Data Collection

This slice rebuilds the Gabaix–Koijen–Richmond–Yogo asset-embedding universe from free public data: I parse the public SEC 13F filings directly and link them to CRSP.

Identity resolution. 13F reports CUSIP natively, so I link CUSIP-8 → CRSP PERMNO through a four-tier point-in-time cascade — a near 1:1, unambiguous map (zero ambiguous keys across the panel).

Modeling unit. I aggregate holdings to PERMCO (firm level). That’s the correct unit because the valuation benchmark — Compustat book equity — is company-level and can’t be split across share classes. I keep PERMNO underneath, since the trading stage needs per-security returns, which keeps the aggregation reversible.

Universe (following GKRY §5.1): US common stock only, drop micro/nano caps (below the NYSE 20th-percentile market equity), and require ≥20 investors per token and ≥20 tokens per investor, iterated to a fixed point. The panel spans 12 quarters, 2014Q1–2016Q4.

Known limitation. 13F is filed at the manager level, so a filer’s separate funds get lumped into a single, sometimes index-like portfolio. The fix is fund-level holdings via CRSP MFDB — but I’d rather finish the whole vertical slice first, since at this stage we can’t know whether that granularity actually moves the strategy’s P&L.

Code: code/asemb/data_in.