Skip to content

Research

What we believe, and what we have found.

Six things we believe and six things we have found, the failures included. The ten platforms we study, and how we tell a good trend from a bad one. Every real figure on this page is a count of Wikipedia pageviews, a measure of attention.

3
tests with the rules written down first. One passed, one failed and one was inconclusive.
52
new Wikipedia topics in the second test: 28 German and 24 Swedish. 47 of them had days to grade.
10
platforms with a reader built. Two are live.
None
No sales outcome has been measured yet.

Six theses

What we believe.

  1. Attention is not sales.

    A view, a search or a post shows that people are paying attention. None of them is an order. Every source we read measures attention, so we report attention and say so. The step from attention to sales can only be measured where the orders are, in a shop.

  2. Most rises fade.

    Interest in a topic swings around its own seasonal norm. In 82 Wikipedia topics we had already studied, about a quarter of a month’s rise was given back the next month. That is a look back at topics we had already studied, not a test, and part of the pull back is day-to-day noise. A spike on a chart is usually not followed by interest that holds.

  3. A rise counts only if it lasts.

    We read a rise on the day it shows, and then we look again two weeks later. Only a rise whose level still holds is worth a closer look. We judge the outcome over the eight weeks after that check.

  4. Beat the simple rival first.

    A method is only interesting if it beats the simplest rule that could work. For ranking which topics will gain attention, last year is the hardest rival: the same weeks a year earlier say a lot about the next few. Our first grade did not beat it, and we say so.

  5. Show the work, failures included.

    Every number on this page comes from a file in our repository, and every result is computed by code there. For each test we write the rules down before we compute any result, run it once and publish the outcome, pass or fail.

  6. Use each source only as its rights allow.

    For every source we record six rights: to collect, store, analyse with AI, train on, show and sell. Unknown counts as no. Today only Wikipedia’s pageview counts may be shown in public, which is why every real chart on this page is drawn from them.

Six findings, with their limits

What we have found.

  1. Finding 01

    Rises that still held two weeks later went on to hold in a little over half of cases in Swedish and six in ten in German. Attention only.

    We wrote the rules down, then fetched 52 new Wikipedia topics and ran the test once. Of rises read on the day, about a third were followed by lasting interest. Of rises whose level still held two weeks later, a little over half were in Swedish and six in ten in German. A simpler rule, picking what did well a year earlier, came close: 52% in German and 45% in Swedish. Any day at random gave about one in five. That is a result about attention. It says nothing yet about sales.

    Edition

    German Wikipedia: 28 topics; 13,609 graded days.

    Real data · Wikipedia pageviews · pre-registered test, one run on 8 Nov 2026Share that held, with its intervalThe share of cases whose interest over days 15 to 70 after the reading was at least 1.25 times the earlier level. The window is the same for every rule.
    • Any graded day (base rate): 18.8% of 13,609 cases went on to hold (95% interval 15.7% to 21.9%). German Wikipedia.
    • Rise read on the day (single signal): 31.0% of 1,852 cases went on to hold (95% interval 27.5% to 34.3%). German Wikipedia.
    • Rise read and level still held two weeks later (confirmed): 59.1% of 447 cases went on to hold (95% interval 54.6% to 63.2%). German Wikipedia.
    • A winner one year earlier (seasonal naive): 52.2% of 2,367 cases went on to hold (95% interval 44.2% to 58.8%). German Wikipedia.
    • The value, with its 95% interval
    • Any graded day, 18.8%

    Rise read and level still held two weeks later (confirmed): 59.1% of 447 cases went on to hold (95% interval 54.6% to 63.2%). German Wikipedia. Source key: editions.de.wikipedia.point.confirmed

    Show as a table
    German Wikipedia: share of cases that went on to hold, by rule
    RuleHeld95% interval
    Any graded day18.8%15.7% to 21.9%
    Rise read on the day31.0%27.5% to 34.3%
    Rise read, and still held two weeks later59.1%54.6% to 63.2%
    A winner one year earlier52.2%44.2% to 58.8%

    Source: research/grade-backtest-2026-11/results-v2.json, key editions.de.wikipedia, computed by scripts/grade_v2_run.py; one run on 8 November 2026, intervals from 1,000 resamples.

