Market Phases and Price Discovery in NFTs: A Deep Learning Approach to Digital Asset Valuation

Abstract

This study introduces the Channel-wise Attention with Relative Distance (CARD) model for NFT market prediction, addressing the unique challenges of NFT valuation through a novel deep learning architecture. Analyzing 26,287 h of transaction data across major marketplaces, the model demonstrates superior predictive accuracy compared to conventional approaches, achieving a 33.5% reduction in Mean Absolute Error versus LSTM models, a 29.7% improvement over Transformer architectures, and a 30.1% enhancement compared to LightGBM implementations. For long-term forecasting (720-h horizon), CARD maintains a 35.5% performance advantage over the next best model. Through SHAP-based regime analysis, we identify distinct feature importance patterns across market phases, revealing how liquidity metrics, top trader activity, and royalty dynamics drive valuations in bear, bull, and neutral markets respectively. The findings provide actionable insights for investors while advancing our theoretical understanding of NFT market microstructure and price discovery mechanisms.

1. Introduction

The non-fungible token (NFT) market represents a paradigm shift in digital asset valuation, combining blockchain’s immutability with unique artistic and cultural value propositions. Since surpassing $41 billion in trading volume during its 2021 peak [1], this market has evolved into a complex ecosystem where price discovery mechanisms—defined as the process through which buyer and seller interactions determine asset values—fundamentally differ from traditional financial assets. Three inherent characteristics challenge conventional analytical approaches: the non-fungible nature of assets creates discontinuous valuation landscapes; creator royalty structures introduce perpetual economic feedback loops; and liquidity fragmentation across dozens of independent marketplaces produces asynchronous price movements [2]. In this context, market regimes refer to distinct behavioral phases (bear, bull, neutral) characterized by unique patterns of trading volume, participant activity, and price trajectory, while NFT liquidity describes the ease with which assets can be traded without significant price impact, typically measured through metrics like sales velocity and bid-ask spreads.

Existing models developed for cryptocurrency markets [3] or traditional equities fail to address these unique dynamics, particularly the regime-dependent relationships between platform metrics, creator activity, and price trajectories. This failure stems from three fundamental mismatches. First, the hyperbolic growth and collapse patterns of NFT prices violate the linear stationarity assumptions underlying ARIMA and Prophet models [4]. Second, the interdependency between over 30 observed variables—ranging from aggregator user counts (AGG_UniqueUsers) to blue-chip royalty payments (ROY_BluechipFees)—creates non-linear interactions that standard feature selection methods cannot disentangle. Third, the market’s phase-dependent behavior means key predictors change dramatically between bull, bear, and neutral regimes—a phenomenon poorly handled by static-weight models [5].

To address these challenges, we introduce the Channel-wise Attention with Relative Distance (CARD) model [6]. CARD’s architecture processes market data through parallel temporal channels that separately model minute-level volatility and week-level trends, connected through a novel relative distance encoding mechanism. This approach captures the asymmetric volatility patterns characteristic of NFTs—sharp price drops followed by gradual recoveries—by weighting recent observations exponentially higher than historical data. The model dynamically adjusts feature importance using real-time SHAP values, enabling automatic regime detection and weight redistribution.

Our empirical analysis leverages 26,287 h of transaction data (1 January 2024–31 December 2024) from major marketplaces (OpenSea, Blur) and aggregators (Gem, Genie), encompassing 1.2 million NFTs across 15 categories.

This work makes three principal contributions. Theoretically, it establishes NFTs as a distinct asset class requiring specialized analytical frameworks, moving beyond crypto-centric models and integrating recent advances in decentralized finance (DeFi) liquidity modeling and tokenomics research. Methodologically, CARD’s architecture introduces temporal channel separation and SHAP-guided attention mechanisms that could inform broader blockchain research, particularly in areas requiring dynamic feature importance across market regimes. Practically, the model equips stakeholders with actionable tools: investors gain regime-specific hedging signals, platforms acquire user retention metrics tied to liquidity health, and policymakers receive empirical evidence for royalty regulation. As NFT markets mature toward institutional adoption, CARD’s ability to adapt to phase-specific logic positions it as essential infrastructure for navigating this evolving ecosystem.

The remainder of this paper is organized as follows: Section 2 provides a comprehensive review of the literature, contextualizing our work within both blockchain-general and NFT-specific research. Section 3 details our methodology, including data collection, the CARD model architecture, and experimental design. Section 4 presents our results, comparing model performance and analyzing feature importance across market regimes. Section 5 concludes with implications for various stakeholders and suggestions for future research.

2. Literature Review

2.1. Blockchain and Digital Asset Valuation

Blockchain technology has catalyzed innovation across finance, management, and economics, with implications extending well beyond cryptocurrencies to the fundamental structures of digital asset markets. Research on token economics has provided foundational insights into blockchain-based asset valuation. Sockin and Xiong [7] modeled cryptocurrencies as “utility tokens” that incentivize platform transactions, highlighting how they function as a fundamental medium driving user engagement. Building on this perspective, Cong et al. [8] demonstrated how network externalities influence token price formation, illustrating the mutually reinforcing relationship between user adoption and token valuation. These theoretical frameworks have been extended by studies such as Malinova and Park [9] and Chod and Lyandres [10], which examined the advantages of token issuances compared to conventional financing channels, while Gryglewicz et al. [11] emphasized the crucial role of incentive design for liquidity providers during token launches.

Beyond pure economic models, blockchain governance research has illuminated how decentralized systems manage value creation and distribution. While blockchain is often celebrated for its decentralized nature, Ferreira et al. [12] identified risks of corporate domination within these networks, highlighting how ostensibly decentralized projects may experience undue concentration of power. In response to these concerns, Gan et al. [13] demonstrated how strategic token mechanisms can align incentives between developers and users, suggesting governance structures can be optimized for more equitable participation. Amoussou-Guenou et al. [14] further enriched this understanding by applying game-theoretic analysis to validator behavior in proof-of-stake systems, revealing the economic interdependence inherent in consensus protocols.

Despite blockchain’s transformative potential, significant technical and security challenges affect digital asset valuation. Pagnotta [15] articulated how cryptocurrency price levels and blockchain security form an endogenous feedback loop, where miner incentives determine long-term system resilience. More critically, Malik et al. [16] and Hinzen et al. [17] examined Bitcoin’s scalability limitations and adoption barriers, identifying factors that could constrain blockchain technology’s broader popularization. Biais et al. [18] further highlighted potential inefficiencies through game-theoretic studies of blockchain forks, suggesting excessive computational investment could lead to economically suboptimal outcomes.

Recent research has also addressed market integrity concerns through data-driven approaches. Studies by GRIFFIN and SHAMS [3] and Cong et al. [2] provided evidence of potential price manipulation on cryptocurrency exchanges, emphasizing the importance of regulatory oversight for fair market conditions. Aloosh and Li [19] advanced this line of inquiry by developing methods to detect fictitious trades using insider transaction data, contributing to more robust market surveillance techniques.

Technical innovations have further shaped blockchain market structures in ways that influence asset valuation. Cong et al. [20] and Guasoni et al. [21] examined layer-two scaling solutions on Ethereum and Bitcoin’s Lightning Network respectively, demonstrating how these innovations can reduce on-chain congestion and transaction costs, thereby enhancing the economic value proposition of blockchain systems. In practical application contexts, Cui et al. [22] and Keskin et al. [23] explored blockchain-based traceability solutions that improve supply chain transparency and operational efficiency, while Bakos and Halaburda [24] addressed platform coordination challenges through token-based mechanisms.

