Tokenized Asset Pricing Models Using Graph Neural Networks

Abstract

Turing tokens into fiat money are truly ushering a new era in monetary systems. The trading of tokenised securities, commodities, real estate, and other assets is often conducted over decentralised networks with highly linked transactional structures. The temporal and relational complexities ingrained into such systems evade representation by traditional asset pricing methods. To overcome this challenge, this work presents a new tokenised-asset-pricing framework that exploits the interdependencies and topological structure of blockchain transactions and is based on Graph Neural Networks (GNNs). We use a dynamic graph model to describe tokenised assets and their interactions, where nodes may represent tokens or wallets, and edges may represent interactions between smart contracts, co-ownership agreements, or transactional connections. To capture market dynamics, understand liquidity flows, and discern asset correlations over time, our proposed system learns expressive node embeddings by applying GNN architectures such as GCNs and GATs. Building upon these methods, temporal graph models are considered to accommodate the changing market environment and transaction profiles. The GNN pricing model proposed is tested on real-world datasets collected from Ethereum-based DeFi platforms and compared with baseline ML models and traditional pricing models. The results point to significantly improved predictive accuracy, resistance to market volatility, and adjustment to different token architectures. This study exemplifies how graph-based deep learning techniques could replace traditional models for digital asset appraisal by one that is more accurate and more scalable. It also sheds light on price discovery in decentralised setups, detection of market manipulation, and contagion of risks.