{"id":"c3ac8aeb-8e90-4a2f-879f-3c763f8199b4","arxiv_id":"2411.13603","paper_version":1,"verdict":"REJECT","confidence":"MODERATE","novelty_score":4.0,"correctness_risk":"high","formal_verification":"none","parameter_count":4,"one_line_summary":"The paper introduces a purported full-history Bitcoin transaction network dataset and proposes several decentralization metrics, but the dataset is not actually provided and the analysis has significant methodological gaps.","lead":"This paper describes building a complete Bitcoin transaction network dataset from 2009 to 2024 and uses it to trace how Bitcoin's wealth and influence have become more or less concentrated over time. The dataset link is a placeholder and key methods are underspecified, so the claims cannot currently be checked.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The network-based decentralization profile is not reproducible: the N×M edge inference in §II-C can create spurious dense cliques, and no feasible algorithm for centrality on 1.4B nodes is provided.","rationale":"The paper's strongest claim is the introduction of a full-history Bitcoin transaction network dataset and the network-based decentralization profile. The load-bearing condition for this claim is that the network analyzed is both faithfully constructed and computationally tractable. Section II-C's all-to-all edge construction is a heuristic known in the literature, but the paper does not quantify how much it inflates the graph, nor does it present any sensitivity analysis. More importantly, the paper reports exact-looking centrality percentiles for 1.4 billion nodes without describing any algorithm or approximation; exact betweenness on this scale is infeasible, so the figures are unverifiable. Even the missing dataset link, while a basic completeness failure, would be partially mitigated if the analysis could be independently reconstructed; it cannot be. The asset-distribution phase analysis (Section IV-A) is based on daily top-5000 balance snapshots, which are on-chain facts and less sensitive to the graph heuristic, and that part is qualitatively plausible. But the paper's stated novelty is the network perspective, so the centrality/community results are central. The prediction experiments also contain an internal inconsistency: Section VII claims Asset and Rank features 'consistently enhance' performance, yet Table V shows Reformer and other models getting worse with Asset/Rank features. This overclaiming reinforces the need for careful verification. Given these issues, the reader's REJECT is appropriate; a revised version that releases the data, describes the computational pipeline, adds sensitivity analyses, and corrects the overclaims could be reconsidered.","tokens_in":15243,"tokens_out":6719,"duration_ms":64702,"concrete_test":"Request the authors' code and run their pipeline on an early subgraph (e.g., blocks 0–200,000) to reproduce the centrality time series; then recompute the same metrics on a graph that uses only the previous-output-to-input spending links (i.e., no all-to-all outputs) or that applies the multi-input/change-address heuristic to consolidate addresses. If the betweenness and PageRank trends in Fig. 1 change materially, the conclusions depend on the N×M edge construction. Additionally, verify whether the 'dataset link' at the end of the abstract resolves to a downloadable file; if not, the central contribution is unavailable.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central network analysis rests on an under-specified and potentially distorting graph construction. Section II-C builds a directed edge from every input address to every output address of each transaction (N×M edges), because the blockchain does not record which input pays which output. On a full-history graph with 2.59B inputs and 2.85B outputs, this creates enormously dense bipartite cliques for multi-party transactions; a single exchange transaction can add millions of edges that do not correspond to any actual value flow. Betweenness, closeness, PageRank, community detection (§VI A2), and HHI all derive from this graph. No algorithm is provided for computing these centrality metrics on a 1.4B-node graph; exact Brandes betweenness is computationally infeasible at this scale without approximation or sampling, and the paper gives neither. Thus the centrality evolution in Fig. 1 and the HHI results in Fig. 5 cannot be reproduced or validated, and the qualitative phase story may be an artifact of the edge-inference heuristic. The asset-distribution phases (Fig. 2) are computed from balance snapshots and are more robust, but the paper's headline contribution, a network-based decentralization profile, is not.