
What Drives Bubbles in the NFT Market? A Feature Analysis Based on the SHAP Method
Mingjun GUO, Siran FANG, Yunjie WEI
China Journal of Econometrics ›› 2025, Vol. 5 ›› Issue (2) : 463-489.
What Drives Bubbles in the NFT Market? A Feature Analysis Based on the SHAP Method
Currently, the non-fungible token (NFT) market is experiencing significant price volatility. This study aims to detect the bubble phenomena in the NFT market and analyze the key features and mechanisms affecting NFT price bubbles. Firstly, the study focuses on the NFT market and three important sub-markets, utilizing the Generalized Supremum Augmented Dickey-Fuller (GSADF) test to identify the occurrence, duration, and dissipation of NFT price bubbles. Secondly, traditional financial asset prices, market sentiment indices, and cryptocurrency prices are incorporated as features to analyze NFT price bubbles using multiple decision tree machine learning models. Finally, the SHapley Additive exPlanation (SHAP) method is employed to visualize the mechanisms influencing NFT price bubbles. Empirical results indicate that there were five instances of bubbles in the NFT market during the observation period, with a significant increase in the duration of bubbles across sub-markets in 2021. Among the three machine learning models, the CatBoost (Categorical Boosting) model demonstrated the best performance in fitting NFT price bubbles. SHAP analysis revealed that gold, the US Dollar Index, and crude oil prices significantly impact bubble formation, whereas the S&P 500 has a relatively weak influence. Additionally, market sentiment indices such as the Chicago Board Options Exchange Volatility Index (VIX) and Google Trend show opposite trends in their influence on bubbles. By incorporating multiple features, this study enhances market participants' understanding of NFT price bubble characteristics and provides datadriven market insights.
NFT market / rational bubbles / bubble prediction / SHAP analysis {{custom_keyword}} /
表1 描述性统计 |
分类 | 指标 | 样本量 | 均值 | 标准差 | 最小值 | 中位数 | 最大值 |
NFT价格 | All Segment | 1580 | 420.72 | 631.96 | 5.55 | 42.76 | 3608.81 |
Art | 1580 | 1190.26 | 2069.50 | 1.28 | 160.05 | 17838.57 | |
Collectible | 1580 | 853.09 | 1441.91 | 1.44 | 17.46 | 8694.45 | |
Game | 1580 | 157.45 | 266.33 | 4.45 | 40.00 | 2588.89 | |
传统金融资产价格 | WTI | 1580 | 62.89 | 20.51 | 61.11 | 123.70 | |
GOLD | 1580 | 1615.03 | 252.81 | 1184.00 | 1721.90 | 2069.40 | |
SP-500 | 1580 | 3446.81 | 564.30 | 2237.40 | 3446.81 | 4796.56 | |
USDindex | 1580 | 95.57 | 2.93 | 88.71 | 95.57 | 108.03 | |
市场情绪指数 | VIX | 1580 | 21.16 | 8.66 | 10.85 | 19.15 | 82.69 |
GEPU | 1580 | 264.71 | 55.84 | 162.13 | 263.61 | 437.24 | |
Google Trend | 1580 | 10.22 | 19.38 | 0.00 | 0.00 | 100.00 | |
加密货币价格 | Bitcoin | 1580 | 21018.02 | 18054.03 | 3228.70 | 10309.15 | 67527.90 |
