; experiments/reverse_rd_v1/, experiments/reverse_rd_v2/; docs/REVERSE_RD.md

Can the missing public timestamp be recovered?

The E1b source-recovery experiment tested each source for a retrospective public instant. This is a shape test — 28 requests, 5 records per source — not a population estimate.

E1b source recovery — public time by source
SourceRetrospective public timeVerdict
SEC EDGAR Yes — acceptanceDateTime, immutable archive, amendments arrive as new accessions STRONG
ClinicalTrials.gov Partial — first-post date only, date granularity, ACTUAL records only; historical content state is not recoverable MEDIUM
openFDA 510(k) / PMA No — no public-posting field exists; decision_date is the event date, not the publication date Prospective observation only

source: docs/refoundation/phase2/E1B_SOURCE_RECOVERY.md; README.md

Result 03 · early signal

Five firms, and a lead-time gate that did not pass.

5 firmsPhase 0 · CONDITIONAL-GO

Early detection of corporate technology shifts

Phase 0 is CONDITIONAL-GO: the deep-case study can continue, but only under the stated conditions (sustained positive, earlier source, human adjudication).

lead-time gatePhase 1 · NO-GO

The change was not shown to be detectable earlier than the market

The gate that matters is the lead time. It did not pass, so the study is recorded as a NO-GO rather than reported as an early-warning capability.

source: research/studies/early_signal_5firms/PHASE0_REPORT.md; PROJECT_STATE.yaml → research_studies; docs/RESEARCH_INDEX.md

Limitations

Nine limitations, none of them decorative.

These change how every number on this page may be read.

  1. Event time only

    published_at = available_at = event_date for 375,378 / 375,378 rows; independent public time = 0 rows.

  2. No genuine point-in-time claim

    Unsupported until source publication timestamps are re-acquired.

  3. Partial entity resolution

    35.2 % company key, 52.4 % track; unmatched events are counted, never guessed.

  4. The revision ledger is not tamper-proofing

    It records the application write path; a database owner can still replace the file.

  5. 510(k)/PMA rows carry no public timestamp

    They enter industry counts and never return studies — a deliberate loss of sample.

  6. The AdCom collector parses zero rows

    The calendar is rendered client-side; it returns empty / source_changed, never a silent ok.

  7. The simulated book is in-sample and long-only

    38 trades; shorts are alerted but never traded; fees, squeeze-outs and delisting are unmodelled.

  8. No causal claims

    Difference-in-differences and synthetic control are roadmap items, not results.

  9. NMPA is deliberately not scraped

    Without a user-supplied CSV every China value is null, never zero.

source: README.md → Limitations; docs/ASSUMPTIONS.md; docs/METHODS.md