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The manuscript describes the construction of a Bitcoin transaction network dataset spanning from the genesis block to block 845,652 and uses it to profile Bitcoin's decentralization through centrality metrics, asset-distribution concentration, ranking-stability measures, and Herfindahl-Hirschman Index (HHI) calculations. It further evaluates whether adding these network/decentralization features improves transaction-fee and MVRV-Z forecasting. The paper's three headline claims are that this is the first full-history Bitcoin transaction network dataset, that Bitcoin's decentralization evolved in three distinct phases, and that network properties and decentralization metrics consistently improve prediction accuracy. The raw-data statistics are substantial (about 1.41 billion addresses and 1.01 billion transactions), but the dataset itself is not actually made available, the methodology for computing centrality metrics at this scale is not specified, and the forecasting claim is contradicted by the paper's own Table V.","tokens_in":15412,"tokens_out":6795,"duration_ms":63690,"significance":"If the dataset were supplied and the methodology made reproducible, this could be a genuinely useful longitudinal resource for Bitcoin research: the full-history extraction, the daily top-5000 balance snapshots, the ranking-stability measures, and the attempt to tie decentralization to downstream financial forecasting are all potentially valuable. The paper also deserves credit for addressing decentralization from a network perspective rather than only from mining data. However, in the submitted form the central artifact is unavailable, the network measures cannot be reproduced or validated, and one of the headline conclusions is internally inconsistent. The contribution is therefore not yet verifiable, and the significance of the claimed findings cannot be assessed independently.","major_comments":[{"comment":"The paper's central contribution is a 'full-history BTC transaction network dataset,' but the only access point is the placeholder text 'dataset link' in the abstract. No URL, DOI, checksum, file schema, or repository instructions are provided anywhere in the manuscript. Without a persistent link and a data description, none of the network-property tables, centrality plots, or downstream experiments can be independently verified. A dataset paper must provide a working access mechanism and a clear description of the data format.","section":"Abstract and Section I"},{"comment":"The manuscript reports betweenness, closeness, in-degree, and PageRank on a graph with 1,411,482,182 nodes, but it gives no algorithm or approximation for computing these metrics at this scale. Exact betweenness and closeness are computationally infeasible on such a graph, and the paper states neither sampling fractions, number of PageRank iterations, nor the software implementation. Readers cannot reproduce Fig. 1 or the subsequent dispersion-rate discussion. The authors should specify the exact computation pipeline, including whether centrality is computed on the full graph or on per-window subgraphs, and should make the code available.","section":"Section III-A and Fig. 1"},{"comment":"The N×M edge-inference rule creates a complete directed bipartite connection between every input address and every output address of each transaction. For transactions involving exchanges, mixing services, or other multi-party payments, this can generate very dense cliques that do not correspond to any actual value flow. Because all centrality results in Fig. 1, the community detection in Section VI, and the HHI computations in Fig. 5 are based on this inferred graph, the reported decentralization phases may be artifacts of the edge-inference heuristic. The paper needs to justify this construction with robustness checks, such as comparing against subgraphs that exclude multi-input/multi-output transactions or using an amount-proportional edge-elimination heuristic, and should explicitly discuss the known limitations of all-to-all input-output edges.","section":"Section II-C and Section VI"},{"comment":"The text states that adding Asset and Rank features 'improves the performance of models' and that incorporating both 'consistently enhances prediction accuracy across all models.' Table V contradicts this claim: for Reformer, the MSE worsens from 0.252 with Centrality only to 0.437 with Centrality+Asset, and to 0.308 with Centrality+Asset+Rank; for Transformer, Centrality+Rank (0.297) is worse than Centrality only (0.283). The conclusion must be reworded to identify the specific models and settings in which each feature combination helps, and the blanket 'consistently enhances' claim should be removed or replaced by a proper statistical comparison of the relevant configurations.","section":"Section VII-B, Tables III and V"},{"comment":"The three phases in Section IV-A are defined by visual inspection of the top-x proportion curves in Fig. 2, and the decentralization degree D_A in Section IV-B is computed from the same asset-distribution curves. The statement in Section IV-B that D_A's patterns 'align well with the boundaries of the three phases' is therefore not an independent validation of either the phase boundaries or the definition of D_A. An objective procedure, such as changepoint detection on the concentration series with confidence intervals, is needed if the three-phase narrative is to be load-bearing.","section":"Section IV-A and Section IV-B"}],"minor_comments":[{"comment":"The