    Real data · Wikipedia pageviews · pre-registered test, one run on 8 Nov 2026How real rises endedThe same cases, sorted into the five outcome tiers. All four rules are shown, the two rivals included.
    • 13,609 cases.
    • 1,852 cases.
    • 447 cases.
    • 2,367 cases.
    • Winner
    • Mid tier
    • Flat
    • Softening
    • Loser

    Any graded day: Winner 18.8%, Mid tier 15.3%, Flat 30.8%, Softening 9.3%, Loser 25.8%. 13,609 cases. Source key: editions.de.wikipedia.tier_shares.base

    Show as a table
    German Wikipedia: outcome tier of every case, by rule
    RuleWinnerMid tierFlatSofteningLoserOut of
    Any graded day18.8%15.3%30.8%9.3%25.8%13,609 cases
    Rise read on the day31.0%18.4%28.8%7.4%14.4%1,852 cases
    Rise read, and still held two weeks later59.1%18.1%17.2%3.1%2.5%447 cases
    A winner one year earlier52.2%21.3%17.0%2.5%7.0%2,367 cases

    Source: research/grade-backtest-2026-11/results-v2.json, key editions.de.wikipedia, computed by scripts/grade_v2_run.py; one run on 8 November 2026, intervals from 1,000 resamples.

    The test we had written down, confirmed against read on the day: passes in both editions.

    Limits: this is attention, not sales. Part of the gain comes from waiting itself, because a level that has held for two weeks tends to keep holding; the plain rival “the level held, with no signal” was not tested. The confirmed rule acts two weeks later and rarely (447 of 13,609 German graded days). Against “a winner one year earlier”, the confirmed rule leads by +6.8 points in German (interval +0.2 to +14.1) and by +9.1 in Swedish (−3.7 to +20.5, which includes zero): last year’s outcome alone gets most of the way. The Swedish intervals rest on 17 topics. 52 topics were picked by hand, and cases from one topic overlap in time, so the intervals are likely too narrow. One run.

  2. Finding 02

    Our first grade showed no skill.

    Our first grade ranked the next four weeks of attention less well than a rule that just looks at last year. We had set the bar before the test, and the grade missed it. We kept the result and then tested a different rule (finding 01).

    Real data · Wikipedia pageviews · pre-registered test, one run on 5 Nov 2026The grade against its rivalsHow well each rule ranked the next 28 days of Wikipedia attention, as a rank correlation: 1 means the same order and 0 means no relation. Higher is better; the grade had to clear the bar.
    • The 0 to 10 grade: rank correlation 0.256 with the next 28 days (95% interval 0.222 to 0.290), over 3,660 pairs.
    • The plain ratio (last weeks over the weeks before): rank correlation 0.368 with the next 28 days (95% interval 0.328 to 0.410), over 3,660 pairs.
    • Last year's same change (seasonal naive): rank correlation 0.427 with the next 28 days (95% interval 0.334 to 0.521), over 3,660 pairs.
    • The value, with its 95% interval
    • The bar the grade had to clear, 0.477

    The 0 to 10 grade: rank correlation 0.256 with the next 28 days (95% interval 0.222 to 0.290), over 3,660 pairs. Source key: test.spearman.grade

    Show as a table
    Rank correlation with the next 28 days, test years 2023 to 2026
    RuleRank correlation95% interval
    The 0 to 10 grade0.2560.222 to 0.290
    The plain ratio (last weeks over the weeks before)0.3680.328 to 0.410
    Last year's same change (seasonal naive)0.4270.334 to 0.521

    Source: research/grade-backtest-2026-11/results.json, computed by scripts/grade_backtest.py; one run on 5 November 2026.

    Verdict, in the result file’s words: no demonstrated skill.

    Test years 2023 to 2026: 3,660 pairs, 41 Swedish Wikipedia topics. The grade trailed last year’s change by between 0.068 and 0.284 (95% interval).

    Limits: attention only; only larger articles (9,673 of 16,154 test pairs were under the floor of 20 views a day); overlapping windows, so the intervals are narrower than the real uncertainty; hobby topics picked by hand. It neither proves nor disproves the framework.