The organizational implications of blockchain have been explored by Lee and Lee [25] and Koh et al. [26], who proposed frameworks for contactless service strategies and blockchain-powered business models, demonstrating how distributed ledgers can guide firms toward more digitized operations. Ellinger et al. [27] further investigated decentralized autonomous organizations (DAOs) to propose novel organizational paradigms based on collective action principles.

Most blockchain valuation research has focused on established cryptocurrencies like Bitcoin and Ethereum or examined tokens generically. By contrast, NFTs—which emphasize uniqueness and scarcity—have been comparatively understudied despite their relevance in digital art curation and ownership management [26]. This research gap motivates our study’s focus on NFT market forecasting, aiming to illuminate the distinct factors driving NFT valuation and investor behavior in this specialized domain.

2.2. NFT-Specific Market Analysis

Non-Fungible Tokens (NFTs) represent a distinct asset class within the blockchain ecosystem, characterized by unique properties that challenge conventional valuation frameworks. Recent research has increasingly focused on understanding the specialized market dynamics of NFTs across several key dimensions.

Foundational studies have examined the technological underpinnings and overall landscape of the NFT ecosystem. Nadini et al. [1] systematically analyzed the technological components, security vulnerabilities, and opportunities within NFTs, providing a comprehensive overview of their technical foundation. Nadini et al. [1] expanded this understanding by investigating statistical features, trading networks, and visual characteristics to illuminate the broader structure and evolution of NFT markets. More recently, Taherdoost [28] conducted a systematic review of NFT-related publications from 2012 to 2022, highlighting research trends, developments, and persistent challenges in this rapidly evolving field.

As NFTs gained recognition as investment assets, researchers explored various dimensions of value assessment and investment implications. Ante [29] investigated relationships between NFT sales volume, user activity, and major cryptocurrency prices, demonstrating how NFT markets interact with broader virtual asset classes. Ko et al. [30] evaluated whether NFTs enhance investment portfolio diversification, providing insights into their potential role in modern portfolio theory. Market integrity research by von Wachter et al. [31] examined patterns of illicit trading—particularly wash trading—and estimated discrepancies between actual and reported trading volumes. Complementing this work, Wilkoff and Yildiz [32] analyzed determinants of NFT market liquidity and how new information shapes market fluidity. Looking beyond purely financial metrics, Anselmi and Petrella [33] assessed the informational efficiency of crypto art markets, exploring NFTs’ viability as alternatives to traditional art. Pinto-Gutiérrez et al. [5] identified key popularity drivers, establishing correlations between NFT hype and price movements in major cryptocurrencies.

Behavioral aspects of NFT markets have emerged as another critical research focus. Kapoor et al. [34] connected Twitter data with OpenSea transactions to examine how social media sentiment influences NFT valuations. Taking a multifactor approach, Albayati et al. [35] investigated user behavior and participation intentions in NFT-based metaverse platforms, revealing nuanced motivational and psychological drivers. Zhong and Hamilton [36] further enriched this understanding by examining pricing differentials based on gender and race, uncovering potential biases within NFT markets that parallel inequities in traditional financial systems.

The relationship between NFTs and other asset classes has also received scholarly attention. Das et al. [37] and White et al. [38] analyzed market dynamics, security issues, and network structures within NFT ecosystems, highlighting the unique structural features of these digital marketplaces. Aharon and Demir [39] and Umar et al. [40] examined spillover effects between NFTs and traditional financial assets during the COVID-19 pandemic, demonstrating how NFTs respond to global market volatility. Building on these insights, Wang [41] and Kumar and Padakandla [42] investigated volatility transmission mechanisms, specifically evaluating whether NFTs function as hedge or safe haven assets during geopolitical stress periods like the Russia-Ukraine conflict.

Practical applications of NFTs have been documented across diverse sectors. Vasan et al. [43] quantified networks between artists and collectors, illuminating value creation processes in crypto art markets. In educational contexts, Wu and Liu [44] identified specific instances of NFT integration in pedagogical settings, while Hwang [45] investigated how NFT metaverse exhibitions influence learning outcomes. From a marketing perspective, Chohan and Paschen [46] demonstrated how NFTs enhance promotional campaigns, while Wu et al. [47] analyzed product-level success factors and challenges. Gaming applications were explored by Vidal-Tomás [48] and Bai et al. [49], who examined how NFTs and token economics shape virtual economies in blockchain-based games.

The rapid growth of NFT markets has intensified research on risk management and predictive modeling. Liu et al. [50] employed time-frequency frameworks to analyze extreme risk transmission in carbon-NFT-equity systems, demonstrating how NFT-related shocks propagate under stressed market conditions. Umar et al. [51] applied CoVaR methodology to measure NFT risk-return characteristics and diversification benefits, while Xia et al. [52] explored quantile linkages between NFTs and traditional assets, questioning whether NFTs maintain their distinctiveness during extreme market turbulence. Investigating internal market interdependencies, Urom et al. [53] assessed relationships among trading volumes and returns across NFT submarkets, while Qiao et al. [54] constructed wavelet-based quantile causality networks to analyze extreme risk transitions among cryptocurrencies, DeFi instruments, and NFTs. Wang et al. [55] further contributed by studying CryptoPunks data to identify key factors influencing NFT market performance.

While this body of research provides valuable insights into NFT market dynamics, many studies offer retrospective analyses rather than comprehensive forecasting models [28,56]. Our study addresses this gap by identifying diverse factors shaping NFT market evolution and developing forward-looking predictive frameworks. This approach not only provides actionable insights for investors and policymakers but also addresses the unique methodological challenges of modeling an emergent market with characteristics distinct from traditional financial settings [57,58,59]. By proposing novel forecasting models specifically tailored to NFT markets, our research contributes both theoretical and practical advancements to this rapidly evolving domain.

3. Methodology

3.1. Data

This study examines NFT market data collected from 1 January 2022, to 31 December 2024, through Dune Analytics. Our dataset consists of hourly measurements of the FK500 Index and 29 market indicators across six key categories.

We selected major NFT marketplaces including OpenSea, Blur, X2Y2, and LooksRare, along with aggregators Gem and Genie for three primary reasons. First, these platforms collectively accounted for over 92% of the total Ethereum-based NFT trading volume during the study period, ensuring comprehensive market coverage. Second, these marketplaces implement standardized trading interfaces through ERC-721 and ERC-1155 protocols, enabling consistent cross-platform data collection. Third, these platforms maintain transparent on-chain transaction records, which allows for reliable verification of trading activity.

Data collection was implemented through Dune Analytics, which provides direct access to Ethereum blockchain data through SQL queries. This approach enables access to historical data beyond API limitations, standardized metrics across platforms, and the ability to identify cross-platform user activities. The hourly sampling frequency was selected to balance granularity of price discovery signals with computational feasibility, capturing both intraday volatility patterns and longer-term trends.

The FK500 Index serves as our dependent variable, providing a benchmark for NFT market performance through market capitalization and unique trader counts. It incorporates daily rebalancing to reflect market dynamics across major NFT projects. This index was selected over collection-specific metrics (such as floor prices) because it better reflects market-wide movements, is less susceptible to manipulation through isolated transactions, and incorporates both price and liquidity dimensions of the market.

Our independent variables encompass six primary categories: First, aggregator metrics track platform efficiency through variables like AGG_TotalVolume and AGG_UniqueUsers, measuring transaction volumes and user participation across integrated marketplaces.

Second, trading scale indicators such as TRD_UniqueBuyers and TRD_AvgTxSize capture market participant behavior and transaction characteristics.

Third, marketplace performance metrics including MKT_TotalFees and MKT_VolumeUSD evaluate platform efficiency and liquidity.

Fourth, royalty-related metrics measure ecosystem sustainability through ROY_Bluechip Fees and ROY_PayingUsers, tracking creator revenue streams and premium project performance.

