Asset Embeddings

Firm characteristics, based on accounting and financial market data, are commonly used to represent firms in economics and finance. However, investors collectively use a much richer information set beyond firm characteristics, including sources of information that are not readily available to researchers. We show theoretically that portfolio holdings contain all relevant information for asset pricing, which can be recovered under empirically realistic conditions. Such guarantees do not exist for other data sources, such as accounting or text data. We build on recent advances in artificial intelligence (AI) and machine learning (ML) that represent unstructured data (e.g., text, audio, and images) by high-dimensional latent vectors called embeddings. Just as word embeddings leverage the document structure to represent words, asset embeddings leverage portfolio holdings to represent firms. Thus, this paper is a bridge from recent advances in AI and ML to economics and finance. We explore various methods to estimate asset embeddings, including recommender systems, shallow neural network models such as Word2Vec, and transformer models such as BERT. We evaluate the performance of these models on three benchmarks that can be evaluated using a single quarter of data: predicting relative valuations, explaining the comovement of stock returns, and predicting institutional portfolio decisions. We also estimate investor embeddings (i.e., representations of investors and their strategies), which are useful for investor classification, performance evaluation, and detecting crowded trades. We discuss other applications of asset embeddings, including generative portfolios, risk management, and stress testing. Finally, we develop a framework to give an economic narrative to a group of similar firms, by applying large language models to firm-level text data.