FelNeg/bitcoin-transaction-network-2009-2013

Bitcoin transaction network, 2009-01-03 to 2013-01-08

Transaction networks derived from the public Bitcoin blockchain for the window 2009-01-03 to 2013-01-08 — the exact period analysed by Aspembitova, Feng, Melnikov & Chew (2019), PLoS ONE 14(8): e0219346.

Start here: docs/DATA_AND_PROCESSING.md — provenance, every file explained, schemas, the full pipeline, and the caveats you should read before analysing anything.

Headline numbers

quantity	value
cluster-level edges	25,356,719
transactions	10,886,588
address clusters (co-spend heuristic)	4,910,855 (89.6% singletons)
payee identifiers	8,906,547
self-loops	5,878,290 (23.18%)
date span	2009-01-03 18:15:05 → 2013-01-08 23:52:06

What's in the repository

• runs/full_window/out/ — the processed datasets:
  • user_edges.txt — the primary object: tx_id, user_from, user_to, date(YYYYMMDDHHMMSS), value_btc, 25,356,719 rows
  • addr_edges.parquet — address-level edges with per-output freshness flags
  • pk_map.parquet, pk_user.parquet, tx_map.parquet — identifier maps (payee→identity, payee→cluster, tx→txid)
  • clsize.parquet / clsize.tsv — cluster sizes
  • pairs.parquet, change_pairs.parquet, coinjoin_flags.parquet, collapse_events.parquet, variant_stats.parquet, degree_dist_*.parquet
• runs/full_window/step3/ — clusters_h1.npy, clusters_h2.npy (clustering under two change-address rules)
• runs/full_window/*.log — per-step processing logs
• code/btc_py3/ — the extraction pipeline (Python 3 port of the reference implementation)
• scripts/ — generation and analysis scripts, incl. analysis/verify_manifest.py for checksum verification
• docs/ — DATA_AND_PROCESSING.md, data_manifest.json (sha256 for every artifact), clustering_best_practice.md, data_processing_findings.md

Shared raw layer

data/raw/transactions_2009_2013.txt — 7.9 GB, 47,387,418 lines — is the extraction's input layer: the unclustered per-input/output chain records. It is here so you can apply your own clustering heuristics rather than inheriting ours; everything under runs/ already carries our choices baked in.

It is our reformatting, not a pristine chain dump. It was produced by scripts/gen_transactions_txt.py, which reads the AWS Open Data parquet and re-serialises it into the schema the pipeline expects. Two record types, space-delimited:

in  <own_txid> <referent_txid> <input_index> <identity> <YYYY-MM-DDTHH:MM:SS> <sortkey>
out <own_txid> <output_index> <identity> <value_btc> <YYYY-MM-DDTHH:MM:SS> <sortkey>

<identity> is the raw on-chain identifier: a Base58 address, or the raw public-key hex for early pay-to-pubkey outputs. There is no clustering and no integer ID mapping in this file — those appear only under runs/full_window/out/.

Normalisation choices, all deliberate and all visible in the generator:

• timestamps formatted YYYY-MM-DDTHH:MM:SS — the downstream parser slices fixed character positions, so an ISO T separator is required and a space would silently truncate the time
• value_btc printed with 8 decimals
• coinbase transactions emit only their out lines (the parser skips their referent, which is what turns them into self-loop edges downstream)
• pay-to-pubkey outputs are recovered from the script hex (41…ac / 21…ac) because the upstream ETL leaves address NULL for them; only unresolvable identities become (None)
• window-limited to 2009-01-03 → 2013-01-08

Its sha256 is recorded in docs/data_manifest.json. The 15 GB AWS parquet it derives from is deliberately not deposited — it is byte-for-byte re-downloadable from AWS Open Data, as described in docs/DATA_AND_PROCESSING.md.

Loading the data

import duckdb
con = duckdb.connect()
con.execute("""
    SELECT * FROM read_csv('runs/full_window/out/user_edges.txt',
                           delim='\t', header=false,
                           columns={'tx_id':'BIGINT','user_from':'BIGINT',
                                    'user_to':'BIGINT','date':'VARCHAR',
                                    'value_btc':'DOUBLE'})
    LIMIT 5
""").show()

Read before analysing

• Clusters are evidence-based groups of addresses, not confirmed identities. 89.6% are singletons.
• Applying the change-address heuristics naively collapses the network — the largest cluster grows from 261,808 to 3,315,509 addresses. Use the size-guarded variant; four reporting variants are provided so the effect is auditable.
• The largest cluster (261,808 addresses, containing a May–June 2012 consolidation burst) is plausibly a large service but is not attributed.
• Degree-tail exponents are crude Hill estimates, not MLE fits.
• Public blockchain data, but clustering can in principle aid re-identification: prefer aggregate statistics and do not publish identifiable cluster labels.

Provenance and credit

Derived from public Bitcoin blockchain data (AWS Open Data) by porting and running the pipeline of Brugere et al. (Bitcoin-Transaction-Network-Extraction) with the algorithm unchanged, then extended with an independently validated address-level layer.

Validation: the reconstruction reproduces the reference pipeline's own output exactly — 25,356,719 edges in 24,383,640 (tx, from, to) groups with identical hash sums — and distinct transaction counts match blockchain.com's independent daily series within 5%.

Citation

If you use this dataset, cite the work it supports:

Aspembitova A, Feng L, Melnikov V, Chew LY (2019). Fitness preferential attachment as a driving mechanism in bitcoin transaction network. PLoS ONE 14(8): e0219346. https://doi.org/10.1371/journal.pone.0219346

This repository is the only deposit of the derived networks. If you received a copy passed on from elsewhere, verify it against docs/data_manifest.json.