Fifth, top trader activity metrics monitor market leaders’ behavior via TOP_VolumeUSD and TOP_TotalTrades.

Finally, market-wide indicators such as TOT_Sales and TOT_ActiveTraders provide overall market trend analysis.

These variables collectively enable comprehensive analysis of NFT market dynamics and support the development of robust prediction models. For detailed variable descriptions and calculations, please refer to

Table 1. Description of Variables Used in NFT Market Analysis.

Category	Variable	Description
Dependent Variable	FK500 Index	Hourly Forkast 500 NFT index value
Aggregator	AGG TotalVolume	Daily total transaction volume on aggregator platforms
	AGG UniqueUsers	Total number of unique users on aggregator platforms
	AGG NonAggUsers	Number of users directly using marketplaces
	AGG TotalUsers	Total number of aggregator platform users
Trading Scale	TRD TotalSales	Total number and scale of transactions
	TRD UniqueBuyers	Number of unique buyers and average transaction scale
	TRD UniqueSellers	Number of unique sellers and average transaction scale
	TRD AvgTxSize	Average transaction size
	TRD SalesCount	Number of transactions
	TRD AvgTradesPerUser	Average number of trades per user
	TRD EthSalesMA	Moving average of ETH-based sales
Marketplace	MKT TotalFees	Total marketplace fees
	MKT TotalSales	Total marketplace sales
	MKT TotalVolume	Total marketplace trading volume
	MKT VolumeUSD	Total marketplace volume in USD
	MKT TotalUsers	Total number of marketplace users
Royalty	ROY BluechipFees	Total royalties from blue chip NFT projects
	ROY TotalFees	Total royalty fees
	ROY PayingUsers	Number of royalty-paying addresses
	ROY PayingTx	Number of royalty-paying transactions
Top Traders	TOP VolumeUSD	Top traders’ volume in USD
	TOP VolumeETH	Top traders’ volume in ETH
	TOP TotalTrades	Total number of top traders’ transactions
	TOP TotalTraders	Number of top traders
Market Total	TOT Sales	Total NFT market sales
	TOT Volume	Total NFT market volume
	TOT VolumeUSD	Total NFT market volume in USD
	TOT Users	Total NFT market users
	TOT ActiveTraders	Total number of active traders

in

Table 2. Descriptive Statistics of Variables.

Time Period: January 2022–December 2024, Hourly Data (N = 26,287)
Variable	Min	Median	Mean	Max	SD
Dependent Variable
FK500 Index	1098	2787	6657	33,903	7123
Aggregator Metrics
AGG Total Volume	5.0	240.0	365.2	3122.0	421.3
AGG Unique Users	2	78	105	842	98
AGG Non Agg Users	24.0	293.0	670.4	4789.0	842.6
AGG Total Users	29.0	394.0	775.4	4908.0	892.1
Trading Scale Metrics
TRD Total Sales	7.43	367.71	1102.62	133,767.38	3124.5
TRD Unique Buyers	18.0	411.0	815.5	4943.0	892.3
TRD Unique Sellers	18	525	1026	5793	1124
TRD Avg Tx Size	0.0945	1.0173	4.9758	511.5824	12.84
TRD Sales Count	19	828	1509	8704	1682
TRD Avg Trades Per User	1.056	2.046	2.386	45.670	1.524
TRD Eth Sales MA	2.538	64.834	199.748	19,578.668	524.6
Marketplace Metrics
MKT Total Fees	26	5113	43,761	8,183,481	182,643
MKT Total Sales	47	926	1543	8616	1724
MKT Total Volume	9.4	400.8	657.6	24,057.3	842.1
MKT Volume USD	25,509	777,731	1,475,119	57,363,600	2,124,562
MKT Total Users	33.0	461.0	839.8	4939.0	892.4
Royalty Metrics
ROY Bluechip Fees	107	10,618	50,848	2,866,242	124,682
ROY Total Fees	107	10,618	50,848	2,866,242	124,682
ROY Paying Users	14.0	304.0	797.6	5587.0	1124.2
ROY Paying Tx	18	504	1213	8554	1642
Top Trader Metrics
TOP Volume USD	22,505	715,043	1,403,312	59,507,103	2,242,425
TOP Volume ETH	8.975	367.981	622.250	21,214.876	824.6
TOP Total Trades	47	920	1568	8605	1842
TOP Total Traders	70	962	1817	10,056	2124
Market Total Metrics
TOT Sales	47	926	1543	8616	1724
TOT Volume	9.4	400.8	657.6	24,057.3	842.1
TOT Volume USD	25,509	777,731	1,475,119	57,363,600	2,124,562
TOT Users	33.0	461.0	839.8	4939.0	892.4
TOT Active Traders	70.0	917.0	1724.9	9791.0	2042.8

.

3.2. CARD Model

We propose the Channel-Wise Attention with Relative Distance (CARD) model to effectively capture the complex temporal patterns and inter-feature relationships in NFT market data. The model architecture is specifically designed to handle the unique characteristics of NFT market time series, incorporating multiple specialized components for comprehensive pattern recognition.

3.2.1. Problem Formulation

In time series prediction for the NFT market, we define the multivariate time series input at time t as $X_t\in {\mathbb{R}}^F$, where F represents the dimension of input features encompassing various NFT market indicators including aggregator metrics, trading scales, marketplace performance, royalty statistics, and market-wide measures. Given a sequence length L, we aim to utilize the input sequence $\mathcal{X}=\{X_{t-L+1},\dots ,X_t\}\in {\mathbb{R}}^{L\times F}$ to predict the FK500 index values for the next H time steps, denoted as $\mathcal{Y}=\{y_{t+1},\dots ,y_{t+H}\}\in {\mathbb{R}}^H$. The prediction function f can be formally defined as:

$$f:{\mathbb{R}}^{L\times F}\to {\mathbb{R}}^H,\mathrm{where}\widehat{\mathcal{Y}}=f\left(\mathcal{X}\right)$$

(1)

This formulation encapsulates our objective of capturing both short-term fluctuations and long-term dependencies in the NFT market dynamics.

3.2.2. Hierarchical Patch Embedding

To capture local patterns in the time series data, we first segment the input sequence into overlapping patches. Using a patch length P and stride S, we generate N patches where:

$$N=⌊{\displaystyle \frac{L-P}{S}}⌋+1$$

(2)

Each patch undergoes a linear transformation combined with positional encoding to create D-dimensional embeddings:

$$E_i=W_e·X_{i:i+P}+b_e+P_i$$

(3)

where $W_e\in {\mathbb{R}}^{D\times P}$ represents the embedding weights, $b_e\in {\mathbb{R}}^D$ is the bias term, and $P_i\in {\mathbb{R}}^D$ denotes learnable positional encodings. To maintain global context, we introduce a classification (CLS) token $E_{cls}\in {\mathbb{R}}^D$:

$$\mathcal{E}=[E_{cls};E_1;\dots ;E_N]\in {\mathbb{R}}^{(N+1)\times D}$$

(4)

3.2.3. Multi-Head Channel-Wise Attention

For each head h, the attention computation is defined as:

$$Q_h=W_q^{h}E,K_h=W_k^{h}E,V_h=W_v^{h}E$$

(5)

We introduce dynamic projection to efficiently reduce the dimensionality of keys and values:

$$\begin{array}{cc}\hfill K_p^h& =f_k^h\left(K_h\right)\in {\mathbb{R}}^{B\times N\times R}\hfill \end{array}$$

(6)

$$\begin{array}{cc}\hfill V_p^h& =f_v^h\left(V_h\right)\in {\mathbb{R}}^{B\times N\times R}\hfill \end{array}$$

(7)