Ethereum | 1580 | 1139.34 | 1282.78 | 85.20 | 388.57 | 4807.30 | |
Ripple | 1580 | 0.50 | 0.31 | 0.14 | 0.37 | 1.84 | |
Litecoin | 1580 | 97.98 | 60.16 | 23.46 | 75.47 | 386.45 | |
NEM | 1580 | 0.13 | 0.11 | 0.03 | 0.10 | 0.80 | |
Dash | 1580 | 136.28 | 84.40 | 39.87 | 107.66 | 530.13 | |
XLM | 1580 | 0.19 | 0.13 | 0.03 | 0.14 | 0.73 |
表2 数据集平衡处理结果频数表 |
变量 | 选项 | 处理前 | 处理后 | |||
频数 | 百分比(%) | 频数 | 百分比(%) | |||
AllSegDummy | 0 | 1527 | 96.65 | 1527 | 50 | |
1 | 53 | 3.35 | 1527 | 50 | ||
ArtDummy | 0 | 1549 | 98.04 | 1549 | 50 | |
1 | 31 | 1.96 | 1549 | 50 | ||
CollectibleDummy | 0 | 1527 | 96.65 | 1527 | 50 | |
1 | 53 | 3.35 | 1527 | 50 | ||
GameDummy | 0 | 1537 | 97.28 | 1537 | 50 | |
1 | 43 | 2.72 | 1537 | 50 |
表3 模型参数设置 |
模型 | 参数名称 | 参数值 |
XGBoost | 学习率(learning_rate) | 0.01 |
最大树深度(max_depth) | 4 | |
树的数量(n_estimators) | 100 | |
最小子节点权重(min_child_weight) | 30 | |
样本子采样(subsample) | 0.6 | |
L1正则化参数(reg_alpha) | 0.1 | |
L2正则化参数(reg_lambda) | 0.7 | |
目标函数(objective) | binary: logistic | |
LightGBM | 学习率(learning_rate) | 0.01 |
最大树深度(max_depth) | 4 | |
树的数量(n_estimators) | 100 | |
最小子节点权重(min_child_weight) | 30 | |
样本子采样(subsample) | 0.6 | |
L1正则化参数(reg_alpha) | 0.1 | |
L2正则化参数(reg_lambda) | 0.7 | |
目标函数(objective) | binary | |
CatBoost | 学习率(learning_rate) | 0.01 |
最大树深度(max_depth) | 4 | |
树的数量(n_estimators) | 100 | |
最小子节点权重(min_child_samples) | 30 | |
样本子采样(subsample) | 0.6 | |
L2正则化参数(l2_leaf_reg) | 0.1 | |
损失函数(loss_function) | logloss | |
特征随机采样比例(rsm) | 0.8 |
表4 NFT市场泡沫检测结果 |
NFT市场 | 起始日期 | 结束日期 | 持续时长(天) | |
AllSeg | 1 | 2020/7/25 | 2020/8/4 | 11 |
2 | 2020/8/14 | 2020/8/20 | 7 | |
3 | 2020/9/22 | 2020/10/1 | 10 | |
4 | 2021/1/31 | 2021/2/11 | 12 | |
5 | 2021/2/13 | 2021/2/25 | 13 | |
Art | 1 | 2021/3/9 | 2021/3/18 | 10 |
2 | 2021/8/20 | 2021/9/9 | 21 | |
Collectible | 1 | 2020/9/11 | 2020/9/21 | 11 |
2 | 2020/9/22 | 2020/10/1 | 10 | |
3 | 2021/1/31 | 2021/2/11 | 12 | |
4 | 2021/2/13 | 2021/2/25 | 13 | |
5 | 2021/7/27 | 2021/8/2 | 7 | |
Game | 1 | 2019/9/7 | 2019/9/16 | 10 |
2 | 2021/1/26 | 2021/2/11 | 17 | |
3 | 2021/8/12 | 2021/8/21 | 10 | |
4 | 2022/6/26 | 2022/7/1 | 6 |
注: 本表显示了至少持续5天的泡沫情况. |
表5 机器学习模型对NFT价格泡沫分类效果对比表 |
指标 | 模型 | Allseg | Art | Collectible | Game | |||||||
训练集 | 测试集 | 训练集 | 测试集 | 训练集 | 测试集 | 训练集 | 测试集 | |||||
准确率 | XGBoost | 0.949 | 0.955 | 0.991 | 0.985 | 0.963 | 0.964 | 0.919 | 0.934 | |||
LightGBM | 0.958 | 0.961 | 0.992 | 0.990 | 0.965 | 0.977 | 0.927 | 0.937 | ||||
CatBoost | 0.948 | 0.953 | 0.995 | 0.990 | 0.959 | 0.971 | 0.940 | 0.953 | ||||
精确率 | XGBoost | 0.914 | 0.922 | 0.983 | 0.975 | 0.945 | 0.945 | 0.867 | 0.884 | |||
LightGBM | 0.931 | 0.932 | 0.985 | 0.983 | 0.936 | 0.956 | 0.897 | 0.912 | ||||
CatBoost | 0.907 | 0.916 | 0.991 | 0.981 | 0.925 | 0.944 | 0.895 | 0.913 | ||||
召回率 | XGBoost | 0.992 | 0.993 | 0.999 | 0.996 | 0.984 | 0.985 | 0.991 | 0.996 | |||
