PaperlayerView exact revision on arXiv

Source-linked AI summary

Blockchain Lifecycle Prediction - Dead Coins

Uwe A. Kuehn, Syed Muhammad Adnan

arXiv:2610.01379v1cs.DC

TL;DR

Cryptocurrency failure is widespread, but prior work lacks a granular framework for diagnosing lifecycle stages and forecasting transitions. This paper applies binary prediction to market time series and reports discriminative performance alongside gradual value erosion and elevated volatility before inactivity.

  • Problem

    The paper addresses limited stage-based diagnosis of cryptocurrency failure, a concern because dead coins can impose investor losses and impair exits as liquidity disappears.

    Source: §II · p. 1, §I · p. 1

  • Method

    The study uses a binary classifier on daily market data, scoring each asset from a rolling 90-day window with an LSTM.

    Source: A. Data source and Data collection · pp. 5–6

  • Results

    Failing assets exhibit gradual value erosion and elevated volatility before inactivity, and the study reports discriminative performance using two market variables and sequence learning.

    Source: §VI · pp. 8–9

  • Takeaways & Limitations

    Under current data constraints, the paper justifies binary classification as a practical alternative to multi-stage lifecycle modeling.

    Source: §VI · pp. 8–9

    Conclusions are constrained by the small dead-coin sample, single-provider data, coin-level labels, and two-feature input.

    Source: §VI · p. 9

Abstract

from arXiv · show

The cryptocurrency ecosystem has experienced extraordinary growth alongside an equally remarkable rate of failure, with over 52 percent of all tokens launched since 2021 ceasing to trade by early 2025. Despite the scale of this phenomenon, predictive modeling of cryptocurrency death remains an underdeveloped area of research, constrained by definitional ambiguity, data scarcity, and the absence of granular lifecycle frameworks. This work investigates whether the failure of cryptocurrency assets can be predicted using publicly available market data and deep learning methods. A Long Short-Term Memory (LSTM) recurrent neural network was trained on 90-day sequences of daily reference price and estimated market capitalization for 82 cryptocurrency assets (41 alive and 41 dead), sourced from Coin Metrics over the period 2020 to 2026. The model was evaluated using a strictly chronological train-test split to prevent look-ahead bias. The LSTM classifier achieved in best cases a Receiver Operating Characteristic Area Under the Curve (ROC AUC) of 0.98 on the held-out test set. Finally, the model was applied to unseen data, and it was observed that the ROC AUC decreased between 0.59 and 0.65. The findings demonstrate that temporal patterns in price and market capitalization alone contain sufficient discriminative signal to identify assets on a trajectory towards economic inactivity. Diagnostic analyzes reveal that failing assets exhibit gradual value erosion and elevated volatility in the months preceding inactivity, rather than sudden catastrophic collapse. This work also documents the practical infeasibility of a multi-stage lifecycle model under current data conditions and justifies the transition to a binary classification approach.

I. INTRODUCTION

Cryptocurrency failure is widespread, creating financial losses and liquidity problems for investors. The paper identifies a gap in lifecycle-stage diagnosis and presents a six-stage framework alongside a model for alive-to-dead prediction.

I. INTRODUCTION

  • 52.7 percent of nearly seven million GeckoTerminal tokens were no longer trading by early 2025, with 1.8 million failing in the first quarter alone.Source: §I · p. 1
  • Dead coins can cause investor losses and leave holders unable to exit positions when trading volume and liquidity disappear.Source: §I · p. 1

II. PROBLEM AND CONTRIBUTION

  • Prior research addressed death criteria or predicted death probability, but did not distinguish intermediate stagnation and decline stages.Source: §II · pp. 1–2
  • The paper presents a six-stage cryptocurrency lifecycle model and a prediction model for detecting transitions from alive to dead.Source: §II · pp. 1–2

III. BACKGROUND AND RELATED WORK

The background reviews crypto-asset categories and prior work on cryptocurrency longevity. It situates the paper’s lifecycle framework among market-cycle and project-stage accounts.

