Artificial intelligence and big holdings data: Opportunities for central banks

Summary

Focus

Why do asset prices fluctuate? How do central bank interventions affect asset prices? And which investors or assets are exposed to the same risk factors? When aggregate demand for securities must equal their total supply, their sensitivity to asset prices can be estimated using data sets on investor portfolio holdings. Such demand systems help explain price movements, providing valuable insights for central banks. Moreover, artificial intelligence tools traditionally used for language tasks offer new analytical methods to study these asset demand systems.

Contribution

A key ingredient to understanding price elasticity is knowing which assets are close substitutes and which investors tend to behave similarly. However, observed characteristics such as sector or balance sheet variables do not tell the whole story. For example, before the Covid-19 pandemic, data on companies' sensitivities to lockdowns were not available, but investors responded in real time to their assessment of "winners" and "losers". The paper explains how the artificial intelligence method of "embedding" assets and investors in a vector space helps uncover such market reactions. Asset embeddings enable central banks to better understand how asset prices change. And investor embeddings offer insights on investors' likely response to central bank interventions or other market movements.

Findings

The paper illustrates the use of embedding techniques in a number of use cases. Since asset embeddings represent investors' views of securities that are close substitutes, it can predict what investors buy after selling some of their portfolio to central banks in asset purchase programmes. Similarly, these embeddings offer a more nuanced glimpse into so-called "crowded trades", by finding companies that investors judge to be exposed to the same factors, even in the absence of data showing direct similarity. These models can also be used to design stress testing models. And beyond financial markets, embeddings uncovered using these techniques can provide insights on the dynamics of relative prices and consumer heterogeneity.

Abstract

Asset demand systems specify the demand of investors for financial assets and the supply of securities by firms. We discuss how realistic models of the asset demand system are essential to assess ex post, and predict ex ante, how central bank policy interventions impact asset prices, the distribution of wealth across households and institutions, and financial stability. Due to the improved availability of big holdings data and advances in modelling techniques, estimating asset demand systems is now a practical reality. We show how demand systems provide improved information for policy decisions (eg in the context of financial contagion, convenience yield or the strength of the dollar) or to design optimal policies (eg in the context of quantitative easing or designing climate stress tests). We discuss how recent AI methods can be used to improve models of the asset demand system by better measuring asset and investor similarity through so-called embeddings. These embeddings can for instance be used for policymaking by central banks to understand the rebalancing channel of asset purchase programs and to measure crowded trades.

Non-technical background paper based on main paper "Asset embeddings"

JEL classification: C5, G11, G12

Keywords: asset prices, central bank policies, artificial intelligence, embeddings