How Noiseradar works

What the number means

Each market gets one reading per day. 1.00 means the topic is at its own 90-day normal. 2.00 means twice the usual attention, 0.50 means half.

The baseline is calculated separately for each topic, so readings are comparable across markets of any size.

Examples

The first three are dates with known events: May the 4th, the anniversary of the first Star Wars release, and Pokémon Day. The method has no calendar or event list. These peaks come from the source data.

Sources

The index uses four public sources. It uses no data from Noise, including prices.

Weights are the same for every market and add up to 1.00. Half-life is the smoothing period used in step 2 below.

How a reading is calculated

1. Every morning each source is queried for the past 7 days. Figures that arrive late are assigned to the day they belong to.
2. Each source's daily value is smoothed with an exponential moving average.
3. The smoothed value is divided by the topic's median over the previous 90 days. The result is the source's factor, where 1.00 is normal for that topic on that source. A source needs 21 days of history before it produces a factor.
4. The factors are combined with a weighted geometric mean.

The full history is recalculated from stored raw data on every run.

Why a geometric mean

With a geometric mean, one source cannot produce a high reading alone. If news is at 10 times normal and the other three sources are at 1.00, the index is 2.00. An arithmetic mean would give 3.70.

Coverage

Coverage is the share of total weight that reported on a given day. At 0.60, sources carrying 60% of the weight contributed to the reading. It is shown as coloured dots on each chart and as coverage in the API.

Data rules

• For Hacker News and Wikipedia edits, a day with no activity is stored as zero.
• If a source is late, its last factor is carried forward for up to 3 days. After that the source is left out of the reading until new data arrives.
• A market with less than 120 days of history is marked provisional.
• Weights, windows and thresholds are identical for all 49 markets.

Daily schedule

The pipeline runs once a day at 07:30 UTC in four steps: collect, recalculate, check freshness, post.

The freshness check verifies that every source returned new data. If one did not, nothing is posted that day.

History starts on 2025-08-18. The longest series is 439 days.

Posts on X

The account posts a chart when a market's index changes by:

• 50% or more in 24 hours
• 100% or more in 7 days
• 40% or more in 7 days, with at least 5 of those days moving in the same direction

A post is published only if all of the following are true:

• At least 2 sources reported on both the first and the last day of the period.
• The data for both days is current.
• The index reached 1.00 or higher during the period.
• The market is not provisional.
• No single headline makes up more than half of that day's news articles on the topic. This removes cases where one story was reprinted by many outlets.

There are at most 2 posts per day. A market is not posted again for 5 days unless the new change is at least 2 times larger than the one already posted.

Beta index

A second index adds YouTube as a fifth source. It is available at ?preset=beta on the site and in the API. Posts on X use the four-source index.

API

All data is available as JSON. No key is required.

Add ?preset=beta for the five-source index. Data updates once a day. Responses support ETag.

Limits

• The index measures attention. It is not a price feed or a trading signal.
• Instagram and TikTok are not included. Neither offers an API for public data on how often a topic is posted about. Topics that are discussed mostly there are measured through news and Wikipedia only.
• Topics without a specific name, such as 90s-nostalgia, are harder to match and are measured less precisely.