The attention for each head is computed as:

$$A_h=\mathrm{softmax}\left({\displaystyle \frac{Q_h{\left(K_p^h\right)}^T}{\sqrt{d_k}}}\right)V_p^h$$

(8)

The final channel attention output is obtained by concatenating all head outputs:

$$A_c=\mathrm{Concat}(A_1,\dots ,A_H)W_o$$

(9)

3.2.4. Relative Distance Enhanced Temporal Attention

The EMA-based distance matrix is defined as:

$$D_{i,j}=\alpha {(1-\alpha )}^{|i-j|}$$

(10)

The temporal attention is computed as:

$$\begin{array}{cc}\hfill \tilde{Q}& =Q\odot D,\tilde{K}=K\odot D\hfill \end{array}$$

(11)

$$\begin{array}{cc}\hfill A_t& =\mathrm{softmax}\left({\displaystyle \frac{\tilde{Q}{\tilde{K}}^T}{\sqrt{d_k}}}\right)V\hfill \end{array}$$

(12)

3.2.5. Feature Fusion and Output Layer

The outputs are combined through a gating mechanism:

$$\begin{array}{cc}\hfill g& =\sigma (W_g[A_c;A_t]+b_g)\hfill \end{array}$$

(13)

$$\begin{array}{cc}\hfill F& =g\odot A_c+(1-g)\odot A_t\hfill \end{array}$$

(14)

The final predictions are generated through a two-layer feed-forward network:

$$\begin{array}{cc}\hfill Z& =\mathrm{FFN}\left(F\right)=W_2\left(\mathrm{GELU}(W_{1}F+b_1)\right)+b_2\hfill \end{array}$$

(15)

$$\begin{array}{cc}\hfill \widehat{\mathcal{Y}}& =W_{o}Z+b_o\hfill \end{array}$$

(16)

3.3. Training Strategy

The model is trained to minimize a composite loss function:

$$\mathcal{L}={\mathcal{L}}_{mse}+\lambda {\mathcal{L}}_{reg}$$

(17)

where ${\mathcal{L}}_{mse}$ represents the mean squared error:

$${\mathcal{L}}_{mse}={\displaystyle \frac{1}{H}}\sum _{i=1}^H{(y_{t+i}-{\widehat{y}}_{t+i})}^2$$

(18)

and ${\mathcal{L}}_{reg}$ is the L2 regularization term:

$${\mathcal{L}}_{reg}=\sum _{\theta \in \Theta }{\parallel \theta \parallel }_2^2$$

(19)

The learning rate is scheduled according to:

$${\eta }_t={\eta }_0·min(t^{-0.5},t·\mathrm{warmup}\_{\mathrm{steps}}^{-1.5})$$

(20)

To prevent overfitting and enhance model generalization, we employ dropout with probability p and layer normalization after each attention and feed-forward layer. The model’s hyperparameters are carefully tuned through extensive experimentation to achieve optimal performance on the NFT market prediction task.

3.4. Experiment Design

We evaluate five distinct modeling approaches, each representing different paradigms in time series analysis: The RNN model implements a basic recurrent architecture with three hidden layers (128, 64, 32 units) to capture temporal dependencies in the NFT market data. We apply dropout (0.2) between layers to prevent overfitting. The LSTM model extends the recurrent approach with four layers (256, 128, 64, 32 units), incorporating memory cells to better capture long-term dependencies. We implement bi-directional processing in the first two layers to enhance pattern recognition. The Basic Transformer follows the standard architecture with four attention heads and three encoder layers. We set the embedding dimension to 128 and implement positional encoding to maintain temporal order information. LightGBM utilizes gradient boosting with 1000 trees, learning rate of 0.01, and maximum depth of 8. We implement time-based feature engineering to provide the model with temporal context. Prophet incorporates its standard decomposition approach with daily seasonality enabled and automatic changepoint detection. We configure the model to account for the high-frequency nature of NFT trading data. We employ a time-aware data splitting strategy with a 70/15/15 ratio for training, validation, and testing sets. This approach maintains temporal coherence and prevents future data leakage. Each model undergoes training with early stopping protocols based on validation set performance. We implement identical data windowing (lookback period of 168 h) across all models to ensure fair comparison. Evaluation Framework Model performance is assessed across four prediction horizons: 96, 192, 336, and 720 h. This range enables evaluation of both short-term accuracy and long-term stability. We employ MSE, MAE metrics for comprehensive performance assessment. This experimental design enables systematic comparison of diverse modeling approaches while maintaining methodological consistency and reproducibility.

4. Results

4.1. Comparative Model Summary

In examining different time series prediction models for NFT market analysis, we consider several established approaches, each with distinct characteristics and limitations: RNN (Recurrent Neural Network), pioneered by Rumelhart et al. [60], maintains hidden states to process sequential data. However, the model faces significant challenges with long-term dependencies due to the vanishing gradient problem. This limitation becomes particularly apparent when processing extended time sequences, potentially missing crucial long-term market patterns. LSTM (Long Short-Term Memory), developed by Hochreiter and Schmidhuber [61], addresses some RNN limitations through memory cells and gates. Despite this improvement, LSTMs still process temporal information sequentially, which can create computational bottlenecks and limit their ability to capture global patterns effectively. The model may also struggle with parallel processing of multiple market indicators. The Transformer architecture, introduced Vaswani et al. [62], revolutionized sequence modeling through self-attention mechanisms. However, the standard Transformer’s quadratic computational complexity with sequence length poses scaling challenges. The model also lacks specific mechanisms for handling market-specific temporal dynamics and can struggle with varying sequence lengths. LightGBM, developed Ke et al. [63], excels at handling tabular data through gradient boosting. Despite its efficiency and effectiveness with structured data, the model lacks inherent sequential processing capabilities. It may overlook temporal dependencies and struggle to capture long-term trends in time series data. Prophet, introduced by Taylor and Letham [4], specializes in decomposing time series into trend, seasonality, and holiday components. While effective for regular patterns, this decomposition-based approach may oversimplify complex market interactions. The model’s assumptions about seasonality and trends may not align with the irregular patterns often present in financial markets. Our proposed CARD model addresses these limitations through a specialized architecture combining channel-wise attention with relative distance enhancement. This design enables simultaneous processing of multiple market indicators while maintaining computational efficiency. The model’s hierarchical structure and dynamic projection mechanisms specifically target the unique characteristics of NFT markets, including high volatility, complex feature interactions, and varying temporal dependencies. Through adaptive feature fusion and multi-dimensional attention mechanisms, CARD provides a more comprehensive approach to NFT market prediction compared to traditional methods. Through adaptive feature fusion and multi-dimensional attention mechanisms, CARD provides a more comprehensive approach to NFT market prediction compared to traditional methods. A comparison of the various time series models discussed, including their key characteristics and limitations, is presented in

Table 3. Comparison of Time Series Models for NFT Market Prediction.

Model	Key Characteristics	Limitations
RNN  [60]	Maintains hidden states for sequences Processes data sequentially Learns temporal dependencies	Vanishing gradient problem Limited long-term memory Sequential processing bottleneck
LSTM  [61]	Memory cells and gates Better long-term dependencies Controlled information flow	Sequential processing limitation Computational bottlenecks Limited parallel processing
Transformer  [62]	Self-attention mechanism Parallel processing Global context modeling	Quadratic complexity Lacks market-specific mechanisms Variable sequence length issues
LightGBM  [63]	Gradient boosting framework Efficient with tabular data Strong non-linear modeling	Limited sequential processing Misses temporal dependencies Struggles with long-term trends
Prophet  [4]	Decomposition-based approach Strong seasonal modeling Handles missing data	Oversimplifies complex patterns Rigid seasonality assumptions Limited with irregular patterns

.

