Projecting Non-Fungible Token NFT Collections A Contextual Generative Approach

2025-05-02 0 0 1.09MB 12 页 10玖币

侵权投诉

Projecting Non-Fungible Token (NFT) Collections: A Contextual

Generative Approach

Wesley Joon-Wie Tann, Akhil Vuputuri, Ee-Chien Chang

Department of Computer Science, National University of Singapore

ABSTRACT

Non-fungible tokens (NFTs) are digital assets stored on a blockchain

representing real-world objects such as art or collectibles. An NFT

collection comprises numerous tokens; each token can be trans-

acted multiple times. It is a multibillion-dollar market where the

number of collections has more than doubled in 2022. In this pa-

per, we want to obtain a generative model that, given the early

transactions history (rst quarter

𝑄1

) of a newly minted collection,

generates subsequent transactions (quarters

𝑄2

𝑄3

𝑄4

), where the

generative model is trained using the transaction history of a few

mature collections. The goal is to use the generated transactions to

project the potential market value of this newly minted collection

over the next few quarters. A technical challenge exists in that

dierent collections have diverse characteristics, and the gener-

ative model should generate based on the appropriate “contexts”

of the collection. Our method takes a two-step approach. First, it

employs unsupervised learning on the early transactions to extract

characteristics (which we call contexts) of NFT collections. Next,

it generates future transactions of each token based on these con-

texts and the early transactions, projecting the target collection’s

potential market value. Comprehensive experiments demonstrate

our contextual generative approach’s NFT projection capabilities.

KEYWORDS

Contextual generative modeling, non-fungible token (NFT) collec-

tions, blockchain transactions

ACM Reference Format:

Wesley Joon-Wie Tann, Akhil Vuputuri, Ee-Chien Chang. 2023. Projecting

Non-Fungible Token (NFT) Collections: A Contextual Generative Approach.

In Proceedings of ACM Conference (Conference ’23). ACM, New York, NY,

USA, 12 pages. https://doi.org/10.1145/nnnnnnn.nnnnnnn

1 INTRODUCTION

Non-fungible tokens (NFTs) broke into the mainstream and saw

a boom in the past year [

]. The explosion in the popularity of

NFT collections in the digital art space has garnered much interest

among artists, private and institutional investors, and art galleries.

While some collections are highly valuable, many others remain

of little worth. As an immense number of artworks in the form

of NFTs ood the market, it is dicult to determine the potential

Permission to make digital or hard copies of all or part of this work for personal or

classroom use is granted without fee provided that copies are not made or distributed

for prot or commercial advantage and that copies bear this notice and the full citation

on the rst page. Copyrights for components of this work owned by others than ACM

must be honored. Abstracting with credit is permitted. To copy otherwise, or republish,

to post on servers or to redistribute to lists, requires prior specic permission and/or a

fee. Request permissions from permissions@acm.org.

Conference ’23, February 2023,

ACM ISBN 978-x-xxxx-xxxx-x/YY/MM. . . $15.00

https://doi.org/10.1145/nnnnnnn.nnnnnnn

#2087 (ETH 769) #3749 (ETH 740) #4580 (ETH 666)

#7608 (ETH 0.069) #1470 (ETH 0.4) #5037 (ETH 0.19)

Figure 1: The top NFT tokens of two collections with con-

trasting values, measured by market capitalization. While

the Bored Ape Yacht Club (BAYC) collection (upper row) con-

sists of top transactions of around ETH 700 ≃USD 1M, the

top transactions of the NEKO collection (lower row) are usu-

ally less than ETH 1 ≃USD 1550. The reported last sale price

of tokens (accurate as of Jan 20, 2023).

value of any particular NFT token or collection. Projecting its future

value is very dicult [

]. Nevertheless, major art galleries and

investment institutions continue to pour billions into the sector [

], completely changing the landscape of ne art and traditional

investments.

An increasingly large number of new NFT collections are enter-

ing the market [

]. While each collection has a distinct artistic

theme, the tokens of any particular collection can look very similar

(see Figure 1). We wonder if there is an eective method that allows

us to leverage the dierent characteristics of various established

collections to generate transactions of new NFTs.

Given a newly minted NFT collection that has been in the mar-

ket for a few months, our contextual generative approach aims to

generate its future transactions, allowing us to project the potential

market value. However, we observe that because the NFT market

behaves similarly to a limited collectible market, transactions of

each token are few and far between; this peculiar characteristic of

NFT transactions is challenging for machine learning. Mainly, mod-

els trained with dierent collections result in disparate outcomes,

while a model trained on multiple collections merely “averages” the

results. Hence, we take a two-step contextual generative approach

where

(1)

the semantic context of each NFT collection is rst distilled

and extracted, then

arXiv:2210.15493v2 [q-fin.CP] 4 Feb 2023

Conference ’23, February 2023, Tann et al.

(2)

this auxiliary contextual information guides our generative

method

to generate future transaction data, resulting in contextualized

projections.

