Method, version 0.5

A narrative is a claim about something in the world that can be written as one sentence and quoted, in some wording, from dated public text.

Daylight saving time is a topic. “Changing the clocks makes people sick” is a narrative. Both lines were written for this page.

Two texts carry the same narrative when they make the same claim about the same subject, even if the wording differs. The Method tracks the claim. Whether it is true is a separate question.

A series is the count behind a narrative or a topic: distinct items per calendar week, from a written search of one named public source. An item is one published text. A publisher is whoever put it out.

Copies are folded first. A wire story run by forty papers is one item with forty publishers.

A series counts what was published, which is a different thing from what people think.

A narrative or a topic becomes a trend here when it has a series and the series changes. A topic trend shows a subject being written about more or less. It says nothing about any claim in those texts.

A micro trend has risen for 3 weeks running and is still small enough to list on one page: fewer than 10 publishers, or one place, or one language, or one community.

A macro trend has held a weekly count for at least 12 weeks and is carried by 10 or more publishers, under different owners, in at least 3 regions or communities. It must also do this among two kinds of publisher, each tested on its own. On Trends that means passing in both company filings (rule 1.6) and state bills (rule 1.7), each in its own latest window. The counts are never added together.

A trend that meets neither definition is unclassed until it meets one. Micro and macro describe reach alone. They are also the first two trend classes, which are the groups any error rate would be reported by.

How often a micro trend becomes a macro trend is still unmeasured. Getting that figure is part of why this project exists.

These six terms have a fixed meaning on this site. Window A is the last four full weeks before a prediction’s issue date. Window B is the four full weeks ending on its resolution date.

Grow

The count in B is at least 25 percent above the count in A

Hold

The count in B is less than 25 percent above or below the count in A

Fade

The count in B is at least 25 percent below the count in A

Resurface

The series stayed at or under 10 percent of its peak for at least 8 weeks, then a single week reached 50 percent of that peak

Spread further

B holds items from 3 or more publishers that carried nothing in the prior 12 weeks

National coverage

3 or more outlets on the prediction’s written list of national outlets publish a distinct item inside the horizon, the days between issue and resolution. The list is fixed on the issue date

“Peak” stays off this list, because a peak can only be seen afterward. If any of these six is used loosely on this site, that is an error for the Record.

Trends borrows Grow, Hold and Fade from rule 1.4 and reads them over longer windows, because a weekly count of federal documents jumps around. It adds up the last 12 full weeks and the 12 before them. At least 25 percent up is Grow, at least 25 percent down is Fade, and anything between is Hold. For a prediction, the windows in rule 1.4 still decide.

Under 24 items across both windows, about one a week, the reading is Too few to read and no percentage is shown. A rise from one item to three is not a finding. The twelve-week window was set before the first run of the full list and the floor after it. Both numbers are proposed.

A reading describes the window just ended. It says nothing about what comes next, and it does not adjust for the time of year. Trends also gives a longer comparison, the last 52 weeks against the 52 before, which holds every season on both sides. Under 104 items across those two windows, it gives no percentage. That comparison and its floor were set after the first counts were seen. Once or twice a year the government’s regulatory agenda mentions many topics at once, and a reading moves when such a day enters or leaves a window.

In the Federal Register, a publisher is a top-level agency with at least one item in the window, and a bureau or office is folded into its department, following the source’s own list. Nearly every publisher belongs to one government, and all of them publish through one federal journal. So a Federal Register series cannot be a macro trend, however wide it runs, and Trends says which parts of the test it passes. For a micro trend, only the publisher limit in rule 1.3 is applied, because this source is one language and one place.

Trends counts each document as the source returns it. No copy check is run and no item is audited, so these counts have not been through stage 2.2.

A topic stays on the list once it is there. A search can be corrected, with the date and the reason written on Trends.

The second source on Trends is SEC EDGAR full-text search, limited to six report forms: 10-K, 10-Q, 8-K, 20-F, 40-F and 6-K. Amended versions count too. Each topic has its own series there, under the same fixed list. The sources are never added together. The form list was set after test counts were seen, and it is proposed.

An item here is a document, by filing date, so a report and each attachment count separately. A publisher is a filer, one number in the SEC’s list. Filers under one parent are not folded, so owners can be fewer than filers. A region is one of the four US Census regions, or outside the 50 states, by the filer’s principal office. That grouping is proposed.

Companies file around their reporting dates, so a short window mostly shows the calendar. The reading in rule 1.5 and the micro test are not run on this source. Only the comparison of the last 52 weeks with the 52 before is given.

A topic passes the test in company filings when it has a document in each of the last 12 weeks, from 10 or more filers, in at least 3 regions. Passing here alone is not a macro trend. Every publisher here is a company reporting under one law through one federal system, so it shows reach among companies only. A macro trend must also pass in state bills, under rule 1.3.

These counts have not been through stage 2.2 either. Each one can be checked against the source’s own search page, which gives the same number.

The third source on Trends is the monthly database file from Open States, which holds the bills of every state legislature, from 2017 in full. Each topic has its own series there, under the same fixed list. The sources are never added together.

An item here is a bill or resolution, by the date of its first action. It counts when one of the topic’s phrases is in its title or summary. Full bill text is not used, because too much of it is missing from the file. A publisher is one legislature, of the 50 states, the District of Columbia and Puerto Rico, and its owner is that government. Congress is left out. A region is one of the four US Census regions, or Puerto Rico. That grouping is proposed.

Most legislatures sit for part of the year, and more bills are filed in odd years. So the reading in rule 1.5 and the micro test are not run on this source, and a count of the last 52 weeks is compared with the 52 before only as a share of all bills.

A topic passes the test in state bills when it has a bill in each of the last 12 weeks, from 10 or more legislatures, in at least 3 regions. Passing here alone is not a macro trend. These owners differ, but they are one kind of publisher. A macro trend must also pass in company filings, under rule 1.3.

The file has no public search, so a count here cannot be checked with one link. It is checked another way: the dated file, a hash of what was read, the published script and the list of bills behind every count. The bills of one state, Washington, are also compared with that state’s own list on every run. These counts have not been through stage 2.2.

Find is reading public text to locate a narrative. The place and date of the first sighting here get written down.

That records a sighting. Earliest observed source does not establish original authorship. Some content is never public, and some is removed before anyone sees it. Closed groups, private messages and anything behind a login are out of bounds.

Audit is going back to each item and confirming it against the original or an archived copy. Copies are folded. Changed wording is marked. Each claim gets a status word.

Audit records what each item says and leaves true and false to fact-checkers. Forty papers running one wire story are one source, because repetition is not corroboration.

Analyze is reading the checked items as a series. It asks how the wording changed and when its reach widened, then which other explanations fit the same pattern. Trends runs a first, simple part of this on counts alone, under rule 1.5.

Intent, and who is behind a narrative, stay outside it. A pattern that looks coordinated may indicate coordination and requires corroboration. No account is called a bot. As a product stage this is scheduled, with software to come after item collection exists.

Predict is writing down, in advance, how one measure of one narrative will move by a stated date, with the rule that decides it written first.

It is never about prices, demand, public confidence, votes or any other outcome outside the text. Nothing is predicted about a named person or organization. A case can end without a prediction. As a product stage this is scheduled.