Signals

What is the difference between a signal and sentiment?

Sentiment collapses an article to a single axis, which forces unrelated things to compete. A piece about a record harvest that also warns about export tariffs averages out to roughly neutral, and the average describes nothing that is in it.

Signals keep the dimensions separate. The same article can score high on agriculture, high on trade, and high on financial uncertainty simultaneously — which is a description you can filter on and act on.

Industry

How strongly an article reads as being about a sector, rather than merely mentioning one.

Business language

The register a piece is written in — how markets and lawyers talk, rather than the subject.

Risk context

Operational and systemic risk language, for monitoring rather than sentiment.

Filtering on them

Each signal is a number. Filter on a range, not a value, and stack constraints when you need more than one to hold at once.

# one signal, lower bound
GET /search?signal=cyber_incident&signal_min=0.25

# several at once — every constraint must match
GET /search?signals=energy:0.3,financial_stability_stress:0.2

# combine with any other filter
GET /stories?country=GM&signal=supply_disruption&signal_min=0.2

# the top 5% of the match set on one signal
GET /search?signal=cyber_incident&signal_percentile_min=95

A fixed signal_min admits more in a busy week than in a quiet one. signal_percentile_min keeps a share instead: 95 keeps the articles at or above the 95th percentile of that signal over the articles the rest of the query matched. The response reports the cutoff it used as signal_threshold, so you can see where the slice fell, compare it across runs or pin it as a fixed bound.

Every article response also carries its full signal map, so you can retrieve broadly and threshold client-side while you work out where the useful cut is.

Why does a signal filter return nothing at all?

Because the range excludes everything the other constraints matched, or the window is empty. A wrong name is no longer one of the causes. The engine stores signal_finance and accepts finance, and any other name is refused with unknown_signal and the nearest valid name, so a leading signal_ is an error you see rather than an empty set you misread. If a signal filter returns zero results where you expected some, loosen the range before changing anything else.

Choosing a threshold

There is no universal cutoff. Scores are relative to how strongly the language appears, and the useful threshold depends on how noisy your query already is. Retrieve a week without a signal filter, look at the distribution for the signal you care about, and cut where the results stop being about what you meant. GET /aggregate?interval=week&metrics=signal:cyber_incident reports the mean and the 50th and 95th percentiles for each week without retrieving the articles.

Re-check it quarterly. Editorial vocabulary shifts, and a threshold tuned during a quiet period will admit far too much during an active one. A percentile cut moves with the volume on its own, and in exchange it returns the same share of the match set whether or not anything unusual happened.

What they are not

Signals measure language, not events. A high risk score means an article is written in the register of that risk — it does not confirm the risk is real, material, or new. Treat them as a retrieval and triage tool, not as a finding.

They are also versioned. Responses carry the version that produced them, so a series computed across a version change is comparing two different things and should be recomputed rather than concatenated.

Reference

Every parameter, what it accepts and how it behaves against the index: the signal taxonomy.