Methodology

Search queries (Google, Iran): «خرید طلا», «بیت کوین», «فروش طلا», «وام», «قسط», «گرانی», «خرید خودرو», «قیمت خودرو», «خرید خانه», «اجاره خانه», «استخدام», «کاریابی», «مهاجرت», «ویزا», «آیلتس», «یخچال», «تلویزیون», «لپ تاپ», «ماشین لباسشویی», «گوشی موبایل». Queries are requested in volume-matched groups of five, because Google scales each request to its largest term and a small term next to a large one collapses to an integer 0–3. Market instruments (tgju, daily closing prices): 18k gold, Emami coin, Gold ounce (global), TEPIX (total index), Brent crude.

Collecting Economic news volume: count of ISNA, IRNA and Mehr homepage headlines mentioning each topic; recorded from launch day and excluded from indices until two weeks of data exist. اخبار اشتغال و حقوق: 51 Day, اخبار تورم و قیمت: 51 Day, اخبار طلا و سکه: 51 Day, اخبار مسکن: 51 Day.

• Weekly return = percent change of the week’s last price versus the prior week; 30-day volatility = Standard deviation of daily returns over the past 30 days.
• Durable goods basket = Average search interest for refrigerators, TVs, laptops, washing machines and mobile phones.

For each component in a 52 -week window: z = (value − mean) ÷ SD, times a sign (+ means the component lifts the index). Composite = simple mean of z’s; smoothing with a 3 -week moving average; index value = median percentile of the composite in the same window (0–100; a flat series scores 50). The pulse is built from unsmoothed sub-index composites then smoothed once. Mean and SD use the whole window, so history is redefined each new week (in-sample). Under 20 weeks of data, the index is “Collecting”. Weights are equal; no parameter is fitted.

Validation: for each index, correlation of the composite with subsequent gold weekly returns (leads 1–4) and direction accuracy on the final 40% of data are reported. To date none of the indices has shown meaningful predictive power; they describe the present and are labeled as such.

For each weekly series, this week’s value is compared with the mean and SD of the prior 12 weeks (excluding this week); |z| ≥ 1.5 is reported. For prices we use weekly return, not price level, because a trending level always looks “unusual”; for series that are already percentages (return, premium, volatility) the change is in percentage points, not percent-of-percent. A flat baseline (SD = 0) is not judged. The incomplete current week is dropped from every series so “this week” is a full week for all. Composite indices are omitted here; their moves appear on their own tiles.

For each registered pair: Pearson correlation at lags ±6 weeks (or ±14 trading days); positive lag means the first series leads. p from a t approximation, and p_adj = p × 13 (Bonferroni over all pairs × all searched lags — choosing the best lag is itself a search). “Found” = |r| ≥ 0.2, p_adj < 0.05, n ≥ 30. “Weak” = only raw p significant. “Out-of-sample: better than a coin flip” = lag chosen on the first 60%, repeated with the same sign on the final 40%, and direction accuracy (with relationship sign; for level targets, on the target’s change) beats a coin flip in a binomial test at p < 0.05, with at least 30 observations. Correlation is never reported as causation.

• Google search interest is relative and rescaled on every fetch; low-volume queries have integer resolution.
• Index windows are one year; “strongest week” means strongest within that year.
• tgju prices are daily closing prices with no holiday gaps; a week closes on that week’s last price.
• With 13 pairs tested, without multiple-testing correction one false relationship is expected per run; that is why Bonferroni is applied.
• All indices are experimental and descriptive. No figure on this site is a forecast, and no relationship is causation.

Last calculation: 2026-10-31T20:50 · Week ending ۳ آبان · 6.1 seconds