Glossary

The finding that topics the media cover most are the topics the public comes to rate as most important: media shape what people think about, not what they think.

A record store that only ever adds rows: corrections are new rows that point at the ones they replace, so history, including the mistake, stays readable.

A delivery guarantee under which every message or job arrives one or more times, never zero — so duplicates are possible and must be tolerated.

Bias that arises when units drop out of a study for reasons related to what is measured, so the units that remain no longer represent the original group.

Loading or recomputing data for past periods after the fact — and recording when each value was really loaded, so history is not made to look earlier-known than it was.

Evaluating a rule or model by replaying it on historical data as if it had been used at the time. Informative only if the replay is honest.

How common something is in the whole relevant population, before looking at any specific evidence.

A simple method, such as repeating the last value, against which a new method is compared. A result means something only relative to it.

Data that records two times for every fact: when it was true in the world, and when the system learned it.

A ranking function that scores documents for a query from the query words they contain, with a cap on the credit for repetition and an adjustment for length.

Estimating how much a statistic would vary from sample to sample by resampling the observed data, with replacement, many times.

The average squared difference between forecast probabilities and what happened. Lower is better; always saying 50% scores 0.25.

Events forecast at p% happen about p% of the time. Calibration is separate from accuracy: an overconfident forecaster is not calibrated.

A guard that stops calling a failing service after repeated errors, waits, then tests carefully before resuming.

How much two raters agree beyond what chance alone would produce: 1 is perfect, 0 is no better than chance.

Grouping items or people by a shared starting point, such as the month they first appeared, and following each group over time.

Storing a table column by column instead of row by row, so analytical queries that need a few columns read far less data.

A change over time in the relationship between inputs and the outcome, so that a model or rule that once worked slowly stops working.

A range calculated from a sample that expresses how precisely a quantity has been estimated. A 95% interval comes from a procedure that captures the true value in 95% of repetitions.

Looking for, noticing and remembering evidence that supports what we already believe, and discounting what contradicts it.

A third factor that influences both the supposed cause and the outcome, creating or hiding an association between them.

A method that turns texts or images into countable categories using explicit coding rules, so that claims about content can be checked and repeated.

Storing each object under a name derived from a hash of its bytes, so identical content has one address and any change creates a new one.

Correlation means two quantities move together; causation means changing one would change the other. Data alone cannot tell which, or whether a third factor drives both.

Confirming a claim with evidence independent of the original source. Reports that trace back to the same origin count as one.

How closely two vectors point in the same direction, as the cosine of the angle between them. It ignores length.

Estimating how well a method will do on new data by repeatedly leaving part of the data out, training on the rest, and rotating.

A function that turns any input into a short fixed-size fingerprint: the same bytes always give the same fingerprint, and any change gives an unrelated one.

A written, versioned agreement between the producer of a dataset and the people who use it, covering structure, meaning, timing and quality.

Information that would not exist at prediction time finds its way into the data used to build or evaluate a model, so results look better than they can ever be in use.

The terms that say what you may do with a dataset or feed: keep a copy, compute from it, publish results, pass it on. Access alone grants none of these.

The recorded path of a piece of data: where it came from, how it was changed, and what depends on it.

Named aspects — accuracy, completeness, consistency, timeliness, uniqueness — that turn the vague idea of clean data into checks you can measure.

A complete record of a list or table as it stood at one moment, kept as a unit with its time, scope and completeness, so later versions can be compared against it.

A short, dated note made at the time of a decision: options, evidence, assumptions, the choice and what would change it.

Changing a price frequently, often automatically, in response to demand, stock, time or competitors. One observed price is only a snapshot.

How big a difference or relationship is in practical terms, as opposed to whether it can be detected statistically.

Deciding which records, in one source or many, refer to the same real-world thing, and linking them under one identifier.

Two orders for moving data into an analysis store: transform it first and then load it (ETL), or load it raw and transform it afterwards (ELT).

Finding, in a stream of documents, the moments when something new has happened, usually as a burst of stories about the same thing.

Storing every change as an immutable event and working out the current state by replaying the events.

Returning, with every score, the reasons behind it: which rules fired, what each criterion contributed and what data were missing.