III. BACKGROUND AND RELATED WORK

  • By 2025, GeckoTerminal had registered almost seven million assets and tokens, and CoinGecko recorded 3.7 million projects ending trading.Source: §III · p. 2
  • Survival analysis linked longevity to project attributes: standalone-chain coins had higher collapse risk, while known headquarters were associated with better survival chances.Source: §III · p. 2

A. Crypto Assets and their Categories

B. 6-staged-lifecycle

  • Industry lifecycle accounts span ideation through listing and post-listing growth, while market-cycle stages connect to the paper’s proposed lifecycle transitions.Source: B. 6-staged-lifecycle · p. 2

C. Dead coin predictive models

The related work describes varied criteria for defining dead coins and factors associated with failure. It motivates sequential deep learning and explains the choice of LSTM for fixed-window prediction.

IV. METHODOLOY

The methodology treats cryptocurrency survival prediction as binary classification over daily market time series. It uses a rolling 90-day window and market variables to produce a dead-class risk score.

  • The model frames cryptocurrency survival prediction as supervised binary prediction on multivariate daily market time series.Source: §IV · p. 4
  • Each asset is scored from its most recent 90-day window, which is recast and rescored as new daily data arrive.Source: §IV · p. 4
  • The design relies on market variables because non-market signals such as developer activity, governance, and community data are often inaccessible for failing assets.Source: §IV · p. 4
  • LSTM is selected to train nonlinear temporal effects while being less susceptible to vanishing gradients than vanilla recurrent networks.Source: §IV · p. 4

A. Data source and Data collection

The study uses Coin Metrics daily market data to build an asset-level dataset of price and estimated market capitalization, then applies within-asset missing-value handling and chronological evaluation. Its cohort and labels are shaped by an inactivity rule, history requirements, and market-cap ranking.

A. Data source and Data collection

  • Coin Metrics supplied daily reference prices and estimated market capitalizations, which were selected for availability across established and marginal assets.The authors note that other predictors may be more effective but are often unavailable or inconsistently defined, and restricting features also reduces feature-based survivorship bias.Source: A. Data source and Data collection · p. 4
  • Short gaps were filled forward and backward within each asset, and rows still containing missing values were removed before sequence construction.This assumes short gaps reflect reporting noise rather than actual economic jumps; remaining incomplete windows were excluded.Source: A. Data source and Data collection · pp. 4–5
  • Raw missingness was 0.63% for reference price and 7.76% for estimated market capitalization.Market-capitalization gaps were more frequent because coverage can lag until circulating-supply estimates become available; price gaps mainly reflect non-trading days or short coverage gaps.Source: A. Data source and Data collection · pp. 4–5
  • The aggregate row-level missingness rate was approximately 4.3%, while structurally sparse assets could be excluded by the inactivity rule and minimum-sequence-length filter.The authors characterize imputation errors as localized because gaps occur within individual asset histories.Source: A. Data source and Data collection · p. 5
  • Assets were labelled dead if they had no Coin Metrics observations during the final 365 days of the study period; otherwise, they were labelled alive.The cohort was limited to up to 100 high-ranked assets per class, and history and sequence-length requirements reduced the prototype cohort to 121 assets.Source: A. Data source and Data collection · p. 5
  • Chronological evaluation placed training windows before 2024-01-01 and windows ending on or after that date in the test set.A stratified split within the training pool supplied validation data; later analyses also tested alternative split methods.Source: A. Data source and Data collection · pp. 5–6

V. RESULTS

The results section introduces examples of experiments that vary data selection, training, and testing details.

V. RESULTS

  • The authors began the reported experiments with a prototype described earlier in the paper.The results section introduces selected examples rather than reporting findings in this passage.Source: §V · p. 6
  • The section presents selected examples of modified experimental setups.The passage does not state outcomes for those examples.Source: §V · p. 6
  • The described setup variations concern data selection, model training, and testing.Source: §V · p. 6

A. Experiments Overview

The experiment overview compares data-splitting and time-span configurations. Because the fixed-date split left some assets short of the history needed for training or prediction, subsequent analyses used percentage-based and optimal-date splits and extended data collection.

