Vegard Høghaug Larsen
Associate Professor of Data Science and Economics at BI Norwegian Business School
I work in the department of Data Science and Analytics at BI Norwegian Business School (BI). I have a Master of Science in Economics from the Norwegian University of Science and Technology (NTNU), and a PhD in Economics from BI. I am a Distinguished CESifo Affiliate and also affiliated with the Centre for Applied Macroeconomics and Commodity Prices. Before my current roles, I held a position as a Senior Researcher at Norges Bank, the central bank of Norway.
My work is at the intersection of economics and data science, where I use machine learning and natural language processing techniques to study the transmission of economic shocks, understand how agents form their expectations, and develop methods to measure unobserved concepts such as sentiment, uncertainty, and climate risk. My papers have been published in journals such as Journal of Econometrics, American Economic Journal: Macroeconomics, Journal of Monetary Economics, and International Economic Review.
See my CV for more.
Building on recent advances in Natural Language Processing and modeling of sequences we study how a Transformer-based deep learning architecture with multimodal features can be used to narrate the business cycle. The framework we propose combines text (news) and (macroeconomic) time series information using embeddings and multihead cross-attention mechanisms and incorporates important empirical regularities, such as potential differences in data frequencies, reporting delays, and time-dependent non-linearities. The model is trained on the dual task of tracking (future) business cycle fluctuations and providing a more structural decomposition of these fluctuations by classifying the underlying news flow into either demand, supply, or noise.
We study how speeches given by Federal Open Market Committee (FOMC) members and regional Federal Reserve presidents influence private sector expectations. Speeches emphasizing inflationary pressures lead households and professional forecasters to raise their inflation expectations, suggestive of Delphic effects. We then investigate whether FOMC speeches signaling speakers' firmer determination to respond to the communicated inflationary pressures offset these Delphic responses. While professional forecasters lower their expectations in response to such Odyssean communications, households do not, leaving the Delphic effects dominant. This pattern is consistent with a model in which agents vary in their ability to interpret Odyssean communications.
Climate change increases the likelihood of extreme climate- and weather-related events, but also the pressure to adjust to a lower-carbon economy. We propose a measure of climate change transition risk, based on neural-network word embedding models for large-scale text analysis, and document that when it unexpectedly increases, major commodity currencies experience a persistent depreciation in line with economic theory. Expanding the analysis to a richer set of countries confirms a negative correlation between a country’s carbon export dependency and exchange rate response to transition risk. Word embeddings have been crucial for scientific advances and improvements on down-stream tasks in the Natural Language Processing literature the last decade. Our study shows how they can be used to quantify an important but hard to measure concept in economics.
Working paper available here.
Topic based uncertainty measures for Norway (daily series, updated March 2023, 14.5 MB) based on the paper Components of Uncertainty.
Norwegian Economic Policy Uncertainty (EPU) Index, (monthly series, updated March 2023, 12 KB) see the Online appendix of Components of Uncertainty for details.