Project 4 Signal Preparation Data Dictionary

Field names match configs/signal_contract.json and signal_layer.py. All data here are constructed (frame.constructed = true, has_real_company_data = false). The real-source column describes inputs needed for a complete Project 4, not data already obtained.

A. Input tables

A1. `versions`: filing versions

Primary key: security_id + period_end + available_at. An original filing and its restatement for the same fiscal year occupy separate rows.

Field	Meaning	Unit	Time role	Missingness/convention	Real source needed
`security_id`	Security/company identifier	String	—	Required	SEC `cik` mapped to a security master; missing
`period_end`	Fiscal year-end/reporting period	Date	event_time	Required; not availability time	SEC `period`
`available_at`	Public availability of this version	Zoned UTC timestamp	available_at	Fiscal year-end 16:05 UTC +60–110 days; restatements 130–260 days	SEC `accepted`; time zone still to verify in `source_feasibility.md`
`ingested_at`	Entry into this database	UTC timestamp	ingested_at	Constructed as `available_at` +1 day	Actual download/ingestion timestamp
`form`	`10-K` or `10-K/A`	String	—	Keep both original and amendment	SEC `form`, `prevrpt`
`revenue`, `cogs`, `sga`, `interest`	Revenue, cost of goods sold, selling/general/administrative expense, interest expense	Reporting currency, common scale	Same version	Banks have NaN `cogs` and `ebitda`	XBRL tag mapping still to verify
`net_income`, `cfo`	Net income, operating cash flow	Same	Same	Same	Same
`assets`, `assets_prev`, `equity`, `equity_prev`, `debt`, `cash`, `ebitda`	Current/prior year-end assets/equity, debt, cash, EBITDA	Same	Same	About 3% of company-years have near-zero/negative equity; prior comparatives are stored in the same filing	Same

Every numeric value in a selected row must come from the same filing. latest_available_values selects the complete row together.

A2. `returns`, `market_cap`, `listed`

Primary key: month × security_id.

Field	Meaning	Unit	Time role	Missingness/convention
`returns`	Monthly simple total return	Decimal	Return month; event_time within month, no available_at field	NaN outside listed life; delisting-month−0.25 is constructed, followed by NaN
`market_cap`	Month-end capitalization	Currency	Month-end	NaN outside listed life
`listed`	As-of month-end universe membership	Boolean	Known at `signal_time`	Include later delistings; never backfill with today's membership

A3. `industry`

Primary key: security_id. Six nonfinancial industries plus Bank, static in the constructed universe. Real applications need the version known at signal time. SEC sic is the value at filing.

B. Signal-layer outputs

Each output is a month × security panel. Row $t$ uses information available by $t$ only. Every raw field is NaN for unlisted names.

Output	Meaning	Direction	Missingness
`book_to_market`	Equity / market cap	Larger is cheaper	Nonpositive equity/cap → NaN
`earnings_yield`	Net income / market cap	Larger is cheaper; keep losses negative	Nonpositive cap → NaN
`ev_to_ebitda_cheapness`	Negative same-date percentile of EV/EBITDA; EV=cap+debt−cash	Larger is cheaper	Nonpositive EV or EBITDA → NaN; banks NaN
`mom_12_1`	$\prod(1+r)-1$ over months $s-11..s-1$	Larger is stronger	Any missing window month → NaN
`gp_a`	(Revenue−COGS) / average assets	Larger is better	Banks NaN
`neg_accrual`	−(Net income−operating cash flow) / average assets	Larger is better	Any missing input → NaN
`roe_avg`	Net income / average equity	Larger is better	Both year-end equity values must be positive; average equity must exceed 5% of average assets
Processed $z$	Same-date winsorization → industry/log-cap residual → z-score	—	Too few valid names → entire row NaN
Style/composite scores	Equal weights within styles, available-term renormalization and minimum 1 field; across styles minimum 2 styles	Larger is more bullish	Fewer than 2 styles → NaN, no zero filling
Quintile long–short weights	Top $+1/m$ , bottom $-1/m$ , gross 2; optional listed-support industry projection with gross restoration	—	Listed count below 60, no long–short direction, or degenerate projection → cash; listed names without alpha may hedge, off-support weights zero

The EV bridge assumes minority interest, preferred stock, and nonoperating assets are zero; the universe lacks these fields. Real use must complete Chapter 24's bridge and reconcile it.

C. Time fields and rules

available_at ≤ signal_time < trade_time; signal time is the last instant of month-end UTC. A version is usable in month $t$ only if both available_at and ingested_at strictly precede the first instant of the next month. Future returns are labels only and enter no signals, preprocessing, or parameter selection.

D. Current gates and output identity

The listed-name gate 60, winsorization count 30, z-score count 10, and within-style 1/across-style 2 coverage gates are separate. OLS uses the same valid pairs for signal, all controls, and group, with at least 2 residual degrees of freedom. Combination availability counts only nonzero-weight terms. Industry-projected holdings may include no-alpha hedges and cannot be described as pure extreme-quintile holdings. A5 full-column gross 2.11 is historical; current C13 support targets restore gross 2 and report execution residuals separately.