Ethereum Transfers

One Parquet table with 24 columns: every native ETH and token transfer on Ethereum, one row each, with the amount, the token's details, the transaction it settled in and the fee. ETH moved by contracts in internal calls is included, and so are NFTs.

About 7.5 billion rows over 134 months, roughly 380 GB as Parquet, estimated from the sample files. The last 24 months hold about 3.3 billion rows and the last 12 months about 2.0 billion. Row counts come from the live table; exact figures arrive with the delivery manifest.

To the start of indexed history: from 7 August 2015, when the first value transfer landed at block 46,147. The Full window runs from there to yesterday.

The export applies no filter of its own: spam tokens, lookalike tokens, zero-value rows and failed transactions all stay in, and the columns let you drop them. Value credited outside a transaction or a contract call has no row: the 2015 genesis allocation, mining rewards and validator withdrawals. Token balances rebuilt from these rows hold, rebasing tokens aside; native ETH balances will not match the chain.

Prices come from our own hourly Ethereum series, which has covered about 8,300 tokens over its life and about 2,900 in the last month. ETH itself is priced from 12 August 2015, so the first five days carry 0. In the published sample 81% of rows have a USD value; the rest are zero-value rows and lookalike tokens. A value of 0 means either a $0 transfer or no price; the column cannot tell them apart.

Decoded from Ethereum by our own indexer, which reads token transfer events and traces internal calls for ETH. Before release we read the sample back: 26,563 rows over blocks 24,053,500 to 24,053,549, every amount carrying exactly its token's decimals. The live table returns 26,552 rows for the same blocks, 11 fewer than the file.

The amount column is decimal adjusted. Multiply it by 10 to the power of the token's decimals for the raw on-chain integer. We ship the adjusted value because decimals differ per token and the raw figure is wrong without them.

Amounts are decimal strings, so casting to float loses precision on large values. Symbols and names repeat, and lookalike tokens copy them, so join on the contract address. Failed transactions are kept; filter TransactionStatus_Success for settled flow. Zero-value rows, 7.8% of the sample, move nothing, so drop them before counting activity. USD values are amount times price with no cap. In the sample, 26 rows of the SORA token (XOR) move amounts in the octillions and are valued above $100 quintillion each, so drop outliers before summing.

Filter Transfer_Currency_SmartContract to the token's contract, lower case with 0x. For native ETH alone, filter Transfer_Currency_Native to true; its contract column holds the string 0x.

One line in DuckDB: SELECT * FROM read_parquet('transfers/*.parquet') LIMIT 10. Pandas works too: pandas.read_parquet(path). The sample file opens without credentials.

Refreshed daily with T+1 latency, so a purchase made today includes everything to yesterday. Delivered as Parquet named by block range, 50 blocks per file and sorted ascending, under a stable S3 layout with a JSON manifest listing every file and its sha256. Signed HTTPS links are emailed once the files are prepared, or on request we deliver into your own S3, GCS or R2 bucket.