LightGBM | 0.991 | 0.993 | 0.999 | 0.998 | 0.999 | 1.000 | 0.966 | 0.965 | ||||
CatBoost | 0.999 | 0.996 | 1.000 | 1.000 | 1.000 | 1.000 | 0.999 | 1.000 | ||||
XGBoost | 0.951 | 0.956 | 0.991 | 0.985 | 0.964 | 0.964 | 0.925 | 0.937 | ||||
LightGBM | 0.960 | 0.961 | 0.992 | 0.990 | 0.966 | 0.977 | 0.930 | 0.938 | ||||
CatBoost | 0.951 | 0.954 | 0.995 | 0.990 | 0.961 | 0.971 | 0.944 | 0.955 | ||||
AUC值 | XGBoost | - | 0.993 | - | 0.999 | - | 0.998 | - | 0.992 | |||
LightGBM | - | 0.992 | - | 0.999 | - | 0.997 | - | 0.989 | ||||
CatBoost | - | 0.991 | - | 0.999 | - | 0.998 | - | 0.997 | ||||
KS值 | XGBoost | - | 0.950 | - | 0.978 | - | 0.963 | - | 0.955 | |||
LightGBM | - | 0.950 | - | 0.981 | - | 0.974 | - | 0.909 | ||||
CatBoost | - | 0.918 | - | 0.991 | - | 0.987 | - | 0.963 |
表6 NFT市场总体及子市场特征重要性排序对比 |
排序 | Allseg | Art | Collectible | Game |
1 | Gold | Google Trend | Gold | Bitcoin |
2 | USDindex | XLM | Google Trend | Google Trend |
3 | VIX | Dash | NEM | WTI |
4 | WTI | GEPU | GEPU | GEPU |
5 | Google Trend | WTI | USDindex | Gold |
6 | Ethereum | Litecoin | VIX | NEM |
7 | GEPU | Bitcoin | Ripple | SP-500 |
8 | XLM | NEM | Bitcoin | Litecoin |
9 | Ripple | Ethereum | WTI | XLM |
10 | NEM | SP-500 | Ethereum | Ethereum |
11 | Bitcoin | Ripple | Litecoin | Dash |
12 | SP-500 | Gold | XLM | Ripple |
13 | Litecoin | USDindex | Dash | VIX |
14 | Dash | VIX | SP-500 | USDindex |
曹睿, 廖彬, 李敏, 孙睿娜, 基于XGBoost的在线短租市场价格预测及特征分析模型[J]. 数据分析与知识发现, 2021, 5 (6): 51- 65.
{{custom_citation.content}}
{{custom_citation.annotation}}
|
陈曦泽, 贾俊峰, 白玉磊, 郭彤, 杜修力, 基于XGBoost-SHAP的钢管混凝土柱轴向承载力预测模型[J]. 浙江大学学报(工学版), 2023, 57 (6): 1061- 1070.
{{custom_citation.content}}
{{custom_citation.annotation}}
|
方思然, 郭明君, 魏云捷, 加入舆情信息是否可以有效提高汇率预测效果?[J]. 计量经济学报, 2023, 3 (2): 464- 486.
{{custom_citation.content}}
{{custom_citation.annotation}}
|
郭春宁, 元宇宙的艺术生成: 追溯NFT艺术的源头[J]. 中国美术, 2021 (4): 14- 19.
{{custom_citation.content}}
{{custom_citation.annotation}}
|
洪永淼, 汪寿阳, 大数据、机器学习与统计学: 挑战与机遇[J]. 计量经济学报, 2021, 1 (1): 17- 35.
{{custom_citation.content}}
{{custom_citation.annotation}}
|
江哲丰, 彭祝斌, 加密数字艺术产业发展过程中的监管逻辑——基于NFT艺术的快速传播与行业影响研究[J]. 学术论坛, 2021, 44 (4): 122- 132.
{{custom_citation.content}}
{{custom_citation.annotation}}
|
秦蕊, 李娟娟, 王晓, 朱静, 袁勇, 等. NFT: 基于区块链的非同质化通证及其应用[J]. 智能科学与技术学报, 2021, 3 (2): 234- 242.
{{custom_citation.content}}
{{custom_citation.annotation}}
|
苏刚, 数字艺术收藏品及NFT迷局[J]. 中国美术, 2021 (4): 33- 37.
{{custom_citation.content}}
{{custom_citation.annotation}}
|
孙少龙, 魏云捷, 黎建强, 基于在线外汇新闻情感挖掘的汇率预测研究[J]. 计量经济学报, 2022, 2 (2): 441- 464.
{{custom_citation.content}}
{{custom_citation.annotation}}
|
谭本艳, 吴艳, 甘子琪, 中国家庭债务风险测度及其预警研究[J]. 金融发展研究, 2022 (12): 38- 48.
{{custom_citation.content}}
{{custom_citation.annotation}}
|
王功明, NFT艺术品的价值分析和问题探讨[J]. 中国美术, 2021 (4): 38- 43.
{{custom_citation.content}}
{{custom_citation.annotation}}
|
王娟娟, 陈昊, NFT数字藏品价值影响因素分析及评估框架——基于FM矩阵分解算法[J]. 中国资产评估, 2022 (7): 18- 28.