abstract dates the genesis block as '1st March, 2009,' but Table I gives the start date as 01-03-2009 and the Bitcoin genesis block is conventionally dated January 3, 2009; the paper should use the correct date consistently.","section":"Abstract and Table I"},{"comment":"The block count is inconsistent: the abstract refers to the '845651-th block,' while Section II-B and Table I refer to block-845,652 and 'Number of blocks 845,652.' Clarify whether the last included block is 845,651 or 845,652.","section":"Abstract and Section II-B"},{"comment":"The baseline list in Section VII does not include Informer, but Informer appears in Tables II, III, IV, and V. Add Informer to the baseline descriptions or explain why it is treated separately.","section":"Section VII"},{"comment":"The expression 'normalized(HHI A1/A2)' in Eq. (2) is not defined. Specify the normalization procedure for HHI under the address-as-entity (A1) and community-clustering (A2) approaches so that D_HHI is computable.","section":"Section VI, Eq. (2)"},{"comment":"The lower panels of Figures 3 and 4 show breakdowns by top-N group, but the legends are not described in the captions. Add explicit caption text indicating the interval lengths and top-N groups represented in each panel.","section":"Figures 3 and 4"}],"recommendation":"major_revision","confidential_remarks":"To the editor: the manuscript has a potentially valuable dataset and a broad set of analyses, but it is not ready for publication in its current form. The most serious issues are the missing dataset link, the unexplained centrality pipeline on a 1.4-billion-node graph, and the internal contradiction between the forecasting claim and Table V. These are fixable within the scope of a major revision if the authors can provide the dataset, a detailed reproducible methodology, and corrected claims. If the dataset and code are not supplied, the paper's central contribution remains unverifiable and a rejection would be warranted."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Here's the short version: this is a dataset paper whose dataset isn't actually accessible, and the network analysis that gives the paper its title isn't reproducible. The asset-distribution part — daily top-5000 balance shares, the D_A curve, ranking stability — is a reasonable descriptive exercise and could be useful if the data were shipped. But the centrality and community analysis rests on an N×M edge-inference step that is known to create spurious dense cliques in multi-party transactions, and the paper gives no algorithm for computing betweenness, closeness, PageRank, or community detection on a 1.4B-node graph. Exact centrality at that scale is infeasible without approximation or sampling; none is described. So the network-based decentralization profile — the main hook — cannot be reproduced or validated.\n\nWhat's genuinely new is the 15-year coverage through May 2024 with daily snapshots and the three-phase narrative: early decentralization, a re-centralization around 2010–2012, then long-term stability. Prior full-history analyses (Ron–Shamir 2013, Kondor et al. 2014) exist and are cited, but the combination of metrics and the extended time span is new. The forecasting experiments are a reasonable idea, but the text's claim that adding Asset and Rank features 'consistently enhances prediction accuracy' is contradicted by the paper's own Table V: Reformer's MSE worsens from 0.252 to 0.437 when Asset is added, and several other models also get worse. That's not a minor slip; it's a load-bearing overclaim.\n\nThe asset distribution analysis is the healthier part. The balance data are straightforward, and the observed phases are plausible. But the paper's headline contribution is the network-based decentralization profile, and that is exactly the part that is under-specified and potentially distorted by the graph construction. The three-phase story is also defined by visual inspection of the same curves used to compute D_A, so the claimed agreement is not independent evidence.\n\nWho is this for? Someone looking for a descriptive timeline of Bitcoin balance concentration over 15 years might find the figures relevant — but only if the underlying data and code are actually provided. As submitted, the dataset link is literally the text 'dataset link,' and the methods for the central computation are missing. I would desk-reject this version. A resubmission that ships the data, describes the computational pipeline (or uses approximations with error bars), and drops the overclaims would deserve serious referee time.","headline":"A dataset paper with no accessible dataset and an unreproducible network-centrality analysis; only the asset-distribution part has real descriptive value, and the forecasting overclaim is contradicted by the paper's own table.","tokens_in":16015,"tokens_out":5311,"would_cite":false,"duration_ms":44664,"reading_group":"no","serious_thinker":"no","would_accept_peer_review":false},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"Full Bitcoin history maps asset decentralization into three phases","keywords":["Bitcoin","transaction