How do professionals sanity-check data?

Final lesson of the course — and of the entire Foundations core. It assembles this unit into a habit set that data professionals (ourselves included — this platform's business IS data) run on reflex.

The five-point sanity check

Before trusting any number enough to reason from it:

1. Timestamp — as of when? Market open? Feed delayed? (Lesson 2 — the where/when.)
2. Adjustment — raw or adjusted? Split/dividend cliffs lurking? (Lesson 1.)
3. Sample — who's missing? Survivors only? Window cherry-picked? (Lesson 3.)
4. Source agreement — does a second independent source roughly concur? Two sources agreeing within rounding is confidence; two sources disagreeing wildly is a finding in itself (usually: different venue, different adjustment, different timestamp — now you know all three suspects).
5. Plausibility — does the number pass the smell test? A stock "up 4,000%" today is a decimal error, a split artifact or a micro-cap sneeze until proven otherwise. Extraordinary numbers demand ordinary explanations first.

Thirty seconds, five questions — and the vast majority of data errors, artifacts and manipulations die before touching your reasoning. The Quant & Coding path will automate pieces of this checklist; the reflex, though, has to live in you.

Run it once, here

A checklist you have only read is a checklist you do not have. Here is a live quote, row by row, and the five points against it:

Timestamp. The update time is a count of seconds in UTC, so the card can tell you exactly when. Convert it and subtract from now. Fifteen minutes is the delay you were promised; nineteen hours is a market that shut yesterday, and the two look identical until you do the subtraction.

Adjustment. Look for an adjusted price and find that the card has none: a real-time quote is the raw tape, which is what a quote should be. It becomes a problem the moment you compare this number against a price from before a split, and that comparison is the mistake the adjustment lesson exists to prevent.

Sample. One row. The question does not apply, and knowing why it does not is the point: there is no selection to be biased, because nothing was selected. The moment you pull a list instead of a quote, it applies again.

Source agreement. You do not need a second source to start — the card disagrees with itself if anything is wrong. The latest price minus the previous close must equal the change in dollars. Then divide that difference by the previous close and multiply by 100, because the change row is already a percentage: on a recent Apple quote, −0.17 over 325.13 is −0.000523 as a ratio and −0.0523 as the percentage row reports it. Getting that factor of a hundred backwards makes a perfectly good quote look broken, which is its own lesson about checking a convention before checking a number.

Plausibility. Read the percentage change and ask whether a company of this size moves that far in a day. Single digits are a Tuesday. Anything past twenty demands a reason, and the reason is usually a decimal, a split or a ticker you did not mean to fetch.

In the data

Point four of the checklist has help in the data: a record of ticker changes. Here is one you will recognise:

A price history for META that appears to begin abruptly in June 2022 is the same company under its previous ticker, FB, and two sources that "disagree" about Meta's past are often one that followed the rename and one that did not. The other common cause is the same company listed in more than one place; a company profile names its other listings, and two sources quoting different exchanges will disagree honestly, in different currencies and at different hours.

Try it now

1. Now run the five points on a number nobody has pre-chewed for you: Apple's delayed quote above, at the date beside its title. To run it on your own anchor company instead, open it in the Terminal and change the symbol there. Work down the list on what you see — first dropping any row that reads "not available", because on a shut or thinly quoted instrument most of them do: convert the timestamp and say how old it is in minutes; name what is missing where an adjusted price would be; say why sample does not apply; check the latest price minus the previous close against the change in dollars, and the ratio times a hundred against the percentage change; then read the percentage change and decide whether it is a Tuesday.

Open AAPL.US in the EODHD Terminal

2. Write down which point failed, or came closest to failing, and why. On a liquid US name in market hours the honest answer is usually "none, and the timestamp was eleven minutes old" — which is worth knowing as the baseline, because it is what a clean answer looks like.
3. Then run it again after the close, or on a company listed somewhere else, and watch which point moves first. Knowing which of the five is load-bearing for the number in front of you is the whole exercise; a checklist that always passes has told you nothing about where it was working.

Under a minute for all five? Then the reflex is installed, which is what this course existed for. The Foundations core is complete — three courses and one architecture: what things are, how the machine runs, and how to observe it honestly. Everything ahead speaks that language.

The exam is next, and then the paths from Course 1 stand open for real: Investor → Fundamental Analysis · Trader → Technical Analysis · Builder → Quant & Coding, your first market tool with no code experience needed · or stay in the domain with Macro for Markets.