Waiting longer after each failed attempt, usually doubling the delay and adding randomness, so retries relieve a struggling service instead of piling on.

Turning raw data into inputs a model or rule can use, such as ratios, rates of change and counts over windows. Each must be computable from what was known at the time.

Floating-point numbers are fast but cannot hold many decimal fractions exactly; money and other exact quantities are better stored as whole numbers of a small unit.

Choosing which aspects of a situation to emphasise so as to suggest a particular problem, cause or remedy. The same facts can be framed in different ways.

A real web browser run without a visible window and controlled by code; it runs the page's scripts, so it can read pages assembled in the browser.

A measure of market concentration: the sum of the squared market shares, in percent, of all sellers. It rises sharply when a few sellers dominate.

A ranking of kinds of evidence by how well their design guards against error: reviews and randomised experiments above observational studies, anecdotes at the bottom.

Once an outcome is known, believing it was predictable and that you knew it all along. It makes the past look more certain than it was.

Data set aside before any building begins and used once, at the end, to estimate performance on unseen data.

An operation is idempotent when doing it several times has the same effect as doing it once — which is what makes retries safe.

A table mapping each word to the documents that contain it, so a search engine can find matches without scanning every document.

Records that reach a pipeline after the period they describe has already been processed or reported. They are normal, so a pipeline needs a rule for them.

The many niche items that each sell little but together make up a large share of sales, typical of markets where shelf space costs almost nothing.

A study error where information that was not yet available at the time of a decision is used to evaluate that decision.

The average, over many queries, of one divided by the position of the first correct result. It rewards finding one right answer early.

The mean is the sum divided by the count; the median is the middle value. Extreme values pull the mean but not the median.

How many times a topic, brand or person is mentioned in a defined set of sources over a period. Its meaning depends entirely on how the count is defined.

Values that should exist but were not recorded. Why they are missing decides whether ignoring them is harmless or biases the result.

Replacing each point of a series with the average of the points around it, to smooth short-term noise. A longer window smooths more but reacts later.

The more tests, rules or variants you try on the same data, the more likely some look significant by chance. The number of tries has to be counted.

Finding mentions of real-world things in text, such as people, organisations and places, and labelling their type. Linking each mention to one identifier is a separate step.

Finding documents that are almost, but not exactly, the same, such as a news story republished with a new headline, by comparing fingerprints of their word sequences.

A score for a ranked list that adds the graded relevance of results with lower positions discounted, divided by the best score possible. It runs from 0 to 1.

John Boyd's model of decision-making as a repeating cycle of observing, orienting, deciding and acting. Orientation, how we interpret what we see, is where bias enters.

Judging a decision by how it turned out instead of by how well it was made with what was known at the time.

An observation far from the rest. It may be an error, a rare but real event, or a sign of two kinds of things mixed together; look before deleting.

When a model learns the noise in its training data instead of the real pattern, and so fails on new data.

How surprising the data would be if the null hypothesis were true. It is not the probability that the hypothesis is true.

Splitting a long list into pages fetched one at a time. How you page decides whether items can be skipped or repeated when the list changes.

Observations of the same units, such as products, firms or people, repeated over time, so change can be studied within each unit.

A percentage point is the difference between two percentages; a percent change is that difference relative to the starting value.

The value below which a given percentage of observations fall. The 50th percentile is the median; the 90th marks the value that 90% do not exceed.

Reading data as it was known at a past moment — not as it looks today after later corrections and additions.

The population is everything you want to know about; the sample is the part you observed. Statistics asks what the second says about the first.

People click results partly because they are high on the page. Click counts therefore reflect the ranking as well as the relevance.

A distribution in which a few items are enormous and a long tail of items are tiny; the average is a poor summary of such data.

A planning exercise: imagine the project has already failed and write down why, before starting. Treating failure as certain makes people name risks they would otherwise keep quiet about.

Writing down, before seeing results, the question, the data and exactly how it will be analysed, and keeping that record unchanged.

Two measures of search and classification quality: how clean the returned results are, and how complete they are.

How strongly the quantity bought responds to a price change: the percentage change in quantity divided by the percentage change in price.