network","decentralization","blockchain dataset","network centralities","Herfindahl-Hirschman Index","time series forecasting","UTXO"],"falsifier":"Recompute the daily top-5000 concentration curves and phase boundaries using only single-input-single-output transactions or a standard change-address heuristic; if the phase boundaries at February 2010 and March 2012 disappear or shift substantially, the three-phase decentralization history is an artifact of the all-to-all edge construction.","tokens_in":14974,"feed_emoji":"📊","tokens_out":6240,"duration_ms":56022,"temperature":0.7,"pith_summary":"The paper's central claim is that a complete Bitcoin transaction-network dataset spanning from the genesis block in January 2009 through block 845,651 in May 2024, with roughly 1.4 billion addresses and 1 billion transactions, can support a systematic longitudinal profile of Bitcoin's asset decentralization. Using this dataset, the authors report that Bitcoin's ownership history separates into three phases: rapid decentralization during 2009-2010, a re-centralization period from early 2010 to early 2012, and a long phase of stable top-holder shares from 2012 to the present. They also introduce several decentralization metrics, including centrality dispersion, an area-based asset-distribution degree, ranking stability measures, and a Herfindahl-Hirschman-based concentration index, and show that adding these network features to forecasting models reduces prediction error for transaction fees and MVRV-Z scores. The payoff for a reader is a shared longitudinal resource and a measurable answer to the question of whether the world's largest cryptocurrency has actually become more decentralized over time.","feed_headline":"Full Bitcoin history maps asset decentralization into three phases","feed_subtitle":"A network of 1.4B addresses from the genesis block to May 2024 shows concentration stabilizing after 2012.","key_machinery":"The load-bearing object is the directed graph built from Bitcoin transactions. Because the blockchain records only which addresses appear as inputs and which as outputs, the paper creates an edge from every input address to every output address of each transaction, with weights proportional to output values; this graph is then used to compute betweenness, closeness, in-degree, and PageRank for 1.4 billion addresses. The other central mechanism is the daily top-5000 ranking list, which feeds an area-based decentralization degree (comparing the real cumulative BTC distribution against an ideal equal distribution), Spearman and retention-rate measures of ranking stability, and a Herfindahl-Hirschman Index computed both per address and per detected community. The downstream experiments feed these features into transformer and recurrent forecasting models.","core_discovery":"On the paper's own terms, the discovery is that decentralization is not monotonic in Bitcoin's history: the network was most evenly distributed in its first year, then concentrated sharply around the Mt. Gox era, and has since plateaued into a stable regime in which the top-5000 addresses hold a roughly constant share of total supply. The evidence comes from daily top-5000 rankings, cumulative distribution curves, and four centrality measures computed on the full transaction graph. The authors further find that address-level concentration (treating each address as an independent holder) is consistently low, while entity-level concentration (clustering addresses into communities) shows a pronounced rise and fall, implying that large players control many addresses rather than one dominant address. These phase boundaries align across the asset-distribution plots and the proposed decentralization degree, which the authors take as mutual support for the three-phase reading.","pith_inferences":["Because the all-to-all edge inference can create spurious connections in multi-input transactions, re-running the three-phase analysis on single-input transactions or with change-address heuristics is a direct robustness check the paper does not report.","The forecasting gains are correlational; a natural extension is to test whether the proposed decentralization degrees lead or lag price and fee regimes, which would indicate whether they carry predictive rather than descriptive information.","The area-based decentralization degree could be extended to arbitrary wealth quantiles or to other cryptocurrencies, allowing direct cross-asset comparisons of how quickly top-holder shares stabilize.","The paper's phase boundaries (early 2010 and early 2012) coincide with known exchange events, but the causal link is not established; an intervention-style analysis around those dates could test whether specific events caused the shifts."],"forward_implications":["The dataset can serve as a common longitudinal benchmark, letting future studies compare network properties over Bitcoin's full history instead of on snapshots.","The three-phase history implies that Bitcoin's ownership concentration has been roughly stable since 2012, so recent price movements occur without major redistribution