How Gold Desk is built

1. The live price — and why there are two numbers

The big styled number polls public spot-price sources (goldprice.org, with gold-api.com as backup) every few seconds through Gold Desk's own server function. If both are unreachable, the page falls back to the last successful print stored in your browser keeps that number on screen with a small note saying how old it is, and marks the status badge Delayed instead of Live. The page never blanks a price it has already shown. This is a research overlay, not a bank price and not a broker feed.

An earlier version of this page also showed a small TradingView "same feed" quote under the price, meant to guarantee an exact match with the chart. It didn't load reliably enough to be worth keeping, so it's gone. The number above and the chart next to it come from two different data vendors and will normally sit a few dollars apart — that gap is ordinary vendor spread, the same reason two banks' screens never show identical FX quotes down to the cent. If you want the number that's driving the candles, read it off the chart directly.

The percentage under the big number is labelled with what it measures. When goldprice.org answers, it's that feed's daily change and reads today. If that feed is unreachable, the page falls back to the change since the first price it saw on your device in the last 24 hours and reads since HH:MM. The signal meter describes today's move and the headline mood; it is not a forecast, and neither feeds the ranges.

2. Gold and silver, side by side

Silver (XAG/USD) uses the exact same helper-plus-chart-quote pattern as gold, with its own compact chart. The predictions grid, melt calculator, and tax estimator further down the page are gold-only for now — extending them to silver is a natural next step if it'd be useful.

3. The forecasts: Mavia (15 minutes to 2 hours) and Mavia Daily (1 day to 1 year)

One engine serves the home page boxes and GoldDesk AI, so the two never disagree. Each box and each chat card shows a 68% range first, a 95% range, and a direction and expected price only where the model has earned one.

The models

Mavia forecasts 15, 30, 60 and 120 minutes ahead from 5-minute COMEX futures bars (GC=F for gold, SI=F for silver). Mavia Daily does the same from daily closes since 2000, at 1, 2, 3, 5, 10, 21, 42, 63, 126 and 252 trading days; other horizons are interpolated between the two nearest. Each is an ensemble of three members: ARIMA (order chosen by AIC on training data only), XGBoost and LightGBM (tree count by early stopping, depth chosen on validation). The weights are non-negative least squares on out-of-sample validation forecasts, summing to at most 1. The missing share is an implicit weight on "no change", so a member with no skill gets roughly zero weight. A constant-drift candidate is offered to the weights too; validation currently gives it zero.

How it is tested

Every horizon uses the same layout: the first 40% of history for training only, four walk-forward validation folds, and the last 20% as a holdout that no choice ever sees. Each train/test boundary is purged so no training target overlaps the test window. The holdout is scored once, on non-overlapping forecasts, against "no change" with a Diebold–Mariano test. Each horizon then carries the status it earned:

• ok: beat "no change" significantly, after a Benjamini–Hochberg correction for testing many horizons at once.
• weak: beat it, but plausibly by luck. Shown, labelled low confidence.
• no edge: didn't beat it. No direction and no expected price, only the range around today's price.

Confidence above "low" also requires beating a constant-drift forecast and calling direction better than "gold usually goes up". As of the last retrain that bar isn't met at any horizon, so every horizon reads low confidence. The edges are small (0.04–0.6% better than "no change" on gold), and the labels can flip between retrains. The ranges are the dependable part.

The ranges

The bands are the model's own validation errors, divided by volatility at the time and re-scaled by today's volatility × √(horizon). On the untouched holdout the 68% range held 63–71% of the time and the 95% range 94–95%. When a scheduled US release (an FOMC decision or the jobs report) falls inside a short range, that hour's extra volatility is added: the FOMC hour has run about 4× normal volatility and the jobs-report hour about 1.65×, measured over two years of hourly gold prices. The card says when one is coming.

Two improvements were tested and rejected because they didn't beat the current ranges out of sample:

• Blending in the options market's implied volatility (Cboe GVZ). Its interval scores were better, but its ranges held less often than claimed.
• A time-of-day volatility profile. Its interval scores were worse, because the live volatility already adapts.

Which price it anchors on

Mavia forecasts in COMEX futures dollars. The home page applies the same percentage moves to its own anchor, the futures mid when that quote is under three minutes old, else the spot print, so the boxes line up with the price above them. The chat applies them to the Karachi rate, or to the international price in the market you're in. If Mavia isn't serving yet (too few of today's bars), the boxes fall back to a volatility-only range from the last hour of 1-minute bars, with no direction, and say so. When markets are shut there's no band at all.

