Kaggle Gold Price Data

It is the most-assigned practice problem in retail gold analytics, delivered as one flat table. Six fields, 2,290 daily rows, running 2 January 2008 through 16 May 2018. Every row sets the closing price of the SPDR Gold Shares ETF (GLD) — the gold proxy — beside four candidate explanatory series: the S&P 500 index (SPX), the United States Oil Fund (USO), the iShares Silver Trust (SLV) and the euro-to-dollar exchange rate (EUR/USD).

The layout is the point. Where most gold feeds hand you a single long price series, this one ships the standard feature matrix: GLD as the target column, everything else as predictors, already time-aligned session by session. That is why it became the default corpus for gold-price regression and machine-learning tutorials — 142 public notebooks grew around it — and why teams still reach for it when they need a benchmark nobody can argue with.

It is also a fixed snapshot rather than a moving feed, and that is a feature more often than it looks. Benchmark numbers computed against a closed window stay comparable year after year, which is precisely what you want when the deliverable is a methodology comparison rather than a trading signal.

Four rows straight from the table — three consecutive sessions from January 2008, then the final row of the series — exactly as they land in a delivery:

date=1/2/2008    spx=1447.160034  gld=84.860001  uso=78.470001  slv=15.18      eur_usd=1.471692
date=1/3/2008    spx=1447.160034  gld=85.570000  uso=78.370003  slv=15.285     eur_usd=1.474491
date=1/4/2008    spx=1411.630005  gld=85.129997  uso=77.309998  slv=15.167     eur_usd=1.475492

date=5/16/2018   spx=2725.780029  gld=122.543800 uso=14.405800  slv=15.454200  eur_usd=1.182033

Read the first and last rows side by side and the decade tells itself. GLD climbs from 84.86 to 122.54 — up 44 percent in the gold proxy. SPX nearly doubles, 1447.16 to 2725.78. USO collapses from 78.47 to 14.41, because the oil leg is where this matrix keeps its drama, while SLV finishes within two percent of where it started and the euro surrenders roughly a fifth of its dollar value. Five instruments, one join key, zero alignment work left for you.

Six fields, one per column, all read directly from delivered records rather than inferred from documentation. The date column keys each trading session; the five numeric series carry the instruments. GLD is the one to handle with care: it is quoted as a fund share price in dollars, not dollars per ounce, so a share-to-ounce convention needs locking before anyone compares these figures against spot quotes.

The dictionary below is the verified core. Anything beyond the six columns is pinned down against delivered records when your sample is cut, not promised blind.

Geography: coverage is instrument-based rather than country-based — four US-listed series (the S&P 500 index plus three exchange-traded funds) and one currency pair. You will find no mine output, reserve estimates or trade flows here; those live in physical-market sources such as USGS Mineral Commodity Summaries - Gold.

Temporal: the window runs 2 January 2008 to 16 May 2018 — 2,290 trading sessions. It opens mid-crisis, carries through the 2011-2012 stretch when gold peaked and the 2013 reversal, and stops cleanly. Because the window is closed, every analysis built on it is reproducible by construction; because it is closed, anything touching current markets needs a live companion feed layered on top.

Granularity: one row per trading session, six fields per row. Daily closes only — no intraday ticks, no weekly aggregates, no imputed holiday rows. Downstream cuts aggregate from the same atomic unit, which keeps two analysts' answers identical.

API, files, or your warehouse. Daily, weekly, or hourly.

Take the table through whichever channel your stack already speaks: REST for point lookups, flat files sized for overnight warehouse loads, or a direct pipe into Snowflake, BigQuery or Redshift. Cadence is yours to set and to change — a closed historical panel like this one ships whole in one pass, and a live companion feed accrues forward from there.

Every delivery comes with the full field dictionary, sample rows for validation, and a schema that stays put between refreshes.

• Data scientists and ML engineers treat it as the shared reference corpus for gold-price prediction exercises. A model trained on these 2,290 rows produces numbers anyone can reproduce, which makes it the lingua franca of forecasting tutorials and interviews alike — 142 public notebooks set the precedent.
• Quant researchers and investors prototype cross-asset ideas on a window whose quirks are already known: gold versus equities through a crisis, gold versus oil across a supply shock, silver's stubborn flatness next to gold's run. Findings survive review because the window never shifts underneath them.
• Journalists, academics and students get a teachable decade in one file — financial crisis, bull market, bust — with clean session-aligned values for coursework, lecture charts and worked examples.
• Developers and data-product builders embed the panel in demos, dashboards and take-home exercises precisely because it is small, coherent and self-explanatory.

Deeper application notes sit on the data scientists, investors quants and journalists academics use case pages.

Inside the gold slice, three neighbours answer three different questions. Stooq - XAUUSD Historical Gold Data owns depth: spot-denominated dollars-per-ounce reaching back centuries, still current to the previous session. SPDR Gold Shares (GLD) Fund Data owns authority: the official fund-level series for the very ETF this table proxies, running daily since November 2004. USGS Mineral Commodity Summaries - Gold owns the physical market: production, recycling and reserves by country, annually.

This dataset occupies the remaining quadrant: the compact multi-asset matrix. Nothing else hands you gold already joined to equities, oil, silver and FX in one session-aligned table. The practical stack is to normalise units against Stooq's spot series, verify recent behaviour against the official GLD feed, and let this file carry the teaching, benchmarking and prototyping load.

Three things, stated up front.

First, this is a closed historical window, not a moving feed: the last row is dated 16 May 2018. Treat it as a benchmark and practice corpus, and pair it with a live gold feed when the question turns to current markets.

Second, mind the gold unit. GLD is a fund share price in dollars, not an ounce quote — the opening row's 84.86 is a share price, not a metal price. Lock the share-to-ounce convention before comparing against spot-denominated series, or ask us to ship the conversion as a derived column.

Third, the schema is six documented columns, verified against delivered records. Anything beyond them gets defined and exemplified when your sample is cut, so nothing ambiguous reaches your pipeline. Name the columns and conventions you need and the sample comes back shaped to them.