A.8.1 Data Sources from Zero

Basic Prerequisites: Financial Statements & Accounting (the three statements). No prior data or programming knowledge needed.

---

1. Intuition & Practical Objective

This page builds the why of fundamental data with no prior knowledge needed. The objective is one idea: a fundamental number is not a fact of nature - it is a claim made by a company, on a date, in a document, under a set of accounting rules; and “where the numbers come from” is a question with a real, inspectable answer.

Start with the dumbest question: where does “FY2022 revenue = 1000” actually live? It is not in a database in the sky. It is a number typed into a 10-K filed with the SEC, tagged with an XBRL element name, stored in EDGAR, and then copied - with each provider’s own adjustments - into Sharadar, Compustat, Bloomberg, Yahoo. Every one of those copies can differ, and the original is always the filing.

Three steps:

1. There is a chain, and it starts at the regulator. Company types numbers into a filing → files it with the SEC → EDGAR stores it (raw HTML plus XBRL-tagged facts) → commercial vendors standardize and resell it. Going upstream buys you fidelity and costs you effort; going downstream buys you convenience and costs you control. (The Corpus’s data-sources-and-corporate-data section is explicit: EDGAR is the club’s primary raw input.)

2. The chain is a four-layer stack, and the layers are not interchangeable.

   Layer	What it is	Example	Who owns the risk
   L0 raw	The filing as published	10-K HTML + XBRL companyfacts JSON	nobody standardizes it for you
   L1 canonical	Tagged concepts mapped to a fixed schema	us-gaap:NetIncomeLoss → net_income	tag ambiguity - see 02 · SEC EDGAR & XBRL
   L2 panel	One row per (firm, period, as-of date)	a PIT table that is never overwritten	look-ahead if you skip the as-of date
   L3 signal	Margins, ROE, growth, screen ranks	ROE = NI / avg equity	bias propagation from every layer below

   The whole discipline of this folder is: never compute L3 on anything but an honest L2. Most people compute L3 directly off an L0/L1 convenience download and never notice the leak.

3. “As reported” is not “as restated,” and the difference is the future. A company’s 2020 revenue will be re-filed in a later 10-K/10-K/A if it is reclassified. If your database shows the restated 2020 figure while you are standing in 2021, your screen is reading a number that did not exist yet - the classic look-ahead leak, and the subject of 05 · Failure Modes.

---

2. Mathematical Ground Truth & Derivations

The provenance tuple. Every fundamental observation is

$$\text{obs}=(\text{entity},\text{concept},[\text{period start},\text{period end}],\text{value},\text{unit},t_{\text{filed}}),$$

and the operational rule is that the observation is usable only from $t_{\text{filed}}$:

$$\text{value is knowable at time }t⟺t_{\text{filed}}\le t.$$

A “point-in-time panel” is one that keeps every vintage of (concept, period) separately, keyed by $t_{\text{filed}}$, instead of overwriting on restatement:

$$\mathcal{P}=\{(\text{firm},\text{concept},\text{period},t_{\text{filed}})\mapsto \text{value}\}.$$

Overwriting on restatement collapses $\mathcal{P}$ to one row per period and destroys the only information that tells you whether a backtest was honest.

The flow/stock distinction that trips up extraction. Income-statement and cash-flow facts are duration facts (they carry a start and an end); balance-sheet facts are instant facts (they carry only an end). Mapping both into one schema without recording which is which is the most common beginner extraction bug:

$$\text{duration: }[\text{start},\text{end}]\text{vs.}\text{instant: }\{\text{end}\}.$$

The look-ahead arithmetic, exactly. Let $R_y$ be revenue for year $y$ as first reported and ${\tilde{R}}_y$ its eventually-restated value. A restatement of year $y$ changes the measured growth of $y+1$ by

$$\Delta \text{growth}=\frac{R_{y+1}}{{\tilde{R}}_y}-\frac{R_{y+1}}{R_y}\ne 0,$$

so even a single restated denominator silently rewrites every growth number that depends on it - which is why the leak is so easy to miss and so corrosive to a screen (§3 measures it).

---

3. Computational Implementation - the stack, and one measured look-ahead leak

Stdlib only. It prints the four-layer stack and then quantifies a single 2020 restatement’s effect on the 2021 growth figure a screen would have computed at the time.

The lesson is not that 5.93 points is enormous - it is that the same firm, the same two years produced two materially different growth readings depending only on which vintage of the database was read. Multiply that across a screen of hundreds of names and it is the difference between a real edge and a data artifact.

---

4. Failure Modes & First-Principles Breakdowns

1. Provenance blindness (first principle: a number is a claim, not a fact). “Revenue is 1000” is incomplete without who tagged it, when, and whether it was later revised. Always carry source and as-of alongside the value. Fully developed in 05 · Failure Modes.
2. Convenience-first extraction (first principle: the downstream copy encodes choices). Starting from a Yahoo quoteSummary scrape means inheriting an unknown standardization and no as-of date - fine for eyeballing a P/E, fatal for a backtest. Start from EDGAR when fidelity matters (02).
3. Duration/instant confusion (first principle: flows span a period, stocks are a point). Storing a balance-sheet “instant” in the same column as a P&L “duration” makes every average-denominator ratio wrong. Tag the fact type at L1.
4. Overwriting on restatement (first principle: $\mathcal{P}$ must keep vintages). A panel that keeps only the latest value cannot reconstruct the past - the exact defect that makes naive Compustat backtests look better than they were (03).

---

5. References

• SEC, EDGAR APIs documentation
• XBRL US / XBRL International taxonomies
• Ittelson, Financial Statements: A Step-by-Step Guide
• Penman, Financial Statement Analysis and Security Valuation
• Sloan, Richard (TAR, 1996)

---

6. Connected Graph Bridges

• Base: Financial Statements & Accounting · Core Financial Ratios
• Continue: 02 · SEC EDGAR & XBRL · Index Hub
• Theory: Accounting Quality & Red Flags (why restatements happen)