What isn't used

Headlines don't move any forecast. The old headline-and-momentum tilt was never walk-forward tested, so it's gone. The headline lean is now only logged next to each forecast, so it can be tested against outcomes after a few weeks before it's ever allowed back. The rupee's own volatility isn't in the Pakistan ranges yet, because no clean history of the USD/PKR rate is available. Real Karachi outcomes are scored instead.

The public track record

Every forecast the home page and the chat show is logged and scored against what actually happened. The scores cover how often each range held, the interval (Winkler) score, the error against "no change", and where outcomes fell in the forecast distribution. See the track record, rolling 30 days.

Costs and limits

Every figure is a mid-market estimate: dealing spreads, commissions and jewellers' premiums aren't included, and a band narrower than a normal dealing spread is marked as noise. A surprise headline can move gold outside the 95% band in seconds. Treat the ranges as a sense of scale, not a trading signal, and not financial advice.

4. The "why" line and headline tags

Headlines are pulled from public RSS feeds (Google News and Yahoo Finance) and matched against five keyword groups — dollar, Fed / rates, geopolitics, inflation / oil, and safe-haven flow. Whichever group appears most often across the last batch of headlines becomes the "dominant theme," and the why-line on the price card combines that theme with the direction of the live print. This is pattern matching on headline text, not editorial judgment — an ambiguous or sarcastic headline can be tagged into the wrong bucket.

5. Pakistan rates (PKR)

There are two panels here, on purpose, because there's no live public feed for the actual local rate — so rather than fake one, the page shows both pieces honestly and separately:

International equivalent (live). Calculated, not quoted: live international spot (USD/oz) is divided down to a per-gram price, converted at the live open-market USD→PKR rate, then multiplied up to a tola (1 tola = 11.6638 g). The FX rate comes from a public exchange-rate API; if it can't be reached, the panel falls back to a fixed rate after 8 seconds and says so. This auto-refreshes every few seconds like the rest of the page.

Local Sarafa Bazar rate (live). This is the actual Sarafa Bazar number, the one Pakistani jewellers trade on, which includes the local premium the calculated panel above deliberately doesn't add. There's no official API, so Gold Desk reads it from public listings (hamariweb.com, falling back to gold.pk) every 20 minutes. Each reading is checked before use: karat ratios in range, and 24K within ±10% of the international equivalent. A source that fails is treated as down, never shown. Expect the two panels to sit close but not identical: the gap between them is the local premium.

Karachi vs. Lahore. There isn't a meaningfully different base rate between Pakistani cities — Sarafa Bazar publishes one national benchmark, and city-to-city spread on the raw metal price is typically well under 0.1%. What actually varies by city and by shop is the making charge (ujrat) on jewellery, not the tola rate itself.

ARY-style bar estimate. ARY and other certified-bar retailers price off the same Sarafa benchmark plus a retail premium; reported figures range from roughly 0% to 3% depending on the source and product, with jewellery running far higher (15–25% making charges). The page uses a 1% midpoint as a labeled estimate, not a live ARY quote.

6. The buy-tax estimate

The calculator applies three rules, in order:

• Item type and karat first. Jewelry is always treated as taxable, at any karat. Among items marked "investment," only 24K bars/rounds and 22K legal-tender coins (the two forms that typically meet a state's purity or legal-tender test) are even eligible for an exemption — 18K and 21K investment pieces are treated as taxable everywhere.
• Then the state. Hawaii, Maine, Maryland, New Mexico, Vermont, Washington, and Washington DC currently have no investment-bullion exemption at all, so the full rate applies regardless of order size. Virginia is also included in that list here, because the most recent reporting available when this page was built indicated its exemption lapsed on July 1, 2026 — this specific exemption has been enacted, repealed, and reinstated more than once, so treat this one as the most likely to be stale and confirm it directly if it affects your order.
• Then the threshold. California ($2,000), Connecticut, Massachusetts, and New York (each $1,000) only exempt a single transaction that clears that dollar line; smaller orders in those states are taxed in full.

The tax rate field defaults to each state's general statutory base rate and is fully editable — it deliberately does not try to guess your county or city add-on, since those vary by ZIP code and change independently of state law. This tool is a general guide built from public reporting at one point in time, not tax advice, and sales-tax treatment of bullion is genuinely one of the more volatile corners of state tax law. Confirm with your state's department of revenue, or a tax professional, before a purchase where the answer matters.