How the scores work

Inputs: Google Trends (literal search term, US), English Wikipedia pageviews (human users), and reviewer-title video counts (yt-sportswear-reviewer-panel-v1 for Sportswear, yt-reviewer-panel-v1 for Automotive). Videos posted by brands are excluded (brand publishing, not public attention). Legacy YouTube search estimates and social data are excluded.

Reference window: fixed 2026-04 to 2026-09 (6 months; minimum 6). Every reference month must be complete and collected under one method: one Trends export batch and scale group, one reviewer panel version. A new window means a new version, never a silent change.

Normalization: mid-rank percentile. P = 100 × (number of reference months below the latest value + 0.5 × number equal to it) ÷ number of reference months. Ties share half credit; the latest month counts as a tie with itself when inside the window, so scores run from about 8 to 92 with six months.

Combining: Attention = equal-weight mean of eligible source percentiles. Equal weights are a transparent design choice, not an empirically validated one. At least 2 sources must qualify and all must share the same latest complete month; otherwise the score is unavailable and the reason is shown.

Source index = 100 × month value ÷ that source’s reference-window average. Activity index = equal-weight mean of source indexes for the same source set. Change = 100 × (index this month − index last month) ÷ index last month. Both months must be complete under identical sources, panel versions and Trends batches. A missing or zero previous month gives “Unavailable”. We do not report a change in the 0–100 percentile score because it is bounded and would distort size.

Tone = 100 × (positive + 0.5 × neutral) ÷ N, where N = decisive positive + neutral + negative headlines. Headlines must pass relevance rules relevance-v4 and are deduplicated by publisher URL and headline. Scoring uses word lists (tone-v2, AFINN-165 with overrides); mixed or unscorable headlines are excluded and counted separately.

Minimum coverage: at least 20 usable headlines from at least 3 distinct publishers, and a successful news collection within 3 days. This is an experimental coverage rule, not a confidence level. The automatic word scoring has not been validated by human reviewers; earlier AI-assisted review is not proof of accuracy. A change versus the prior 30 days is shown only when both samples meet the minimum under the same rule versions.

Version forecast-v1 (2026-10-08). Models: last month’s value (baseline), 3-month average, simple exponential smoothing (alpha fitted only on each fold’s training months). Rolling-origin backtest: each fold trains on all earlier months (at least 12) and predicts the next one; every model uses the same folds. Score: mean absolute error in score points. A more complex model is picked only if it beats every simpler one by at least 2%; otherwise the baseline is used.

The shaded historical-error range is the forecast ± the 80th percentile of past absolute errors, clipped to 0–100. Its future coverage has not been validated. Model selection and error measurement use the same retrospective folds, and historical values come from a later normalized download, so accuracy may look better than it will in practice. Seasonal effects and unexpected events may cause larger errors.

Forecasts are saved as dated repository records by the manual import workflow, not recalculated on page load. A rerun preserves the original prediction for each brand, method and target month. A monthly export update is required to produce a new outlook. Past forecasts remain visible as archived records. New Google Trends downloads can change scale and revise history: forecast-versus-actual accuracy is unavailable until a comparable-scale source or a separately validated calibration is available. Historical tests are never labeled as issued forecasts.