{{custom_citation.content}}
{{custom_citation.annotation}}
|
魏丽婷, 郭艳, 贺梦蛟, 非同质化代币(NFT): 逻辑、应用与趋势展望[J]. 经济研究参考, 2022 (4): 130- 140.
{{custom_citation.content}}
{{custom_citation.annotation}}
|
吴一楷, 金融法维度下非同质化通证的属性研究与监管构建[J]. 上海金融, 2022 (11): 32- 41.
{{custom_citation.content}}
{{custom_citation.annotation}}
|
肖争艳, 陈衎, 陈小亮, 陈彦斌, 通货膨胀影响因素识别——基于机器学习方法的再检验[J]. 统计研究, 2022, 39 (6): 132- 147.
{{custom_citation.content}}
{{custom_citation.annotation}}
|
谢波, 韩梓龙, 非同质化通证(NFT) 洗钱犯罪的生成逻辑与侦查路径探讨[J]. 中国人民公安大学学报(社会科学版), 2023, 39 (5): 22- 32.
{{custom_citation.content}}
{{custom_citation.annotation}}
|
谢勇, 项薇, 季孟忠, 彭俊, 黄益槐, 基于Xgboost和LightGBM算法预测住房月租金的应用分析[J]. 计算机应用与软件, 2019, 36 (09): 151- 155.
{{custom_citation.content}}
{{custom_citation.annotation}}
|
赵子龙, NFT艺术品: 风口还是泡沫?[J]. 艺术市场, 2021 (6): 30- 31.
{{custom_citation.content}}
{{custom_citation.annotation}}
|
{{custom_citation.content}}
{{custom_citation.annotation}}
|
{{custom_citation.content}}
{{custom_citation.annotation}}
|
{{custom_citation.content}}
{{custom_citation.annotation}}
|
{{custom_citation.content}}
{{custom_citation.annotation}}
|
{{custom_citation.content}}
{{custom_citation.annotation}}
|
{{custom_citation.content}}
{{custom_citation.annotation}}
|
{{custom_citation.content}}
{{custom_citation.annotation}}
|
{{custom_citation.content}}
{{custom_citation.annotation}}
|
{{custom_citation.content}}
{{custom_citation.annotation}}
|
{{custom_citation.content}}
{{custom_citation.annotation}}
|
{{custom_citation.content}}
{{custom_citation.annotation}}
|
{{custom_citation.content}}
{{custom_citation.annotation}}
|
{{custom_citation.content}}
{{custom_citation.annotation}}
|
Guo Q, Lei S, Ye Q, Fang Z. 2021. MRC-LSTM: A Hybrid Approach of Multi-scale Residual CNN and LSTM to Predict Bitcoin Price[C]// International Joint Conference on Neural Networks (IJCNN).
{{custom_citation.content}}
{{custom_citation.annotation}}
|
{{custom_citation.content}}
{{custom_citation.annotation}}
|
{{custom_citation.content}}
{{custom_citation.annotation}}
|
{{custom_citation.content}}
{{custom_citation.annotation}}
|
{{custom_citation.content}}
{{custom_citation.annotation}}
|
Lundberg S M, Lee S I. 2017. A Unified Approach to Interpreting Model Predictions[C]// 31st Annual Conference on Neural Information Processing Systems (NIPS), 30: 4765-4774.
{{custom_citation.content}}
{{custom_citation.annotation}}
|
{{custom_citation.content}}
{{custom_citation.annotation}}
|
McNally S, Roche J, Caton S. 2018. Predicting the Price of Bitcoin Using Machine Learning[C]// 26th Euromicro International Conference on Parallel, Distributed, and Network-Based Processing (PDP), 339-343.
{{custom_citation.content}}
{{custom_citation.annotation}}
|
{{custom_citation.content}}
{{custom_citation.annotation}}
|
{{custom_citation.content}}
{{custom_citation.annotation}}
|
{{custom_citation.content}}
{{custom_citation.annotation}}
|
{{custom_citation.content}}
{{custom_citation.annotation}}
|
{{custom_citation.content}}
{{custom_citation.annotation}}
|
{{custom_citation.content}}
{{custom_citation.annotation}}
|
{{custom_citation.content}}
{{custom_citation.annotation}}
|
{{custom_citation.content}}
{{custom_citation.annotation}}
|
{{custom_citation.content}}
{{custom_citation.annotation}}
|
{{custom_citation.content}}
{{custom_citation.annotation}}
|
{{custom_citation.content}}
{{custom_citation.annotation}}
|
{{custom_citation.content}}
{{custom_citation.annotation}}
|
{{custom_ref.label}} |
{{custom_citation.content}}
{{custom_citation.annotation}}
|
/
〈 |
|
〉 |