among top addresses.","Adding decentralization and ranking-stability features consistently lowers forecast error for transaction fees and MVRV-Z across several models, supporting the view that network structure carries financial signal.","The HHI-based distinction between address-level and entity-level concentration shows that measured decentralization depends heavily on how addresses are grouped into owners.","The authors claim the framework transfers to other blockchain assets, enabling cross-chain comparisons of decentralization."],"supporting_citations":[{"why":"Supplies the blockchain parsing method used to extract all blocks and transactions from the BTC Core full node.","marker":"[6]"},{"why":"Provides the graph-primer convention that each transaction becomes directed edges from input addresses to output addresses.","marker":"[7]"},{"why":"Prior empirical analysis of the Bitcoin transaction network that this full-history dataset extends.","marker":"[8]"},{"why":"Earlier full-transaction-graph quantitative analysis whose snapshot approach this paper goes beyond.","marker":"[43]"},{"why":"Earlier quantitative centralization-trend measure in Bitcoin's distributed network, used as comparison in related work.","marker":"[2]"},{"why":"Defines the Herfindahl-Hirschman Index used to construct the market-concentration decentralization degree.","marker":"[12]"},{"why":"Supports the HHI concentration interpretation applied to the address and community groupings.","marker":"[13]"},{"why":"The transformer-based forecasting baseline used in the downstream transaction-fee and MVRV-Z experiments.","marker":"[14]"},{"why":"The recurrent baseline compared against when testing feature combinations.","marker":"[20]"}],"fun_headline_variants":["Bitcoin decentralization: not monotonic, three phases from genesis to 2024","Full-history Bitcoin network shows three-phase asset concentration","Bitcoin's decentralization: early diffusion, Mt. Gox crunch, stable plateau","Fifteen years of Bitcoin transactions reveal non-monotonic decentralization","Entity-level Bitcoin concentration rises and falls; addresses stable"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The analysis depends on the assumption that drawing an edge from every input address to every output address of each transaction faithfully represents how Bitcoin flows, even though the blockchain does not record which input paid which output.","fun_headline_variants_meta":{"raw":{"variants":["Bitcoin decentralization: not monotonic, three phases from genesis to 2024","Full-history Bitcoin network shows three-phase asset concentration","Bitcoin's decentralization: early diffusion, Mt. Gox crunch, stable plateau","Fifteen years of Bitcoin transactions reveal non-monotonic decentralization","Entity-level Bitcoin concentration rises and falls; addresses stable"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000845,"raw_usage":{"total_tokens":3647,"prompt_tokens":885,"completion_tokens":2762,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":501,"completion_tokens_details":{"reasoning_tokens":2673}},"tokens_in":501,"tokens_out":2762,"duration_ms":19841,"temperature":1.0,"reasoning_tokens":2673,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-12T17:32:55.687637+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Recompute the daily top-5000 concentration curves and phase boundaries using only single-input-single-output transactions or a standard change-address heuristic; if the phase boundaries at February 2010 and March 2012 disappear or shift substantially, the three-phase decentralization history is an artifact of the all-to-all edge construction.","supporting_citations":[{"cited_title":"Blocksci: Design and applications of a blockchain analysis platform,","cited_arxiv_id":null,"evidence_quote":"Supplies the blockchain parsing method used to extract all blocks and transactions from the BTC Core full node."},{"cited_title":"Do the rich get richer? an empirical analysis of the bitcoin transaction network,","cited_arxiv_id":null,"evidence_quote":"Prior empirical analysis of the Bitcoin transaction network that this full-history dataset extends."},{"cited_title":"Quantitative analysis of the full bitcoin transaction graph,","cited_arxiv_id":null,"evidence_quote":"Earlier full-transaction-graph quantitative analysis whose snapshot approach this paper goes beyond."},{"cited_title":"Trend of centralization in bitcoin’s distributed network,","cited_arxiv_id":null,"evidence_quote":"Earlier quantitative centralization-trend measure in Bitcoin's distributed network, used as comparison in related work."},{"cited_title":"Similarities and differences between the cr and hhi as an indicator of market concentration and market power,","cited_arxiv_id":null,"evidence_quote":"Defines the Herfindahl-Hirschman Index used to construct the market-concentration decentralization degree."},{"cited_title":"Concentration in us labor markets: Evidence from online vacancy data,","cited_arxiv_id":null,"evidence_quote":"Supports the HHI concentration interpretation applied to the address and community groupings."}],"review_version":1}