HNDS-R Collaborative Research: Measuring Belief Distortions to Improve Predictive Outcomes

Systematic expectational errors embedded in beliefs (belief distortions) can have important effects on the economy. This project aims to explore the relation between belief distortions and economic outcomes. The research will explore to what extent beliefs pertaining to key variables closely related to monetary policy, such as interest rates and financial aggregates, may be distorted.

The project will also examine how the beliefs and sentiments voiced in polling, such as those pertaining to elections, may produce biased forecasts of election outcomes. In exploring these questions, the project will utilize textual data from written documents such as online news outlets and social media, and data from betting markets, which will further be combined with other economic indicators.

By utilizing this data, the project will account for biases in surveyed expectations about the future path of monetary policy as well as the performance of polls and surveys in predicting financial market variables and election outcomes. The project will develop machine learning based methods to improve prediction and estimation in a range of settings that rely on surveys by uncovering systematic errors in survey responses, and by correcting these errors using artificial intelligence algorithms.

These tools will provide considerable potential improvements in prediction accuracy of surveys in a variety of contexts in the economy.

A fundamental challenge in addressing whether beliefs are biased is that no objective measure of such distortions exists. This research aims to address this challenge by leveraging advancements in machine learning. A general premise of the approach followed in this project is that big data algorithms can be productively employed to reveal subjective biases in human judgements in multiple contexts, thereby facilitating more accurate objective forecasts.

The project will construct and study a comprehensive, methodologically consistent, econometric measure of belief distortions in expectations about future monetary policy and electoral outcomes, among other variables. This objective requires employment of large amounts of data related to real-time decision making and machine-learning tools to reduce sampling noise.

Data scraped from written documents will be analyzed using a Latent Dirichlet Allocation type of model to extract high-frequency measures of the topics covered by news outlets and social media. Data about election outcomes, betting markets, and financial market futures contracts will be added to a large real time dataset of economic information. The research will then incorporate these methodologies and data to study beliefs and possible judgmental errors found in survey expectations related to the future conduct of monetary policy, the behavior of financial markets, and the outcomes of elections.

The inclusion of fast-moving variables taken from options and futures markets and scraped from online documents will enable high-frequency tracking of revisions in beliefs and objective forecasts. This framework will allow high-frequency revisions in beliefs to be connected to specific news and announcements to improve prediction accuracy.

This award reflects NSF's statutory mission and has been deemed worthy of support through evaluation using the Foundation's intellectual merit and broader impacts review criteria.