A code that names a product so records can be matched: a seller's SKU, a global GTIN (the number under a barcode), or a marketplace's own catalogue number.

The hierarchy of categories a platform uses to organise products. It differs between platforms and changes over time, so a category means something only with its version.

One version of a product that differs in a few attributes such as size or colour, usually grouped under a parent listing. Counts may refer to the parent or to a variant.

A measurable stand-in for something you cannot measure directly. Proxies work until they become targets: then the number stops tracking what it was meant to represent.

Removing training observations whose time span overlaps the test period, and skipping a short stretch after it, so no information leaks between the two.

Capping how many requests a client may make in a period. Servers apply it to protect themselves; responsible clients also impose it on themselves.

A rating is a score a customer gives; a review is the written text. The average score reflects the people who chose to rate, not all buyers.

The higher price shown beside the current price to suggest a saving. It is the seller's claim, to be stored separately from the price you pay.

Extreme measurements tend to be followed by less extreme ones, because part of any extreme value is chance. The worst look improved and the best look worse without any cause.

Human assessments of how well each result answers a query, recorded as grades. They are the answer key against which ranking methods are scored.

A result is reproducible when someone else, with the same data and code, gets the same result. It needs the data version, code, settings and environment to be recorded.

A plain-text file at a site's root in which the owner states which paths automated clients are asked not to fetch. A convention, not a lock.

A marketplace's ordering of products within a category by recent sales. It is relative, changes constantly, and is not a count of units sold.

The list or source a sample is actually drawn from. Whatever the frame leaves out cannot appear in the sample, however large the sample is.

Changing the structure of data over time while old data stays readable and existing consumers keep working.

What the person behind a query is trying to achieve: find a place, learn something, or do something such as buy. Relevance depends on it.

A regular pattern that repeats at a fixed period, such as higher sales every December. Without separating it, a normal seasonal rise is mistaken for growth.

The units in your data differ systematically from the group you want to describe, because of how they were chosen or kept. Results hold for the selected group only.

Automatically classifying text by the attitude it expresses, usually positive, negative or neutral. It reads wording, not truth, and has well-known blind spots.

One party's mentions as a share of all mentions of the parties being compared. It is relative, so it can fall while your own mentions rise.

A trend that appears in every subgroup can reverse or vanish when the groups are combined, because group sizes differ and the grouping is linked to both choice and outcome.

A descriptive attribute, such as a product's category, that changes rarely. Type 2 handling keeps every version with validity dates.

How far a source deserves to be believed on a given subject: its expertise, access to the facts, track record and incentives. It informs, but never replaces, checking the claim.

How far values typically lie from their average, in the units of the data. Small means tightly clustered; large means spread out.

A selection error from studying only what survived a process and ignoring what dropped out along the way.

Material produced once and published by many outlets, such as a news-agency story. Counting every publication overstates how many independent sources reported something.

A list of numbers representing a text so that texts with similar meaning lie close together. Useful for matching by meaning; hard to interpret.

A weight for a word in a document: how often it occurs there, multiplied by how rare it is across all documents.

Logic with a third value, unknown, besides true and false. A rule that depends on a missing value is unknown, not false.

Observations of one quantity recorded in time order. Neighbouring values are related and order matters, so time series need methods that respect time.

Splitting text into the units a search or analysis system works with, usually words. It decides what can match, and in Chinese finding the word boundaries is a task of its own.

A platform that connects two groups, such as buyers and sellers. Its rules for listing, ranking and fees shape the numbers it publishes.

UTC is the world's reference time scale and a time zone is a local offset from it. Store instants in UTC; convert for display.

Evaluating a forecasting method by training on data up to a date, testing on the period just after, then moving the date forward and repeating.

A standard file format that stores the raw requests and responses of web captures together with when and how each was made.

A marker of how far a pipeline has progressed — usually the latest time or sequence number fully processed — so the next run only asks for what comes after.

Collecting data by having a program request web pages meant for people and pull out the fields it needs. A fallback with real fragility.

Rates each candidate on several criteria, multiplies each rating by an agreed weight and adds them up. Trade-offs become explicit, and the answer depends on the weights.

How many standard deviations a value lies above or below the average of its group. It puts different measures on one scale.