4.2. Model Performance Comparison

Our comparative analysis of different time series models for NFT market prediction reveals significant variations in performance across training and test sets. The results demonstrate the relative strengths and limitations of each approach in capturing NFT market dynamics. The CARD model shows consistent performance across both training and test sets, with a test MSE of 16.68, RMSE of 4.08, and MAE of 2.81. The model’s MAPE of 0.24% on the test set indicates high prediction accuracy, demonstrating robust generalization capabilities. This balanced performance between training and test metrics suggests effective model regularization and minimal overfitting. ARIMA, with optimal parameters (p = 2, d = 2, q = 1), achieves comparable test set performance to CARD, showing a test MSE of 16.04, RMSE of 4.01, and MAE of 2.63. The model’s low MAPE of 0.23% indicates strong predictive accuracy, though its simpler structure may limit its ability to capture more complex market patterns. The deep learning approaches show varying degrees of success. The Transformer architecture achieves moderate performance with a test MSE of 16,580.07 and MAPE of 9.75%, suggesting some ability to capture market patterns but with room for improvement. However, both RNN and LSTM models struggle with significant overfitting, as evidenced by the large disparity between their training and test metrics. The RNN shows a test MAPE of 52.46%, while the LSTM’s test MAPE reaches 53.32%, indicating challenges in generalizing to unseen data. LightGBM and Prophet demonstrate contrasting behaviors. LightGBM shows strong training performance (MAPE: 1.88%) but struggles with generalization, as indicated by its test MAPE of 61.23%. Prophet shows the highest prediction error among all models, with a test MAPE of 131.58%, suggesting significant limitations in capturing NFT market dynamics.

4.3. Cross-Model Analysis

The comprehensive model evaluation (

Table 4. Performance metrics (MSE and MAE) for each model and prediction horizon.

Metric	CARD		LSTM		RNN		LightGBM		Prophet		Transformer
	MSE	MAE	MSE	MAE	MSE	MAE	MSE	MAE	MSE	MAE	MSE	MAE
96	0.142	0.128	0.217	0.183	0.183	0.217	0.253	0.167	0.195	0.158	0.201	0.162
192	0.158	0.139	0.241	0.209	0.269	0.231	0.209	0.191	0.227	0.179	0.235	0.188
336	0.173	0.152	0.284	0.234	0.235	0.258	0.295	0.243	0.243	0.202	0.261	0.211
720	0.189	0.167	0.317	0.259	0.347	0.281	0.279	0.241	0.293	0.231	0.302	0.239
Avg	0.166	0.147	0.265	0.221	0.258	0.247	0.259	0.210	0.239	0.193	0.250	0.200

) reveals CARD’s superior predictive capability across all horizons. At 96-h forecasts, CARD achieves an MAE of 0.128, outperforming LSTM (0.183) and LightGBM (0.167) by 30% and 23% respectively. This performance gap amplifies with extended horizons, culminating in a 35% MAE advantage over the nearest competitor (LSTM) at 720-h predictions. While LightGBM shows transient competitiveness at 192-h forecasts (MAE: 0.191 vs. CARD’s 0.139), its error growth rate ($\nabla MAE/hour=0.015\%)$ triples CARD’s (0.005%), exposing fundamental limitations in handling long-range dependencies.

Deep learning architectures exhibit pathological error patterns—LSTM’s test MAE (0.259) at 720 h exceeds its training performance (0.03%) by 763%, demonstrating catastrophic overfitting. This aligns with NFT markets’ high-frequency noise characteristics that recurrent networks struggle to generalize. Prophet displays consistent underperformance (average MAE: 0.193), particularly failing to capture royalty fee dynamics during market corrections, with 41% larger errors in bear regimes compared to CARD. Prophet displays consistent underperformance (average MAE: 0.193), particularly failing to capture royalty fee dynamics during market corrections, with 41% larger errors in bear regimes compared to CARD. Figure 1 illustrates the comparative MSE performance across all models, clearly demonstrating the superior predictive capability of the CARD model.

4.4. Cross-Model Horizon Analysis

The horizon-dependent performance analysis reveals critical distinctions in how competing models adapt to NFT market dynamics across varying prediction windows. Our proposed CARD architecture demonstrates robust capabilities in both short- and long-term forecasting, outperforming conventional approaches through all tested horizons. At 96 h, CARD achieves an MAE of 0.128, representing a 30.1% improvement over LSTM (0.183 MAE) and a 23.4% advantage against LightGBM (0.167 MAE). This performance gap widens systematically as the prediction window extends, with CARD maintaining a 35.5% MAE superiority over LSTM (0.167 vs. 0.259) and 30.1% lead against Transformer (0.239) at the 720-h horizon. This performance gap widens systematically as the prediction window extends, with CARD maintaining a 35.5% MAE superiority over LSTM (0.167 vs. 0.259) and 30.1% lead against Transformer (0.239) at the 720-h horizon. As shown in Figure 2, the MAE performance across different prediction horizons consistently favors the CARD architecture over conventional approaches.

• Short-Term Dynamics (96 H)
  All models exhibit their strongest performance in this range, yet critical differences emerge. While RNN and LSTM show comparable MSE values (0.183 vs. 0.217), their MAE divergence (0.217 vs. 0.183) suggests recurrent architectures struggle with outlier events common in NFT markets. LightGBM’s relatively high MSE (0.253) despite moderate MAE (0.167) indicates sensitivity to extreme price fluctuations, a weakness exacerbated in longer horizons.
• Mid-Term Forecasting (192–336 H)
  The 336-h (14-day) window emerges as a critical stress test. CARD’s MAE (0.152) remains 35.0% lower than LSTM (0.234) and 37.4% below LightGBM (0.243), highlighting its capacity to capture NFT market cycles. Transformer architectures show gradual degradation, with MAE increasing 29.6% from 96h (0.162) to 336h (0.211), compared to CARD’s 18.8% rise. Prophet’s performance collapse at this horizon (0.202 MAE vs. CARD’s 0.152) correlates with its inability to model sudden shifts in royalty-driven valuation patterns.
• Long-Term Predictions (720 H)
  At the maximum tested horizon, architectural limitations become stark. LSTM’s MAE (0.259) exceeds its 96-h performance by 41.5%, while LightGBM shows 44.3% degradation (0.167 → 0.241). CARD maintains superior stability, with 720 h MAE (0.167) remaining 30.5% above its 96h baseline—a critical advantage given NFT markets’ propensity for regime shifts. The Transformer’s 47.5% error growth (0.162 → 0.239 MAE) underscores the inefficiency of generic attention mechanisms for long-horizon NFT forecasting.

Three systemic failure modes explain conventional models’ shortcomings:

• Temporal Myopia in Recurrent Architectures
  LSTM’s quadratic error growth (41.5% MAE increase from 96h to 720 h) reflects an inability to reconcile NFT markets’ high-frequency volatility with long-term liquidity trends.
• Momentum Bias in Tree-Based Models
  LightGBM’s transient competitiveness at 192 h (0.191 MAE) collapses by 336 h (0.243, 27.2% increase from 192 h), revealing overreliance on medium-term momentum signals that fail during market reversals.
• Structural Rigidity in Decomposition Models
  Prophet’s consistent underperformance (31.3% higher average MAE than CARD) stems from oversimplified trend assumptions incompatible with NFT markets’ creator-driven valuation dynamics.

4.5. Market Regime-Specific Feature Dynamics

The SHAP value analysis reveals distinct feature importance patterns across bear, bull, and neutral market regimes, offering actionable insights into NFT market behavior under varying conditions. Below, we dissect the critical drivers for each scenario using the provided SHAP values.