  3. Finding 03

    Most single rises do not hold.

    Real data · Wikipedia pageviews · pre-registered test, one run on 8 Nov 2026

    Of rises read on the day, 31.0% in German Wikipedia (1,852 cases) and 31.5% in Swedish (944 cases) were followed by interest that held a quarter above its earlier level. A further 18.4% in German and 15.6% in Swedish ended a tenth to a quarter up, which is mid tier. The landing page gives the first two figures, with the market and the cost of a parcel beside them: why this should exist.

    Source: research/grade-backtest-2026-11/results-v2.json, keys editions.de.wikipedia.point.single and editions.sv.wikipedia.point.single, computed by scripts/grade_v2_run.py; one run on 8 November 2026.

    Limits: as for finding 01, which is the same run.

  4. Finding 04

    Attention returns toward its norm.

    Real data · Wikipedia pageviews · description of articles already seen, not a test

    Across 82 Wikipedia topics over the last five years (14,342 pairs), the next 28 days’ change was −0.255 times the past 28 days’ change (95% interval −0.293 to −0.220), plus 0.248 times the same change a year earlier (0.188 to 0.319).

    Source: research/speculation/opportunity-lab-v1/results.json, key momentum_all_editions_last_five_years, computed by scripts/opportunity_lab_run.py on 5 November 2026.

    Limits: an in-sample description of articles we had already looked at, not a pre-registered test; attention only; hobby topics. The past and the next 28-day changes meet at the same day, so day-to-day noise on that day alone pushes the coefficient below zero; part of the pull back is that noise.

  5. Finding 05

    On the day, a news week looks like a lasting rise.

    Simulation · not real topics, platforms or markets

    In 300 simulated worlds of a one-week surge on every source, 95.3% read as a rise on the day (95% interval 92.3% to 97.2%) and 56.7% scored 8 or more (51.0% to 62.2%); 17.0% passed the two-week check (13.2% to 21.7%).

    Simulation · not real topics, platforms or marketsWhy the second check mattersShare of simulated topics of each kind that look like a rise on the decision day, and the share still up two weeks later. Switch between the two: a news week drops.
    Show
    • Lasting rise: 69.7% read as a rise on the day (95% interval 64.2% to 74.6%); 52.7% still up two weeks later (95% interval 47.0% to 58.2%). 300 simulated worlds.
    • Slow rise: 32.0% read as a rise on the day (95% interval 27.0% to 37.5%); 15.0% still up two weeks later (95% interval 11.4% to 19.5%). 300 simulated worlds.
    • Modest lasting rise: 30.3% read as a rise on the day (95% interval 25.4% to 35.8%); 14.0% still up two weeks later (95% interval 10.5% to 18.4%). 300 simulated worlds.
    • Fad: 67.3% read as a rise on the day (95% interval 61.8% to 72.4%); 36.3% still up two weeks later (95% interval 31.1% to 41.9%). 300 simulated worlds.
    • News week: 95.3% read as a rise on the day (95% interval 92.3% to 97.2%); 17.0% still up two weeks later (95% interval 13.2% to 21.7%). 300 simulated worlds.
    • Same peak every year: 28.3% read as a rise on the day (95% interval 23.5% to 33.7%); 11.0% still up two weeks later (95% interval 7.9% to 15.0%). 300 simulated worlds.
    • One-day spike, all sources: 7.7% read as a rise on the day (95% interval 5.2% to 11.2%); 0.3% still up two weeks later (95% interval 0.1% to 1.9%). 300 simulated worlds.
    • One-day spike, one source: 20.7% read as a rise on the day (95% interval 16.5% to 25.6%); 4.3% still up two weeks later (95% interval 2.5% to 7.3%). 300 simulated worlds.
    • No trend: 18.3% read as a rise on the day (95% interval 14.4% to 23.1%); 6.3% still up two weeks later (95% interval 4.1% to 9.7%). 300 simulated worlds.
    • Slow fade: 7.0% read as a rise on the day (95% interval 4.6% to 10.5%); 2.0% still up two weeks later (95% interval 0.9% to 4.3%). 300 simulated worlds.
    • Lasting fall: 1.3% read as a rise on the day (95% interval 0.5% to 3.4%); 0.3% still up two weeks later (95% interval 0.1% to 1.9%). 300 simulated worlds.
    • Share of simulated worlds read as a rise on the decision day
    • 95% interval over the 300 worlds

