Every forecast on Allora starts with data. Until now, a builder preparing a model for a forecasting question had to source that data on their own, and there was no single place to see exactly which market a question is scored on. Today we are launching Atlas on Forge, a catalog of 363 official market datasets with a dataset pack for every forecasting question.
What Atlas is
Atlas is Allora’s market data service. It collects market prices around the clock, and Forge reads from it directly. Each dataset is the price history of a single market in one-minute candles. A candle summarises one minute of trading with the open, high, low and close price and the volume traded. New candles arrive every minute, so every dataset grows continuously.
Every dataset in Atlas is official and comes from Allora. Builders cannot upload their own. That means a model trained on Atlas starts from the same data a forecasting question is scored on, rather than from a source the builder has to assemble and reconcile themselves.
What is in the catalog
The first release includes 363 datasets. 288 are perpetual futures markets. Of those, 178 are crypto markets and 110 track other assets, including commodities such as gold, silver and oil, as well as currencies, stock indexes and single stocks. The remaining 75 are crypto prices combined across several exchanges.
The deepest history goes back to December 2017. BTC/USD alone holds about 4.7 million rows.
Finding the right data
Builders can browse and search the full catalog on the Datasets page in Forge, filter it by market pair such as BTC/USD, and sort it by how many forecasting questions use each dataset. Every dataset has its own page showing its date range, row count, columns and a short description.
Dataset packs
Every forecasting question on Forge now has a dataset pack, a starting point for building a model for that question.
A pack has two parts. The core dataset is set by rule and is the dataset whose market matches the question, which is the data the question is scored on. Recommended datasets are added by the Allora team as each pack is reviewed, and each one comes with a short reason for its inclusion.
When a pack is published it receives a numbered version with an exact list of the datasets inside it. That list stays fixed, while the data in each dataset keeps updating. Builders can always see which version of a pack they started from.
At launch there are 36 packs, 15 for mainnet questions and 21 for testnet questions.
Why this matters
Allora’s forecasts come from many independent contributors whose models are measured by accuracy. The network is only as strong as the models its contributors build, and good models start with good data. Atlas removes the slowest part of getting started and makes the data behind each question visible to anyone who wants to see it.
It also supports one of the principles Allora is built on. Intelligence on Allora should be inspectable rather than taken on trust. Publishing the exact data each question is scored on is part of that.
What comes next
This is the first release of Atlas. More datasets and packs will follow, starting with suggested datasets in an upcoming release.
Explore Atlas today at forge.allora.network.
