Trend Radar · model radar-1.1
Calls beforethe peak.
A real, documented scoring model built only from data we collect: 125 countries' trending lists, snapshotted about every 30 minutes. Candidates are ranked by momentum, not hype. Every call is timestamped and graded later, in public.
Live candidates
Rising and early, scored now
Top-scored terms on the latest snapshot, sorted by confidence tier, then score (low-confidence scores in amber). Single sightings are "no signal" and are never ranked. Hatched bars mean a feature is still unknown (not enough history yet), which lowers confidence rather than the score.
Calls ledger
Every call, hit or miss
When a candidate clears the bar, we record a call with its timestamp, score and reasons. A random sample of eligible terms the model didn't pick is recorded the same way as a control group, so you can see whether the model beats chance.
Methodology
How the Radar scores a trend
Catalyst-style scoring: the principles come from NYPTID's Catalyst Engine, a YouTube analytics engine that ranks videos on quality signals instead of raw views. This is a fresh implementation of those principles for search trends, not that engine's code.
1 · Data we actually collect
About every 30 minutes (on request when data is stale, plus an hourly scheduled run) we fetch the public Google Trends trending-searches feed for all 125 countries and store one snapshot of the whole world. For each term we keep its approximate-searches bucket over time, when it started, when we first and last saw it, which countries list it and the distinct news headlines attached. Terms are matched across countries by normalized spelling (case, accents and punctuation removed), so translated names count separately.
2 · Catalyst-style principles
- Quality over raw volume. Momentum and relative standing in each country matter more than the biggest absolute number.
- Tiny samples are no signal. A term seen once, in one country, is not scored low; it isn't ranked at all.
- Wait before judging. Velocity, news and persistence need two snapshots at least 30 minutes apart.
- Unknown is not zero. Missing features are left out of the weighted average and lower confidence instead of the score. A country whose feed fails is treated as unknown, not as the trend ending.
3 · Features and weights
| Velocity | Best per-country growth in approx. searches between our snapshots, in doublings per hour (1.5/h or faster = full marks). Zero if flat or falling off its peak. | 24% |
| Cross-country spread | How many countries list the term now (log scale, 16+ = full) plus how many picked it up in the last 6 hours. | 19% |
| News acceleration | New distinct headlines in the Google Trends feed per hour, across all countries (6/h = full). | 10% |
| Coverage acceleration | Articles per hour about the term in the last 6 hours vs the 24 hours before, from Google News search results (8× pace = full). Unknown below 5 dated articles. | 10% |
| Source diversity | Distinct outlets and platforms covering it (Google Trends feed, Google News, Hacker News, Wikipedia, GDELT), log scale, 32+ = full. | 9% |
| Recency | How fresh the breakout is: decays with an 8-hour time constant from the earliest start time. | 10% |
| Persistence | Share of our snapshots it stayed on the list since first seen, scaled up to 3 hours of history. Filters one-snapshot blips. | 10% |
| Relative strength | Its volume relative to the top term in each country where it's listed, not raw volume, so a small country's #1 counts. | 8% |
Score (0 to 100) is the weighted average of the features we could measure:
4 · Confidence, calls and grading
- Confidence comes from sample size: total sightings (2+, 4+), hours of history (1h+, 3h+), country count (2+) and how much of the model is measured. Below 60% measured caps it at low.
- A call is made when a term scores 60+ with medium or high confidence, velocity, persistence and news measured, and at least 75% of the model's weight measured. At most 5 per snapshot, and never the same term twice within 48 hours.
- Hit if within 48 hours its approximate searches rise above the level at call time in any country, or it reaches the trending list in 2+ more countries. Otherwise miss.
- Control group: for every call, one random eligible term the model scored below 60 is graded the same way. The model is only useful if its hit rate beats the control's.
5 · Sources and their live status
Cross-source lookups run for the strongest terms each snapshot, cached and rate-limited. If a source fails, it shows as unavailable here and its features count as unknown, never as zero.
Limits, honestly
Google's feed lists about 10 terms per country with bucketed volumes (100+, 200+, 500+…), so small changes are invisible and a term can only be seen once it's already trending somewhere. The Radar predicts whether a listed trend will keep growing or spread, not trends nobody is searching yet. Early on, with few snapshots, most features are unknown and confidence is low; we show that instead of guessing. No numbers on this page are invented: everything comes from /api/radar.