Hugging Face - global_aluminum_price_daily dataset

One row per trading day, four prices per row, one world benchmark. The Hugging Face - global_aluminum_price_daily dataset is a daily time series of world aluminum prices quoted in US dollars per tonne, organized as two subsets that describe the same series two ways: mart, a wide OHLC table keyed simultaneously by Gregorian and Persian dates, and observations, a long-format table where every row carries its own indicator identity, frequency flag, scope code, unit and compile-time vintage stamp. Each subset holds a single train split of about 3,038 rows — 6,076 in total — running from 6 May 2014 to 28 July 2026.

In Datadory's catalog of 1,744 datasets across 159 viable industries, this record scores 7/10 for quality and sits among 13 cataloged aluminum datasets. Daily granularity is genuinely scarce: exactly 3 of the 1,744 records measure out as true daily series, and 2 of those 3 live in aluminum. It is also one of 48 records in the catalog published under the Hugging Face banner, and the only aluminum one that ships as a typed-column OHLC table carrying our machine-learning tag.

Get a sample of this dataset

Captured during the August 2026 review. The first two rows are the opening pair of the mart subset — the oldest two trading days in the window — and the third is the same first day rendered in the observations subset:

date_greg   period      open      high      low       close
2014-05-06  1393-02-16  2182.75   2205.75   2165.00   2172.75
2014-05-07  1393-02-17  2152.25   2152.25   2146.00   2149.00

indicator_id        period      freq  location_code  value    unit         vintage  source_id
commodity_aluminum  1393-02-16  D     WLD            2172.75  usd_per_ton  1405-04  yahoo

Three things in these rows tell you how the whole dataset behaves. First, every day is dated twice: 2014-05-06 and 1393-02-16 are the same observation, so neither calendar's reporting cycle forces conversion logic on your side. Second, the vintage stamp (1405-04) trails the observation period by twelve Persian years, which identifies it as a compile-time marker on the row rather than a trade date. Third, the breakdown column nests the day's high and low as JSON, keeping the long format lossless next to the wide one. Both shapes quote the same thing: dollars per tonne, world level.

Ten fields carry the verified dictionary — five price-and-date columns on the wide table, five metadata columns on the long one — plus two constant identifiers visible on every observations row. Nothing in the dictionary is nullable in spirit: each field exists to make a row self-describing, which is why the unit travels inside the row instead of living in a README.

The practical grouping: build charts off date_greg, period, open, high, low and close; build pipelines off value, location_code, freq and source_id, joining on dataset_id/indicator_id when you need partition keys.

Coverage is narrow, deep and double-keyed:

• Geography: Global — a single world-level series; the scope code reads WLD on every long-format row.
• Temporal: 6 May 2014 through 28 July 2026, keyed as both Gregorian dates and Persian (Jalali) periods from 1393-02-16 to 1405-05-06.
• Granularity: Daily — one observation per trading day, flagged D in the frequency column.
• Volume: About 3,038 rows per subset, 6,076 overall, held as columnar Parquet files of roughly 88 kB (mart) and 65 kB (observations).
• Unit discipline: Prices in US dollars per tonne, declared in-row through the unit field (usd_per_ton) rather than assumed from context.

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

The series lands however your stack wants it — served from an endpoint, dropped into storage, or synced straight into your database. Name the fields and the cadence when you request the sample; the sample ships first either way.

A twelve-year daily OHLC table earns very different livings depending on who is holding it:

• Investors and quants — genuine OHLC beats a close-only feed for backtest features: overnight gaps from open versus prior close, daily ranges convertible straight into volatility estimates, and candle-level pattern work that a single settlement number cannot support.
• Data scientists and ML engineers — typed float columns, a small enough footprint to iterate on, and a long-format twin ready for group-aware cross-validation and feature-store ingestion with compile order preserved.
• Market researchers — a world dollar-per-tonne reference for cost pass-through narratives, with the range columns supplying uncertainty bands around any single-day callout.
• Procurement and commodity teams — a daily reference curve for contract escalators and supplier-quote sanity checks, where knowing the day's spread matters as much as the close.
• Journalists and academics — citable daily levels whose scope (WLD) and unit (usd_per_ton) travel inside every row, so no caption argues with the chart.
• Developers and builders — two shapes, one series: the wide table for rendering, the long one for ingestion, with zero stitching between them.

Most commodity price sets pick a shape and make you live with it. This one refuses to choose. The wide table is what charting libraries want: six typed columns, one row per day, straight into pandas or Polars. The long table is what warehouses want: identifier, period, frequency, scope, value, unit, breakdown — each row self-describing, so ingestion needs no schema interview. Because every row is dated twice, a team reporting on the Persian fiscal calendar and a team reporting on the Gregorian one can query the identical series natively.

That duality is what earns the record's machine-learning tag in our catalog. Feature engineering happens on the wide table; grouped validation and incremental loading happen on the long one, with the vintage column recording compile order. One series, two idioms, no glue code — which is rarer in commodity data than it should be.

Provenance note — the publisher is Farmaanaa on Hugging Face, and every observations row carries the upstream market-data identifier yahoo in its source_id column. The contract definition behind the world series is the publisher's, not an exchange settlement feed.

Scope note — this is deliberately one world-level series. For country splits, USA Trade Online covers bilateral US aluminum trade by HS code and partner, and the Indian Bureau of Mines covers India's bauxite and aluminium statistics.

Card rail note — this page sits in Datadory's aluminum card rail beside 12 sibling records. The monthly IndexMundi series is the natural baseline pairing, SMM's China panel adds the Chinese spot and futures basis, Trading Economics contributes a longer rolling quote history, and the USGS Mineral Yearbook chapters supply production-side context. Three cadences, one industry, no duplicated schema — ask for whichever combination your model needs.