Financial data are recorded observations and information about financial instruments, transactions, issuers and the economic conditions in which they operate. A share price, an earnings filing, a bond's contract terms and a dividend announcement all qualify. Their usefulness to a quant depends on what they measure, which instrument they describe and when the information could actually have been used.
A model can fit a perfectly tidy table and still answer the wrong financial question. Mixing currencies, confusing a quote with a trade, or attaching a later filing to an earlier decision changes the meaning of the input before estimation begins.
Different objects, different questions
Financial data is an umbrella term; market data is one part of it.
| Data object | Examples | What must travel with it? |
|---|---|---|
| Market observations | Trades, bid and ask quotes, closing prices, trading volume | Venue, field definition, timestamp and quote units; a quote is not an executed trade |
| Security reference data | Instrument identifiers, currency, share class, bond maturity | The exact instrument and the period for which each attribute applies |
| Issuer disclosures and fundamentals | A filing, reported revenue, balance-sheet liabilities | Reporting period, accounting concept, units, filing source and publication/version history |
| Corporate actions | Dividend announcements, splits, redemptions and mergers | Event terms, relevant dates, security identifiers and subsequent corrections |
| Derived measures | Returns, volatility estimates, analyst sentiment scores | Raw inputs, transformation method, estimation window and version |
Economic data, such as inflation releases, describe the wider environment. Alternative data, such as a suitably sourced store-visit measure, may provide another view of economic activity. A measurement of visits does not become revenue simply because the two might be related.
These categories overlap operationally. A dividend event may appear in a corporate-actions feed and also be used to construct a return series. Keeping the event and the resulting calculation separately traceable makes it possible to explain the return. DTCC's corporate-actions and reference-data services illustrate why event announcements and security records are distinct parts of the same infrastructure.
What does “300 observations” mean?
Suppose a fictional dataset records one closing price for each of 100 instruments on three selected dates. Prices are USD per share and there is exactly one row per instrument–date pair. The complete table has 300 rows: 100 instruments multiplied by three dates.
Original synthetic USD closing-price panel. Filled squares represent recorded closes; middle instruments are abbreviated. A row count does not establish independence.
This is panel data: the same instruments recur over time. Each instrument supplies a short time series, and each date supplies a cross-section. It is not evidence of 300 independent observations. Instruments can react to a common market shock, and observations of the same instrument can depend on its past. The usable amount of statistical information depends on that dependence and the question being estimated.
Nor is a total of 300 rows sufficient to establish completeness. If an import repeats instrument 001 on date A while omitting instrument 100 on date C, it still has 300 rows but only 299 distinct instrument–date keys. Checking both uniqueness and coverage finds the problem. In an actual market dataset, the expected coverage must also allow for listings, suspensions and the chosen trading calendar; every absent value is not automatically an import error.
Read the data contract before the values
Identity and coverage. Match the security, share class, contract and venue required by the research question. A company identifier does not by itself identify all its traded instruments. OpenFIGI's allocation rules distinguish instrument and venue levels and explain which changes preserve an identifier. Retain historical mappings and delisted instruments where the intended universe requires them; selecting only today's survivors can create survivorship bias.
Units and adjustments. Record currency and scale, whether a price is per share or another quotation basis, and whether a rate is stored as a decimal or a percentage. Preserve the distinction between raw prices, split-adjusted prices and a total-return series. “Adjusted close” needs the provider's definition: the label alone does not specify dividend treatment. A return calculation must use inputs on a consistent basis.
Time and versions. Separate the time an event happened or a reporting period ended from the time the information was released, received and ready for the strategy. Preserve revisions rather than silently replacing the record used in an earlier decision. Otherwise a historical table can introduce look-ahead bias. A date column alone is insufficient; time zone, timestamp meaning and version selection matter.
The SEC's EDGAR API documentation gives a concrete disclosure example: company-concept facts are separated by unit, while calendar-frame results can contain facts with differing reporting start and end dates. Selecting a convenient frame does not remove the need to inspect periods or reconstruct historical availability.
Source and permitted use. Keep the original provider, retrieval time and transformation history with the data. Check the applicable data licence for the intended storage, computation, display and redistribution. NYSE's connectivity documents explicitly distinguish receiving, using and redistributing its market data. Being able to download a file does not establish all of those permissions.
For the implementation step, the Time-Series Data lesson develops timestamp parsing, duplicate checks, time zones, alignment and resampling. It has a public preview; the full lesson requires membership.
Sources
- Marcos López de Prado, Advances in Financial Machine Learning (Wiley, 2018), Chapter 2, §2.2, “Essential Types of Financial Data,” especially §2.2.1 on release dates and revisions. The categories above are a practical grouping for this entry, rather than a reproduction of the book's taxonomy.
- SEC: EDGAR Application Programming Interfaces: filing data, units and reporting-period differences.
- OpenFIGI: Allocation Rules: identity across instruments, venues and corporate events.
- DTCC: Corporate Actions & Reference Data: event announcements and security-reference services.
- NYSE: Connectivity Documents: documentation governing receipt, use and redistribution of exchange data.