A. Experiments Overview

  • The 2024-01-01 fixed-date split left some training or test assets with fewer data points than required for training or prediction.Table III illustrates this behavior for 7 of the 21 dead coins in setup 0602.Source: A. Experiments Overview · p. 6, Table III · p. 6
  • The later analyses switched from a fixed-date split to percentage splits and tested optimal split dates and different train-test ratios.Table IV provides a limited set of samples for setup 0614, which the authors deemed sufficient for presentation.Source: A. Experiments Overview · p. 6, Table IV · p. 7
  • Extending collection from span A, ending 2025-11-30, to span B, ending 2026-05-30, doubled the dead-coin training data and yielded more dead coins for training.Source: A. Experiments Overview · p. 6
  • Table II presents the experimental data setups, including variations in training-to-test ratios and collection spans.The overview describes a fixed date of 2024-01-01 and later percentage-based and optimal-date split experiments.Source: Table II · p. 6, A. Experiments Overview · p. 6

B. On Train and Test Data

The learned models retain a high tp-tn-rate under small changes to the split or data width, while predictions are skipped for assets with less than 90 days of data.

B. On Train and Test Data

C. Unseen Data

On unseen data, the model performed only a little better than random by ROC, with 78 percent true negative predictions. The paper also plots historical input metrics and labels for selected assets.

C. Unseen Data

  • 78 percent true negative predictions accompanied unseen-data ROC performance described as only a little better than random.Table VI reports results for all available crypto assets; Figure 4 presents the ROC for setup 0614.Source: C. Unseen Data · p. 7, Table VI · p. 7, Figure 4 · p. 7
  • Figures 5–6 plot daily time indexes against normalized CapMrktEstUSD and ReferenceRateUSD, with predicted and original labels; the authors do not use them to explain errors.Source: C. Unseen Data · p. 7, Figure 5 · p. 7

D. On Series of a crypto asset

Asset-level prediction curves vary over time, with HBAR’s changing curve highlighted as a possible avenue for further prediction.

D. On Series of a crypto asset

VI. FUTURE WORK AND CONCLUSION

The study reports discriminative value in minimal market features while acknowledging constraints on the evidence. It outlines future work on richer features, time-specific labels, interpretability, and operational deployment.

VI. FUTURE WORK AND CONCLUSION

  • Feature expansion is the immediate future priority, with volume, on-chain, and off-chain signals proposed in tiers where available.The proposed tiers range from price and market capitalization to models with volume and additional network or project signals.Source: §VI · p. 8
  • Horizon-specific labels, such as 30-, 60-, or 90-day inactivity windows, could replace static coin labels and reduce label noise.The paper proposes classification with horizon-specific labels or survival analysis as possible formulations.Source: §VI · p. 8
  • Future work proposes GRU/BiLSTM comparisons, larger multi-provider samples, and post-hoc explanations such as SHAP.These analyses would examine temporal-model performance, generalization across failure modes, and which inputs contribute to death probability.Source: §VI · p. 8
  • Deployment requires further work on retraining, threshold recalibration, and alert aggregation; richer data could support revisiting a hierarchical lifecycle model.Persistence rules are proposed as a way to reduce false positives.Source: §VI · p. 8
  • Two market variables with sequence learning reportedly match or exceed prior studies using richer features or more complex ensembles.The study identifies price and market capitalization as the two market variables.Source: §VI · p. 8
  • Pre-death behavior is characterized as gradual value erosion and elevated volatility, rather than sudden catastrophic collapse.Source: §VI · p. 8
  • Given current data constraints, the multi-stage lifecycle model was practically infeasible, motivating a binary classification approach.Source: §VI · pp. 8–9
  • Conclusions are constrained by the limited number of dead coins, single-provider source, coin-level labels, and restricted two-feature input.Source: §VI · p. 9
Loading 2610.01379v1…