What we found building the data
Building point-in-time data from raw SEC filings means running into every way the raw data can mislead. These are the problems we measured, with the numbers, so you can check whether they affect your own work.
How much look-ahead bias is in your fundamentals backtest?
Dating fundamentals by quarter end puts 13%–24% of companies in the wrong quintile, and a fixed reporting lag still leaks. Measured on 5,300 companies, 2012–2023.
Read →When the CIK changes and the company doesn't
136 companies have filed under 277 CIKs. Keyed on CIK, Alphabet had one quarter of history at the end of 2015, and every reorganization looks like a delisting.
Read →About 1 in 10 revenue figures is later restated
71% of revenue restatements go down, a median of 16 months after the original. Most datasets keep only the new number.
Read →SEC's bulk filing history has timestamps off by four to five hours for a third of companies
Enough to move half of all earnings releases into the wrong market session. How we found it, corrected it, and why unverifiable times are published as dates only.
Read →Three years after the rule changed, some managers still report 13F values in thousands
Summed as reported, 13F holdings reach $1,255 trillion in one quarter. The checks that bring it back to $82.2 trillion.
Read →Reproduce it
Every figure in these articles comes from the released Varve Data tables. The example notebooks run on the free sample, and the methodology guide documents how each table is built.