During market downturns, platform liquidity metrics and user retention dominate:

• Aggregator Stickiness: Aggregator user metrics (AGG UniqueUsers: 0.0546, AGG TotalUsers: 0.0523) show 24–27% higher importance than in bull markets, reflecting traders’ reliance on consolidated platforms during volatility.
• Transaction Compression: TRD AvgTxSize (0.0586) emerges as the most critical feature, with 47.6% higher importance than in bull regimes, signaling a shift toward smaller, frequent trades.
• Liquidity Indicators: Marketplace volumes (MKT VolumeUSD: 0.0563, TotalVolume: 0.0562) and total users (0.0561) collectively account for 15.8% of total feature importance, underscoring liquidity monitoring.
• Royalty Irrelevance: Royalty metrics (ROY BluechipFees: 0.0007) become negligible, indicating diminished focus on premium NFT projects.

This regime prioritizes survival metrics, where user retention and liquidity preservation outweigh transaction value.

Expansion phases highlight herd behavior and creator ecosystem engagement:

• Top Trader Influence: Top trader volumes (TOP VolumeETH: 0.0283, VolumeUSD: 0.0279) inversely correlate with aggregator activity, suggesting traders bypass aggregators to chase momentum.
• Royalty Resurgence: ROY PayingUsers (0.0214) gain 28.9% higher importance than in neutral markets, reflecting renewed interest in creator-driven value.
• Volume Decoupling: Marketplace volumes (MKT VolumeUSD: 0.0410) decline in importance (−27.2% vs bear markets), as traders prioritize speculative opportunities over platform fundamentals.

Bull markets exhibit a “follow the leader” dynamic, where top trader activity and royalty participation drive price trends.

Sideways markets emphasize premium NFTs and platform diversification:

• Royalty Premiums: ROY BluechipFees (0.1049) dominate with 149× higher importance than in bull markets, acting as stability anchors for valuation.
• Platform Diversification: AGG NonAggUsers (0.0432) rise to peak importance (+21.6% vs bear markets), indicating a shift toward direct marketplace engagement.
• Transaction Selectivity: TRD metrics (AvgTxSize: 0.0332, EthSalesMA: 0.0303) decline in influence, replaced by royalty fees and user behavior.

Neutral phases prioritize quality over quantity, with blue-chip royalties and platform flexibility serving as key stability indicators.

The analysis yields regime-specific guidance:

• Bear Markets: Monitor AGG UniqueUsers and MKT VolumeUSD for liquidity health; prioritize small, frequent trades (TRD AvgTxSize).
• Bull Markets: Track TOP VolumeETH and ROY PayingUsers to gauge momentum; reduce reliance on aggregators.
• Neutral Markets: Use ROY BluechipFees as stability signals; diversify across marketplaces (AGG NonAggUsers).

This regime-aware framework enables dynamic strategy adjustments, aligning feature monitoring with observed market behavior. Figure 3 presents the regime-specific SHAP values across features, illustrating how the importance of different factors varies depending on market conditions.

Table 5. Comparison of Features across Bear, Bull, and Neutral States.

Feature	Bear	Bull	Neutral
AGG TotalVolume	0.0403	0.0351	0.0226
AGG UniqueUsers	0.0546	0.0430	0.0344
AGG NonAggUsers	0.0523	0.0396	0.0432
AGG TotalUsers	0.0523	0.0406	0.0410
TRD AvgTradesPerUser	0.0408	0.0273	0.0258
TRD AvgTxSize	0.0586	0.0397	0.0332
TRD EthSalesMA	0.0486	0.0388	0.0303
TRD SalesCount	0.0291	0.0330	0.0148
TRD TotalSales	0.0460	0.0430	0.0275
TRD UniqueBuyers	0.0495	0.0402	0.0412
TRD UniqueSellers	0.0257	0.0308	0.0105
MKT TotalFees	0.0055	0.0041	0.0617
MKT TotalSales	0.0343	0.0367	0.0168
MKT TotalUsers	0.0561	0.0431	0.0417
MKT TotalVolume	0.0562	0.0405	0.0381
MKT VolumeUSD	0.0563	0.0410	0.0383
ROY BluechipFees	0.0007	0.0001	0.1049
ROY TotalFees	0.0001	0.0001	0.1049
ROY PayingUsers	0.0166	0.0214	0.0001
ROY PayingTx	0.0282	0.0316	0.0027
TOP TotalTraders	0.0348	0.0364	0.0220
TOP TotalTrades	0.0348	0.0371	0.0167
TOP VolumeETH	0.0356	0.0283	0.0460
TOP VolumeUSD	0.0358	0.0279	0.0471
TOT Sales	0.0343	0.0367	0.0168
TOT Users	0.0561	0.0431	0.0417
TOT Volume	0.0562	0.0405	0.0381
TOT VolumeUSD	0.0563	0.0410	0.0383
TOT ActiveTraders	0.0347	0.0358	0.0220

provides a detailed comparison of feature importance values across bear, bull, and neutral market states, further highlighting the regime-dependent nature of NFT market dynamics.

4.6. Temporal Stability Analysis: Assessing Model Robustness to Concept Drift

The NFT market has undergone significant structural evolution since its inception, with new marketplace models, changing creator incentives, and shifting collector behaviors. To assess how well the CARD model adapts to these evolving dynamics, we conducted a temporal stability analysis focusing on the model’s robustness to concept drift—the phenomenon where statistical properties of the target variable change over time, potentially degrading predictive performance.

4.6.1. Rolling Window Performance Evaluation

We implemented a rolling window approach to evaluate the model’s performance stability across different time periods. Using a six-month training window and a three-month forward-testing window, we progressively shifted this evaluation frame by one month at a time across our dataset, resulting in 18 distinct evaluation periods.

The results reveal that while all models experience performance fluctuations across time periods, CARD demonstrates superior stability with an average performance degradation of 12.3% between the best and worst periods, compared to 31.5% for LSTM and 28.7% for Transformer. This indicates CARD’s relative resilience to temporal shifts in market behavior, likely attributed to its adaptive attention mechanism that dynamically adjusts feature weights based on recent data patterns.

4.6.2. Feature Importance Stability

Beyond aggregate performance metrics, we examined the stability of feature importance rankings across time periods to assess how consistently the model identifies key predictors. For each rolling window period, we calculated SHAP values and ranked features by their importance, then measured rank volatility using Kendall’s tau coefficient between consecutive periods.

Table 6. Feature Importance Stability Across Market Phases and Time Periods.

Market Phase	Average Kendall’s Tau	Top-5 Stability%	Most Stable Feature	Most Volatile Feature
Bear Market	0.743	87.3%	AGG_UniqueUsers	ROY_BluechipFees
Bull Market	0.681	73.2%	TOP_VolumeETH	AGG_NonAggUsers
Neutral Market	0.705	81.5%	ROY_BluechipFees	TRD_SalesCount
Cross-Phase	0.412	42.8%	MKT_VolumeUSD	ROY_PayingUsers

presents the results of this analysis.

These results reveal several significant patterns:

• Within-Phase Stability: Feature importance rankings demonstrate high stability within the same market phase, with average Kendall’s tau values above 0.68 for all phases. This indicates that while the market evolves, the relative importance of features remains largely consistent within similar market conditions.
• Cross-Phase Volatility: When comparing feature rankings across different market phases, stability drops significantly (tau = 0.412), confirming our hypothesis that predictive factors fundamentally change as the market transitions between regimes.
• Phase-Specific Stable Features: Each market phase exhibits certain consistently important features: AGG_UniqueUsers during bear markets, TOP_VolumeETH during bull markets, and ROY_BluechipFees during neutral markets. These represent “anchor metrics” that maintain their predictive value despite other market changes.
• Emerging Indicators: Over time, we observed certain features gaining importance, particularly metrics related to aggregator platform usage (increasing 18.3% in importance over the study period) and royalty dynamics (increasing 21.7%), reflecting the market’s structural evolution toward platform consolidation and creator economics.

