Narrative Shift Detector

WHAT IT DOES. It reads the news in the current window for each tracked asset and counts bullish and bearish keywords in each article's headline and description, each keyword at most once per article. Read the run's published computation rule for the score's normalization, rounding, coverage-direction boundaries and thin-coverage precedence. Magnitude is the size of that score, not confidence in the evidence or a price forecast. Distinct keywords and keyword hits are different quantities: one keyword found in ten articles is one distinct keyword and ten hits.

A LEVEL, NOT A SHIFT. No earlier window is measured, so every reading describes coverage in this window and none describes a change from a previous one. Measuring a shift is change-point detection: comparing a recent window against a reference window and asking whether the difference exceeds what the series' own variability would produce, with the threshold set against that noise level rather than chosen for looking reasonable. This analysis does none of that, and a reading here must never be described as a turn, a reversal or a move away from a baseline.

'NEW COVERAGE' IS THIN COVERAGE. It marks sparse article coverage under the run's published rule: little evidence, read on its own terms, not a story that has changed direction.

THE FLAG, AND THE MULTIPLE-COMPARISONS PROBLEM. When flags are enabled, the run's published rule states the eligible readings and the thresholds applied. A flag threshold never changes a label. Scanning more assets increases the expected number of chance extremes whenever readings are noisy. Be precise about what that does and does not imply: it does not follow that a flag appears every run, nor that most are false. Formal control would require a defined null model, calibrated p-values or equivalent scores, and a procedure such as Benjamini-Hochberg. A FIXED MAGNITUDE CUTOFF DOES NOT CONTROL THE FALSE-DISCOVERY RATE; it is a cutoff, not a correction.

WHAT THE MEASUREMENT ACTUALLY IS. Keyword counts from HEADLINES and descriptions, which are written to attract attention and are a compressed and biased sample of an article. Coverage is also episodic: an asset with no qualifying news has MISSING coverage data - not neutral coverage, and certainly not an absence of sentiment among market participants. Missing data and a no-direction reading must not be averaged together. Coverage volume also varies with price, so the coverage series and the price series are not independent to begin with. A scale endpoint cannot rank readings beyond it. The hit and article counts describe how much coverage supports a reading; magnitude is not confidence.

THE CLAIM TO REFUSE, and it is the one every sentiment product is sold on: that catching a turning narrative EARLY is where an edge lies. This analysis does not measure turns at all, and nothing in it establishes that a reading PRECEDES a price move rather than following one. Demonstrating a lead is a specific empirical exercise - a time-ordered test of the signal against SUBSEQUENT returns, out of sample, net of costs - and this analysis does not perform it. Treat a strong reading as a prompt to look.