docs/bitcoin_network_data_notes.md

Bitcoin complex-network study — data-processing notes

Derived from: Aspembitova, Feng, Melnikov & Chew (2019), Fitness preferential attachment as a driving mechanism in bitcoin transaction network, PLoS ONE 14(8): e0219346. Target: extend their framework from 2013 to the 2013–2026 era.

1. What the paper did (must be replicated for comparability)

• Node = user = cluster of addresses (Reid–Harrigan heuristic: all input addresses of a multi-input tx belong to one entity). Source: Brugere's dataset, ref [46].
• Edge = one transaction (directed, weighted by BTC amount), multi-edges retained.
• Periods: 1P = 2009–2010, 2P = 2011–2012. Scale: 11.4M links, 4.7M nodes (2P).
• Windowing: monthly for assortativity / avg degree / clustering (daily "too sparse"); daily used only for size-vs-price.
• Metrics: in/out/total degree + CCDF, power-law exponent (powerlaw pkg; Clauset–Shalizi–Newman MLE+KS), global clustering, avg degree, assortativity, active/inactive/dead/resurrected states, longevity (active months before death), fitness x_{i,n} = Δk_{i,n} / Σ_j Δk_{j,n} (share of new links gained in month n).
• Findings: NO degree PA, NO wealth PA; fitness is power-law (α≈2) and ~constant in time; corr(fitness, degree)=0.47; corr(age, degree/fitness)≈0.02–0.03; links grow ~linearly with nodes (no accelerated growth); generative model (fitness PA + birth/death, 2/3 of new nodes attach to new nodes) reproduces degree distribution but NOT clustering (0.12 vs 0.24).
• Gap: dataset ends 08 Jan 2013. The whole mature era is untouched.

2. Measured properties of the modern data (AWS Open Data, day = 2026-09-01)

Schema per tx: txid, block_hash, block_number, block_timestamp, date, index, size, vsize, lock_time, input_count, output_count, is_coinbase, fee, input_value, output_value, inputs[] = {address, index, value, type, script_hex/asm, sequence, spent_transaction_hash, spent_output_index, txinwitness[], required_signatures}, outputs[] = {address, index, value, type, script_hex/asm, required_signatures}

Day 2026-09-01: 627,345 tx | 1,140,645 inputs | 1,462,070 outputs | 142 coinbase | 72,106 multi-input tx (11.5%) | mean in 1.82, mean out 2.33 | 1,010,472 BTC gross out-value.

issue	measurement	consequence
input address resolution	704 / 1,140,645 missing = 99.94% resolved	clustering heuristic has material
output address resolution	360,057 / 1,462,070 = 24.6% have NO address	OP_RETURN/data-carrier (ordinals) outputs vanish silently in an address-level graph; ~0% in 2013
change / address reuse	449,046 / 627,345 = 71.6% of txs share an address on both sides	after user-clustering most edges become self-loops — must be an explicit policy, report both
multi-input (merging power)	11.5% of txs	only these merge clusters; worth trending over time
coinbase	142/day	flag & exclude or miners are permanent hubs

3. Data-processing decisions to lock before running

1. Entity level: address-level (clean baseline, no heuristic) vs user-level (paper-comparable) vs UTXO-level. Recommend building address-level first, then user-level via clustering.
2. Clustering guards (2013 heuristic breaks in 2026): exchange batching and consolidation create legitimately multi-owner multi-input txs → false mega-clusters; CoinJoin (equal-valued outputs) merges strangers. Need: cluster-size cap, CoinJoin detection, exchange-address lists, and a sensitivity analysis over the merge rule.
3. Self-loop policy: keep vs drop (affects degree tails); report both.
4. No-address outputs: drop, bucket as anonymous, or count as data-carrier activity.
5. Coinbase: exclude from the transaction network; keep for issuance/subsidy series.
6. Weights: BTC amounts are non-stationary in value → use supply-share or USD for wealth attachment; keep raw BTC for volume.
7. Windowing: monthly for structure; daily only for bursts/event studies. Within-block order is miner-determined, not chronological — block_timestamp is the finest honest ordering.
8. Fitness: needs per-window Δk per node + total Δk; zero-activity nodes excluded; requires explicit birth/death/resurrection bookkeeping.
9. Periods: do not inherit 1P/2P. Run data-driven change-point detection (regimes ~2009-10, 2011-13, 2014-16, 2017, 2018-20, 2021-22, 2023+ ordinals).
10. Power-law rigor: MLE + KS + likelihood ratio vs lognormal (powerlaw pkg / CSN); never OLS on log-log.
11. Scale: 1.45B tx / ~2B addresses. Full-graph clustering, path lengths and triangles are infeasible naively → monthly windowed graphs, streaming/approximate algorithms, DuckDB + networkit/graph-tool/igraph. The paper's full graph was 11M edges (≈130× smaller).
12. Reproducibility: paper's dataset is STILL LIVE at https://dynamics.cs.washington.edu/nobackup/bitcoin/bitcoin.tar.gz (679 MB, 2018-01-19) → reproduce first, then extend on AWS parquet (s3://aws-public-blockchain/v1.0/btc/, 2.19 TB total, 146 GB for 2026 YTD).

4. Storage/compute plan

• Paper reproduction: 679 MB (fits anywhere, no GPU).
• 2024→now slice: 561 GB (fits in the 972 GB free).
• Full 2.19 TB history: needs external 4 TB drive or query-in-place over HTTP (DuckDB httpfs, verified working: footer/schema + projection pushdown without full download).