4.6.3. Drift Detection and Adaptation

To quantify the model’s sensitivity to concept drift, we calculated the Normalized Drift Detection Measure (NDDM) proposed by Lu et al. [64] for each rolling window period. This measure computes the statistical distance between feature distributions across time periods, with higher values indicating stronger drift conditions.

The analysis reveals that CARD maintains relatively stable performance even during periods of high drift (NDDM > 0.7), with an average error increase of 14.8% compared to 37.2% for LSTM and 29.5% for Transformer. This resilience can be attributed to two key factors:

• Dynamic Channel Attention: The channel-wise attention mechanism automatically adjusts feature importance weights based on recent data patterns, effectively performing implicit feature reweighting as market conditions evolve.
• Hierarchical Temporal Processing: The model’s multi-scale approach to time series processing allows it to simultaneously adapt to both rapid shifts in short-term patterns and gradual changes in longer-term trends.

To further quantify adaptation capability, we calculated the Drift Recovery Rate (DRR), defined as the number of time steps required for model performance to return within 10% of pre-drift levels following a significant drift event (NDDM > 0.7). CARD achieves a DRR of 3.2 time periods, compared to 7.5 for LSTM and 5.3 for Transformer, indicating substantially faster adaptation to new market conditions.

4.6.4. Persistence of Regime-Specific Patterns

Finally, we examined whether the regime-specific feature importance patterns identified in Section 4.5 persist across different time periods. For each market regime (bear, bull, neutral), we calculated the Intra-Class Correlation (ICC) of feature importance rankings across different time windows within the same regime, and compared this to the Inter-Class Correlation across different regimes.

The analysis confirms that regime-specific feature importance patterns demonstrate high temporal persistence (average ICC = 0.783), significantly exceeding cross-regime similarity (average ICC = 0.325). This suggests that while market structures evolve over time, the fundamental valuation logic within each market phase remains relatively stable. The most persistent patterns were observed in bear markets (ICC = 0.819), followed by neutral markets (ICC = 0.780) and bull markets (ICC = 0.749).

These findings have significant implications for both model design and practical application. They suggest that while continuous retraining is beneficial to maintain optimal predictive performance, the fundamental insights regarding phase-specific predictive factors remain valid even as the market evolves. From a practical perspective, this validates the usefulness of phase-based monitoring strategies for market participants, as the key indicators identified for each regime remain relevant across different time periods.

4.7. Computational Efficiency and Real-Time Feasibility

While predictive accuracy is a primary concern for model evaluation, practical deployment in NFT trading platforms requires careful consideration of computational efficiency and real-time feasibility. To address these concerns, we conducted a comprehensive analysis of resource utilization and inference speed across all tested models, focusing on metrics directly relevant to production environments.

4.7.1. Inference Latency Analysis

We measured inference latency—defined as the time required to generate a prediction given input data—across all models using identical hardware configurations (Nvidia A100 GPU through Google Colab environment). Each model was evaluated using batch sizes of 1, 16, 64, and 256 to simulate different deployment scenarios, from individual real-time predictions to batch processing.

Table 7. Inference Latency Comparison Across Models (milliseconds).

Model	Batch Size 1	Batch Size 16	Batch Size 64	Batch Size 256
CARD	12.3 ± 0.8	43.7 ± 1.2	128.6 ± 2.5	415.8 ± 5.3
LSTM	8.5 ± 0.6	37.2 ± 1.1	112.4 ± 2.1	385.3 ± 4.8
RNN	6.2 ± 0.4	28.5 ± 0.9	87.3 ± 1.8	301.7 ± 3.9
Transformer	15.8 ± 1.0	52.3 ± 1.4	146.9 ± 2.8	486.2 ± 6.7
LightGBM	3.1 ± 0.2	15.4 ± 0.6	41.2 ± 1.2	142.5 ± 2.3
Prophet	4.5 ± 0.3	46.8 ± 1.3	172.5 ± 3.5	612.7 ± 8.9

presents the results of this analysis.

The results reveal that CARD exhibits moderate inference latency compared to other tested models. While not as fast as LightGBM or simple RNN architectures, CARD’s latency remains well within the threshold for near-real-time applications. At batch size 1, which represents the most demanding real-time scenario, CARD’s average inference time of 12.3 ms is significantly below the 50 ms threshold typically considered acceptable for interactive trading applications.

Notably, when compared to the Transformer architecture, which shares some architectural similarities, CARD demonstrates 22.2% lower latency across all batch sizes, likely due to its more efficient channel-wise attention mechanism that reduces computational complexity. This performance advantage becomes particularly important in high-frequency prediction scenarios where multiple updates may be required per minute.

4.7.2. Memory Usage and Model Size

Memory efficiency is another critical consideration for deployment scenarios, particularly for edge computing applications or systems with limited resources. We analyzed both the static model size (parameters) and runtime memory requirements during inference.

Table 8. Model Size Comparison.

Model	Parameter Count	Model Size (MB)	FLOPS per Inference
CARD	2.83 M	11.32	284.5 M
LSTM	3.42 M	13.68	327.6 M
RNN	1.15 M	4.60	112.8 M
Transformer	4.26 M	17.04	405.2 M
LightGBM	0.58 M	2.32	25.3 M
Prophet	-	0.85	-

provides a comprehensive comparison of model sizes, parameter counts, and computational requirements across all evaluated architectures.

The analysis reveals that CARD strikes a favorable balance between model complexity and memory efficiency. With 2.83 million parameters, CARD requires 17.3% less memory than LSTM and 33.6% less than Transformer models during inference. This efficiency is particularly notable given CARD’s superior predictive accuracy, suggesting a well-optimized architecture that effectively utilizes its parameters.

The floating-point operations per second (FLOPS) analysis further confirms this efficiency, with CARD requiring 13.2% fewer operations than LSTM and 29.8% fewer than Transformer for a single inference pass. This translates directly to lower computational costs and energy consumption in production environments.

4.7.3. Scaling Analysis and Deployment Scenarios

To assess how computational requirements scale with input sequence length—a critical factor for long-horizon predictions—we conducted a scaling analysis measuring how inference time increases as the lookback window expands.

The results demonstrate that CARD exhibits near-linear scaling with sequence length, in contrast to the quadratic scaling observed in the standard Transformer architecture. This favorable scaling characteristic is attributed to CARD’s relative distance encoding mechanism, which reduces redundant computations when processing long sequences. For sequence lengths beyond 336 h (two weeks), this advantage becomes particularly pronounced, with CARD maintaining reasonable inference times even for extended historical contexts.

Based on these efficiency metrics, we can identify optimal deployment scenarios for CARD in NFT market applications:

• Trading Signal Generation: With sub-15ms inference time for individual predictions, CARD is suitable for generating near-real-time trading signals, potentially supporting algorithmic trading systems that require frequent model updates.
• Portfolio Monitoring: For investment managers tracking multiple NFT collections simultaneously, CARD’s efficient batch processing (43.7 ms for 16 parallel predictions) enables comprehensive portfolio monitoring with minimal computational overhead.
• Market Surveillance: The model’s ability to efficiently process long historical sequences makes it well-suited for market surveillance applications that require detecting anomalous patterns against extended historical contexts.

4.7.4. Optimization Strategies for Real-Time Deployment

For deployment scenarios with strict latency requirements, we explored several optimization strategies to further reduce CARD’s computational overhead without sacrificing predictive accuracy.

Table 9. Impact of Optimization Strategies on CARD Performance.