Since there are thousands of tokens in any particular NFT col-

lection, these aggregate transactions capture rich market supply

and demand information, which can help us better understand the

economics of NFTs. Never before have such alternative investment

data been so publicly and readily available. Consequently, we ask

this meaningful question:

Can we eectively distill contextual information from var-

ious NFT collections and leverage them to generate repre-

sentative transaction data of unobserved NFTs, reecting

the potential value of new collections?

If we can reasonably answer this question, embracing the recent

advances in deep conditional machine learning, we would have

the means of addressing this challenging market projection puzzle

using nancial data that was once closely guarded but has now be-

come widely available. Moreover, it is a markedly novel investment

class that provides insights into the world of alternative invest-

ments, which was once the exclusive domain of the privileged.

Existing models for forecasting the value of digital collectibles

recorded in the decentralized public ledger, particularly NFTs, are

mainly limited to two categories. Currently, these models are pre-

dominately based on either (1) traditional nancial asset pricing

methods [

] or (2) approaches studying external factors

(e.g., the underlying cryptocurrencies, Twitter inuence, visual fea-

tures) that aect NFT prices [

]. While the traditional

nancial pricing approach performs empirical analysis such as

statistical analysis and economic forecasts on standard nancial

indicators, analyses of external factors focus on how much of a cor-

relation exists between NFT prices and the studied factors. However,

these approaches do not fundamentally consider the underlying

structure of NFT transactions, supply and demand relationships,

and how it impacts NFT prices.

In contrast, we directly address the question of whether market

behaviors captured in transaction series indicate NFT prices. Our

proposed approach leverages conditional generative models [

] and LSTM networks [

] for future transaction series genera-

tion. The rst two stages of development of NFT collections can

be broadly classied as the early stage (initial 3 months) and the

growth stage (next 9 months). First, we analyze each NFT in its

early stage. Then, constructing a context vector

𝐶𝑖

through unsu-

pervised learning for each collection

𝑖

based on its rst quarter

(

𝑄1

) transaction series

𝑇𝑖

provides some particular context of the

collection. The transaction series

𝑇𝑖

consists of the daily values and

transaction count of each token in the collection.

Next, employing these contexts as conditions, it is concatenated

with the transaction series and fed as inputs

(𝐶𝑖,𝑇𝑖)

to our model.

The model then learns from the contexts and transactions of es-

tablished NFT collections. Given a new NFT collection in its early

stage, it eectively generates future transaction series of the new

collection. Lastly, we perform a step-transform procedure on the

generated series, following the piecewise constant series that char-

acterizes each token transaction series. Such a generative modeling

approach produces unobserved future series (See Section 3.3 for

details).

Our experiments are performed on real-world collections in the

NFT market. We used transaction data from ve collections for

training and evaluated the proposed generative approach on ve

other collections (see Tables 3 and 4). In the experiments, we rst

train our model on the ve training collections of various total

market capitalizations. Then, from each collection, we construct

transaction series of size equal to the number of tokens and length

of 365 to reect the early and growth stages. Every collection has its

associated context vector that is characteristic of its transactions in

the early stage (see Figure 5). For example, for the training collection

BAYC, there are 10,000 tokens. The transaction series, size

[

000

365

]

, will have its 6-dimensional context vector concatenated

at the start to make up the inputs to the model. Therefore, the

model learns the transaction characteristics and series of various

collections at dierent developmental stages.

Next, given a new NFT collection in its rst quarter, the model

derives its context through the unsupervised learning method and

uses it to generate the series for the next few quarters of this new

collection. While we use the PCA method [

] here, any other unsu-

pervised learning technique is applicable. Empirical results show

that our approach can accurately generate future series, thereby

projecting the growth of NFTs (see Table 3). Furthermore, we set

up baseline models trained on individual training collections and

an aggregate unconditional model that does not consider the con-

text information. The results show that our approach signicantly

outperforms the baseline comparisons, demonstrating the power

of contextual generative modeling NFT transaction series.

Contribution.

(1)

We identify the signicance of numerous NFT collections

in the market, each with dierent characteristics, and their

corresponding transaction series to estimate the potential

value of such alternative digital assets.

(2)

We introduce a two-step contextual generative approach

that directly leverages the diverse characteristics of NFT

collections to generate future transactions.

(a)

First, using unsupervised learning, we distill key contex-

tual information from early transactions (

𝑄1

) of various

collections.

(b)

Second, given these contexts and early transactions, the

approach generates future transactions (

𝑄2

𝑄3

𝑄4

) of to-

kens over time, projecting the target collection’s potential

market value.

(3)

We present experimental results on real-world NFT collec-

tions to support our proposed approach. Our analysis demon-

strates that the approach is able to generate future trans-

actions that project NFT growth, outperforming baseline

methods.

2 BACKGROUND AND RELATED WORK

In this section, we explain the concept of non-fungible tokens and

discuss NFTs in the domain of blockchains and cryptocurrencies.

Next, we describe the NFT transaction series and valuation methods.

Finally, we investigate existing contextual deep generative models.

Conference ’23, February 2023,

Neko: rst-year growth BAYC: rst-year growth

Figure 2: Mean and variance of Neko (left) and BAYC (right) collections in their rst year, where the daily transaction values

(Y-axis) are plotted against the number of days (X-axis). As shown, the early stages are indicative of potential value. The mean

values of Neko plateau around ETH 0.22 in the growth stage with a converging variance, while BAYC continues to grow with

increasing variance.