    News week: 95.3% read as a rise on the day (95% interval 92.3% to 97.2%); 17.0% still up two weeks later (95% interval 13.2% to 21.7%). 300 simulated worlds. Source key: scenarios.news_week

    Show as a table
    Simulated topics: read as a rise on the day, and still up two weeks later
    Kind of topicRead as a rise on the dayStill up two weeks later
    Lasting rise69.7%52.7%
    Slow rise32.0%15.0%
    Modest lasting rise30.3%14.0%
    Fad67.3%36.3%
    News week95.3%17.0%
    Same peak every year28.3%11.0%
    One-day spike, all sources7.7%0.3%
    One-day spike, one source20.7%4.3%
    No trend18.3%6.3%
    Slow fade7.0%2.0%
    Lasting fall1.3%0.3%

    Source: src/components/demo/fixtures/research-preview.json (scenarios) and research/speculation/opportunity-lab-v1/results.json, computed by scripts/opportunity_lab_run.py with seed 20261105 on 5 November 2026; 300 simulated worlds of each kind.

    Limits: the numbers describe our simulator, which is calibrated on real Wikipedia attention, not the world.

  6. Finding 06

    What we have not done yet.

    • No sales outcome has been measured: no shop’s orders are linked to any reading.
    • How early a reading comes was defined on 10 November 2026, as a number of days; it has never been measured.
    • No model has been run on our data.
    • The holiday signal test has not been run: its rules are not frozen yet.

    What we have not shown: that any reading leads to sales, or how many days a reading comes before anything else. We will say so when a test of ours shows it, and not before.

Names, status and rights only

The ten platforms we study.

Ten platforms, two of them live today. For each one we record what we read, how we reach it and what its rights allow. Only Wikipedia’s counts may be shown in public, so they are the only platform values on this page. We read no platform through a fake account, a solved CAPTCHA or rotated addresses.

Status
  • WikipediaShown in public: YesLive
    What we read
    Daily pageview counts of single articles, per language edition (Swedish, German, English).
    How we reach it
    Wikimedia Analytics API, public and CC0; collector on the owner's server.
    Status
    Four collector jobs healthy.
    Shown in public
    Yes
    Read by AI
    No, under review
    Used for training
    No, under review
  • Google TrendsShown in public: NoLive
    What we read
    The Trending Now list of search terms, with Google's own traffic buckets.
    How we reach it
    Public Trending Now RSS feed (live); the Google Trends BigQuery dataset (reader built, off).
    Status
    Four RSS jobs healthy; the BigQuery reader has no job.
    Shown in public
    No
    Read by AI
    No
    Used for training
    No
  • PinterestShown in public: NoPaused
    What we read
    Pinterest Trends: a relative interest index per keyword. Stored on the server only, never shown.
    How we reach it
    Pinterest API with a short-lived trial token.
    Status
    Its daily and monthly jobs paused: the trial token expired.
    Shown in public
    No
    Read by AI
    No
    Used for training
    No
  • TikTokShown in public: NoWaiting on approval
    What we read
    Aggregate hashtag data only: rank, approximate post and view counts. No account, no user.
    How we reach it
    Creative Center public preview (reader kept, jobs paused); Business API discovery (applied for).
    Status
    Preview jobs paused; Business API discovery access not approved.
    Shown in public
    No
    Read by AI
    No
    Used for training
    No
  • YouTubeShown in public: NoWaiting on approval
    What we read
    Count and reported views of videos that match a search, kept at most 28 days.
    How we reach it
    YouTube Data API; reader built, no job.
    Status
    API audit form submitted; no reply is recorded.
    Shown in public
    No
    Read by AI
    No
    Used for training
    No
  • InstagramShown in public: NoOff
    What we read
    Hashtag counts (planned). The older public-page reader is disabled.
    How we reach it
    Instagram Graph API hashtag search, which needs Meta's App Review; not applied for.
    Status
    Wanted later, by the owner's decision.
    Shown in public
    No
    Read by AI
    No
    Used for training
    No
  • FacebookShown in public: NoOff
    What we read
    Page-level counters, only for Pages that grant access.
    How we reach it
    Meta Graph API; adapter built, no job.
    Status
    No job.
    Shown in public
    No
    Read by AI
    No
    Used for training
    No
  • BlueskyShown in public: NoOff
    What we read
    A bounded sample of public posts that match a search; not a full count.
    How we reach it
    Bluesky public search; adapter built, no job.
    Status
    Public search answered HTTP 403.
    Shown in public
    No
    Read by AI
    No
    Used for training
    No
  • MastodonShown in public: NoOff
    What we read
    Trending hashtags of one Mastodon instance, with daily post and account counts.
    How we reach it
    Public trends API of the instance, no account and no key; reader built, no job.
    Status
    No job.
    Shown in public
    No
    Read by AI
    No
    Used for training
    No
  • Hacker NewsShown in public: NoOff
    What we read
    Headlines, points, comment counts and list position of the first stories in one list.
    How we reach it
    Official Hacker News API, no account and no key; reader built, no job.
    Status
    No job; the site's terms bar scraping.
    Shown in public
    No
    Read by AI
    No
    Used for training
    No
Live
a job runs on our server and its uploads are accepted.
Paused
a job exists but is stopped.
Waiting on approval
we have applied and the platform has not approved.
Off
a reader is built and no job runs.

Rights come from our rights matrix: an unknown answer, a policy block and an answer under review all count as no.

Five outcome tiers

Good, medium and bad trends.

We grade an outcome, not a hope. Over the eight weeks after the check, we compare interest with where it stood before the rise. Up by a quarter or more is a winner; up by a tenth to a quarter is mid tier; within a tenth either way is flat; down by a tenth to fifteen percent is softening; down by fifteen percent or more is a loser.

Definition · from our codeThe five outcome tiersA tier is an outcome, read after the fact: the average interest over the selling window divided by the level before. Press a tier to light its band in the charts of this page.

Source: python/trendfalcon/opportunity_lab.py (the tiers and their edges), python/trendfalcon/trend_rules.py (1.25 and 0.85), src/lib/success-framework.ts (the words). Edge rule: exactly 0.85 is a loser; every other edge belongs to the tier above it.

The reading and the check that come before the outcome

Rise read
The grade is 6 or more, out of 10, on the decision day.
Confirmed
14 days later, the latest week is still at least 1.25 times the level before. This checks the level, not the momentum.
Selling window
The 56 days from day 15 to day 70 after the decision day.
Looks like a winner
Read as a rise on the day, but gone before the selling window: a fad or a news week.
Simulation · not real topics, platforms or marketsWhat each tier looks likeOne simulated case each for four of the five tiers, on the same scale. Ten simulated sources of one topic: each line is one source, with its own level before the change set to 100; the bold line is their median and the shaded band the middle half.

Source: src/components/demo/fixtures/research-preview.json (gallery), built by scripts/build_framework_charts.py with seed 20261108 and almost no noise.

Simulations, labelled as simulations

The same ideas in simulated worlds.

A simulator lets us try the rules on topics whose true course we know. It is calibrated on real Wikipedia attention, and it is not the world: nothing in this section is a real topic, platform or market.

Simulation · not real topics, platforms or marketsTen sources at onceTen simulated sources, not the ten platforms. The same idea with realistic noise: sources jump, dip and disagree, so the clean picture above has to be read through it. Pick a case; the strip under the chart counts how many of the ten agree on each day.

Most sources rise together and stay up.

Scale, ten sources at once
  • Decision day
  • The check, two weeks later
  • The weeks a shop would sell
  • The weeks the level before is read from
  • 100, the level before
  • 125, the winner line

How many of the ten agree. Above the line: sources at 1.25 x or more of their level before. Below it: sources at 0.85 x or less.

Day +42: median 140, middle half 130 to 146; 8 of 10 clearly up, 0 of 10 clearly down (100 is each source's level before).

  1. Decision daygrade 8: read as a rise
  2. Two weeks laterlatest week at 1.55 x the pre-signal level: confirmed
  3. Selling window1.55 x the pre-signal level: winner tier
Show as a table
A lasting rise: the ten simulated sources on five days, 100 = the level before
DayMedianMiddle halfClearly upClearly down
Day −2810396 to 1161 of 100 of 10
Decision day136124 to 1497 of 100 of 10
Day +14, the check155139 to 1908 of 100 of 10
Day +42140130 to 1468 of 100 of 10
Day +70150142 to 1589 of 100 of 10

Source: src/components/demo/fixtures/research-preview.json (tenSources), built by scripts/build_framework_charts.py: one typical simulated world of each kind.

Simulation · not real topics, platforms or marketsHow each kind of topic endsHow simulated topics of each kind end up. A lasting rise is mostly a winner; a fad or a news week looks like one and usually is not; a slow fade and a lasting fall end as losers. With 300 worlds of each kind, a share carries about ±2 to ±6 points.
  • 300 simulated worlds.
  • 300 simulated worlds.
  • 300 simulated worlds.
  • 300 simulated worlds.
  • 300 simulated worlds.
  • 300 simulated worlds.
  • 300 simulated worlds.
  • 300 simulated worlds.
  • 300 simulated worlds.
  • 300 simulated worlds.
  • 300 simulated worlds.
  • Winner
  • Mid tier
  • Flat
  • Softening
  • Loser

Lasting rise: Winner 69.7%, Mid tier 14.0%, Flat 13.0%, Softening 1.0%, Loser 2.3%. 300 simulated worlds. Source key: scenarios.held_rise.tiers

Show as a table
Simulated topics: outcome tier in the selling window, by kind of topic
Kind of topicWinnerMid tierFlatSofteningLoserOut of
Lasting rise69.7%14.0%13.0%1.0%2.3%300 simulated worlds
Slow rise46.7%21.3%22.3%3.3%6.3%300 simulated worlds
Modest lasting rise39.7%19.0%27.0%3.7%10.7%300 simulated worlds
Fad35.7%22.3%21.7%6.3%14.0%300 simulated worlds
News week19.3%17.3%28.3%8.0%27.0%300 simulated worlds
Same peak every year14.0%19.7%32.3%7.3%26.7%300 simulated worlds
One-day spike, all sources19.7%15.3%33.3%8.0%23.7%300 simulated worlds
One-day spike, one source18.3%16.3%30.3%9.7%25.3%300 simulated worlds
No trend20.7%24.0%31.7%5.0%18.7%300 simulated worlds
Slow fade5.3%7.7%26.7%9.0%51.3%300 simulated worlds
Lasting fall2.0%3.7%10.3%5.7%78.3%300 simulated worlds

Source: src/components/demo/fixtures/research-preview.json (scenarios), computed by scripts/opportunity_lab_run.py with seed 20261105 on 5 November 2026; 300 simulated worlds of each kind.

Simulation · not real topics, platforms or marketsExample worldsOne simulated world of each kind: what we would have seen each day, against the simulator’s true level, and what the three checks made of it.

Interest steps up and stays up. The rise is read on the decision day, the level is still up two weeks later, and the selling window that follows is clearly above where it started.

  • What we would have seen
  • The simulator's true level
  • Decision day
  • The check, two weeks later
  • The weeks a shop would sell
  • The weeks the level before is read from
  • 100, the level before
  • 125, the winner line

Day +42: we would have seen 173; the simulator's true level was 149 (100 is the level before).

  1. Decision daygrade 6: read as a rise
  2. Two weeks laterlatest week at 1.63 x the pre-signal level: confirmed
  3. Selling window1.52 x the pre-signal level: winner tier
Show as a table
A lasting rise: what we would have seen and the simulator's true level on five days, 100 = the level before
DayWhat we would have seenThe simulator's true level
Day −2812698
Decision day159125
Day +14, the check144127
Day +42173149
Day +70160162

Source: src/components/demo/fixtures/research-preview.json (examples), built by scripts/build_framework_charts.py with seed 20261107.

Read on.

The market and the cost of a parcel are on the landing page, each with its source.