One account. Every library.
Your free ElkassabgiData key unlocks the whole family: IP Data Library (this site) and HF Data Library — 1-minute intraday U.S. equity data — and Econ Data Library — economic & financial time series.
Living errata. Every discovered issue documented with the date it was found.
Transparency policy. Every data issue we discover is documented here with the date it was found and what was done about it. If you find one, tell us — reporters are credited.
Status: Upstream defect, documented 2026-08-23 — handled, not corrected.
19 of the 9,454,161 patents in patent_measures.parquet (0.0002%) carry a
filing_year that cannot be real: values such as 2982, 7978, 9177, 9183 and 9186.
Each has a perfectly sensible grant_year beside it — patent 4542062 reads
filing 2982 against grant 1985, and 4135654 reads filing 9177 against grant 1979 — so
these are leading-digit typos in the underlying USPTO record (1982, 1977), not a fault in our
processing.
We do not silently repair them, because guessing the intended digit would be fabricating
data. Instead the pipeline refuses to derive anything from them: grant_lag_days is
NULL for all 19 rather than a nonsense value. If your analysis uses
filing_year, filter to a plausible range; if it uses grant_lag_days,
the bad rows are already excluded for you.
Status: Inherent to citation data, documented 2026-08-23.
A patent can only be cited by patents granted after it, so recent patents look uncited whatever their quality. This is not a gap in our coverage — it is the shape of the data, and it is large. Measured on the served bundle, the share of assignee-years with zero forward citations rises monotonically with grant year:
| Grant year | Assignee-years with zero forward citations | Mean forward citations |
|---|---|---|
| 1980 | 2.6% | 66.5 |
| 2000 | 2.4% | 170.9 |
| 2010 | 9.6% | 104.0 |
| 2015 | 19.4% | 63.7 |
| 2020 | 36.1% | 25.4 |
| 2023 | 56.2% | 6.5 |
| 2025 | 87.9% | 0.6 |
Use the windowed counts (fwd_cites_5yr, fwd_cites_10yr), which carry
explicit truncation flags wherever the window runs past the data’s last grant date, and do
not compare raw citation counts across cohorts without a correction.
Status: Methodological difference, documented at launch.
Both measures are computed over CPC subclasses (~679 classes), each patent assigned its
primary class. The original Hall–Jaffe–Trajtenberg variables used roughly 400 USPC
main classes. Finer classes mechanically raise both scores, so levels from this library are not
directly comparable to the NBER files or to papers built on them. Both columns are also
NULL where a patent has no citations to compute over, rather than being filled
with a zero that would read as a real measurement.
Status: Scope limitation, documented at launch.
These measures are computed from PatentsView bulk tables, which are a processed view of the USPTO record rather than the record itself. Assignee disambiguation is PatentsView’s, and inherits its error rate; we do not re-disambiguate. This library is not the official USPTO record and no endorsement is implied.
This page is updated as issues are discovered. Please report anything that looks wrong — a ticker, a patent id, a date, and what you expected.
One account. Every library.
Your free ElkassabgiData key unlocks the whole family: IP Data Library (this site) and HF Data Library — 1-minute intraday U.S. equity data — and Econ Data Library — economic & financial time series.