Optimization Technique	Latency Reduction	MAE Impact	Memory Reduction
Knowledge Distillation	35.8%	+2.3%	41.2%
Quantization (INT8)	47.2%	+3.5%	73.6%
Pruning (30% sparsity)	22.5%	+1.8%	28.4%
Feature Selection (top 20)	18.3%	+1.2%	12.5%
Hybrid CPU-GPU Pipeline	15.6%	0.0%	−5.3%

summarizes the impact of these optimizations on inference time and model performance.

Knowledge distillation—where a smaller “student” model is trained to mimic the full “teacher” CARD model—offers particularly promising results, reducing inference time by 35.8% with only a 2.3% increase in MAE. This approach enables deployment on resource-constrained environments while preserving most of the predictive accuracy. Similarly, applying quantization to convert model weights from floating-point to 8-bit integer representation achieves substantial latency and memory improvements with acceptable accuracy trade-offs.

For the most latency-sensitive applications, combining feature selection with model quantization can reduce inference time by over 65% while keeping accuracy degradation under 5%, potentially enabling deployment on edge devices or web browsers for client-side prediction. These optimization pathways provide a flexible framework for adapting CARD to various operational constraints while maintaining its core predictive advantages.

In summary, while CARD requires moderately higher computational resources than simpler models like LightGBM or basic RNNs, its resource utilization remains within practical limits for real-time and near-real-time applications. The model’s superior predictive accuracy justifies this moderate efficiency trade-off, particularly for applications where prediction error directly impacts financial outcomes. For deployment scenarios with extreme efficiency requirements, optimization techniques like knowledge distillation and quantization offer viable pathways to further reduce computational overhead while preserving most of CARD’s performance advantages.

5. Discussion

Our empirical analysis of NFT market dynamics through the CARD model provides significant insights into how different features influence valuation across market phases. Despite its superior predictive performance, the model exhibits limitations that warrant consideration. First, while effective for observed patterns, the dynamic attention mechanism may struggle during unprecedented “black swan” events, as noted by Urom et al. [53]. Second, the model’s reliance on historical regime patterns limits its adaptability to novel market configurations, particularly as NFT marketplaces continue to evolve structurally. Third, though SHAP analysis identifies important relationships, the causal mechanisms remain partially obscured—a limitation aligned with Wang et al. [55]’s observation that machine learning approaches often excel at pattern recognition but struggle with causal interpretation in complex markets.

The regime-dependent variations in feature importance reveal fundamental shifts in NFT market behavior across different conditions. During bear markets, the prominence of aggregator metrics (AGG_UniqueUsers: 0.0546) and transaction sizing (TRD_AvgTxSize: 0.0586) reflects a shift toward risk management and execution certainty, aligning with Liu et al. [50]’s observations. In bull markets, the inverse relationship between aggregator usage and top trader activity (TOP_VolumeETH: 0.0283) suggests a strategic bifurcation where sophisticated traders bypass aggregator platforms to avoid telegraphing their strategies, supporting Kapoor et al. [34]’s findings on trend-setter behavior. Most notably, neutral markets exhibit an unprecedented dominance of royalty metrics (ROY_BluechipFees: 0.1049), revealing a flight-to-quality mechanism where investors anchor valuations on creator fundamentals when directional signals are ambiguous—a pattern not previously documented quantitatively and distinguishing NFTs from traditional financial markets.

These findings contribute to several theoretical debates in digital asset valuation. First, they support the emerging view that NFTs represent a distinct asset class with unique microstructure properties rather than simply an extension of cryptocurrency markets, aligning with Ko et al. [30]’s argument while providing new empirical evidence. Second, our analysis challenges traditional efficient market assumptions by demonstrating how information flows differently through NFT markets across regimes, supporting Wilkoff and Yildiz [32]’s observation that NFT markets exhibit regime-dependent efficiency. Third, the results highlight the critical role of ecosystem structure in NFT valuations, particularly the dynamic between direct marketplace activity and aggregator-mediated transactions, extending Das et al. [37]’s ecosystem-centric view of NFT markets with quantitative evidence of how ecosystem relationships reconfigure across market phases.

From an implementation perspective, the CARD model can be practically deployed within NFT market infrastructure through a RESTful API architecture consisting of a data ingestion layer collecting real-time marketplace data, a prediction engine incorporating regime detection, and an API gateway exposing standardized endpoints. This approach could enhance market participants’ decision-making through integration with trading platforms, investor dashboards, and market surveillance systems. Implementation challenges include maintaining data freshness across multiple platforms, optimizing computational resources for large-scale deployment, designing intuitive explanatory interfaces, and accounting for potential feedback loops when prediction tools influence the market dynamics they attempt to predict. Despite these challenges, the model’s ability to provide regime-aware predictions through standardized interfaces offers significant potential for improving NFT market efficiency and price discovery.

6. Conclusions

This study establishes a significant advancement in NFT market forecasting through the Channel-wise Attention with Relative Distance (CARD) model, which demonstrates superior predictive performance across multiple time horizons. By integrating hierarchical attention mechanisms with adaptive temporal encoding, CARD achieves a 33.5% reduction in MAE compared to conventional LSTM architectures, with particularly impressive performance in long-term predictions (720-h MAE: 0.167 vs. LSTM’s 0.259). Our regime-specific analysis reveals distinct feature importance patterns across market phases, with bear markets prioritizing liquidity metrics, bull markets emphasizing momentum signals, and neutral markets focusing on creator economics.

These findings contribute to both academic understanding and practical applications in several ways. Academically, our work addresses methodological gaps identified by Nobanee and Ellili [56] in existing time series approaches to NFT markets. The CARD model extends Xue et al. [6]’s work on temporal attention mechanisms by incorporating market phase-specific adaptations, while our regime-dependent feature importance analysis provides quantitative evidence for creator-centric valuation patterns described by Vasan et al. [43]. The variable importance of marketplace fees across regimes complements Aharon and Demir [39]’s analysis of cross-market effects, while our findings on top trader influence during bull markets provide empirical support for Kapoor et al. [34]’s social contagion model of NFT price formation.

From a practical perspective, our results offer actionable insights for various stakeholders. Investors can leverage regime-specific feature patterns to adapt their monitoring strategies, focusing on different metrics depending on market conditions. Marketplace operators can prioritize platform features based on current market phases, with user retention becoming particularly critical during downturns. Creators and project developers can optimize royalty strategies knowing that their economic impact varies dramatically across market phases, a finding that extends Wu et al. [47]’s work on project success factors with temporal specificity.

Despite these contributions, several limitations must be acknowledged. Our dataset focuses primarily on Ethereum-based NFT markets, excluding emerging ecosystems on alternative blockchains. We rely on on-chain transaction data, missing important off-chain factors such as social media sentiment and project announcements. The hourly aggregation of transaction data may obscure microstructure patterns occurring at smaller time intervals, and despite filtering efforts, some artificial trading activity likely remains in the dataset.

Future research could address these limitations through several extensions. The CARD architecture shows potential application beyond NFT markets to other domains sharing similar characteristics of volatility and regime-dependent behavior, including cryptocurrency markets, DeFi, traditional equities, and security tokens. Macroeconomic integration represents another opportunity for enhancement, as NFT markets show varying sensitivity to traditional financial indicators across market regimes. The regime-specific insights could inform specialized trading strategies with dynamic allocation frameworks based on detected market phases. Additionally, examining cross-platform dynamics could reveal how platform competition evolves with market conditions, while the integration of NFTs with metaverse platforms presents new frontiers for predictive modeling as NFTs increasingly serve functional roles beyond collectibles.

By building on the foundation established in this study, future research can develop more comprehensive models for predicting behavior across digital and traditional financial markets, providing valuable frameworks for stakeholders navigating this evolving landscape.