2.1 Non-Fungible Tokens

Non-fungible tokens (NFTs) are digital assets that exist in smart

contracts of the Ethereum [

] blockchain. It can represent real-

world objects such as art, music, in-game items and videos, col-

lectibles, and even real estate. NFT diers from standard cryptocur-

rencies [

] in its fundamental property. While a cryptocurrency

(e.g., Bitcoin [

]) is identical to another, therefore serving as a

medium of exchange, NFTs are uniquely identiable. Specically,

an NFT is a unit of data stored on a blockchain that certies digital

representations of physical assets to be unique. Initially introduced

in the ERC-721 standard [

] for representing ownership of non-

fungible tokens, that is, where each token is unique. Besides pro-

viding a form of digital ownership certication, an NFT introduces

the proof of assets right from inception [19].

Artists and content creators alike can now easily prove their

ownership of digital assets. Although, in essence, NFTs represent

little more than code, the codes to a buyer have ascribed value when

considering its comparative scarcity as a digital object. It secures

the selling prices of these intellectual properties that may have

seemed unthinkable for non-fungible virtual assets [

]. Originally

NFTs were part of the Ethereum blockchain, but increasingly more

blockchains have implemented their versions of NFTs [16].

2.2 Transaction Series

NFT marketplace activities can be categorized into four types: List-

ings, Transfers, Bids, and Sales

. A listing is created when an owner

of a wallet containing the NFT makes it available for sale in the

market. An NFT transfer is an activity that enables an easy way of

transferring tokens to another wallet, which could belong to friends,

fellow community members, or perhaps just another wallet of the

same owner. As NFT marketplaces resemble ne art auction houses,

they operate live bidding auctions where people bid for any partic-

ular NFT of their interest, and the highest bid wins. Finally, a sale is

Further descriptions of various activity types can be found at https://support.opensea.

io/hc/en-us

an actual transaction between the buyer and seller. By taking these

transactions of each NFT token, we construct transaction series that

capture meaningful price behaviors among tokens. For example

(see Figure 2), the early stage transactions of the BAYC collection

distinctly vary from the NEKO collection, which is indicative of

their characteristics at a later growth stage.

2.3 NFT Valuation and Pricing

Existing NFT pricing and asset valuation can be broadly categorized

into two branches. In one branch, the approaches [

]

originate from the nance discipline. NFTs are viewed as alternative

investments and studied with either supply-and-demand models

or empirical methods such as regression models and time-series

economic forecasts. The other branch [

] takes a dierent

approach, studying whether NFT prices are aected by external

factors (e.g., the underlying cryptocurrencies, Twitter inuence,

visual features) and how much of a correlation between the factors.

Nadini et al. [

] showed in linear regression analyses that vi-

sual features of NFTs have signicant predictive power on future

primary sale prices, resulting in regression coecients (Adjusted

R-squared

𝑅2

𝑎𝑑 𝑗 ∈ [

]

). Another work [

] explores the re-

latedness of NFT pricing that is driven by cryptocurrencies. Their

spillover index and wavelet coherence analysis indicate some co-

movement between the NFT and cryptocurrency markets with lim-

ited volatility transmission eects, suggesting that cryptocurrency

pricing behaviors might help understand NFT pricing patterns.

As for other works [

], they studied the eects of social

media on NFT pricing. In particular, Twitter and its social media

features were proposed as a predictive factor for NFT prices. The

presented results show that social media features (e.g., count of

user membership lists, number of likes, retweets) have important

predictive value. However, there has not been a study on the eects

of NFT transaction series and their predictive power over the future

valuation and success of new NFTs.

文档加载中……请稍候！
如果长时间未打开，您也可以点击刷新试试。

下载文档到电脑，查找使用更方便

10 玖币 0人已下载

立即下载

摘要：

ProjectingNon-FungibleToken(NFT)Collections:AContextualGenerativeApproachWesleyJoon-WieTann,AkhilVuputuri,Ee-ChienChangDepartmentofComputerScience,NationalUniversityofSingaporeABSTRACTNon-fungibletokens(NFTs)aredigitalassetsstoredonablockchainrepresentingreal-worldobjectssuchasartorcollectibles.AnNF...

展开>> 收起<<

Projecting Non-Fungible Token NFT Collections A Contextual Generative Approach.pdf

共12页,预览3页

还剩页未读，继续阅读

声明：本站为文档C2C交易模式，即用户上传的文档直接被用户下载，本站只是中间服务平台，本站所有文档下载所得的收益归上传人(含作者)所有。玖贝云文库仅提供信息存储空间，仅对用户上传内容的表现方式做保护处理，对上载内容本身不做任何修改或编辑。若文档所含内容侵犯了您的版权或隐私，请立即通知玖贝云文库，我们立即给予删除！

Projecting Non-Fungible Token NFT Collections A Contextual Generative Approach

相关推荐

开通VIP享超值会员特权

作者详情

相关内容

热门标签

举报选择: