Dr. Gabriel A. Abeyie

This paper proposes a method for enhancing text-based forecasting models, with a specific focus on predicting crude oil prices. Utilizing advanced techniques, including pattern validation and attention mechanisms, the study demonstrates notable improvements in predictive power over traditional approaches. One key finding is that considering the full text of news articles, rather than limiting the analysis to news headlines, leads to significant gains in forecasting accuracy. Furthermore, the model featuring verb-noun and noun-verb collocation pattern validation consistently outperforms benchmarks and models based solely on news headlines across various forecasting horizons. The results suggest that the presence of such collocations as ’price fell,’ ’prices tumbled,’ and ’price dropped’ in crude-oil-related news articles is associated with a decrease in oil price returns. Additionally, integrating macroeconomic data with text-based features enhances predictive performance, demonstrating that combining structured economic indicators with textual features improves forecasting accuracy.

This paper studies whether oil-market narratives have macroeconomic effects, and whether those effects differ depending on whether narratives convey directional information or uncertainty. I construct new monthly oil narrative sentiment and uncertainty indices from Wall Street Journal and New York Times articles from 1953 to 2022 and classify oil-related coverage into five narrative types. Using the typespecific narrative decomposition as external instruments in a Proxy-SVAR, I identify two structurally distinct narrative shocks. A bullish narrative sentiment shock raises the real oil price persistently, increases world industrial production, raises world oil production, and gradually passes through to consumer prices, consistent with a demand-side information channel. By contrast, an elevated narrative uncertainty shock reduces world oil production and world industrial production, while producing only a muted response of the real oil price and consumer prices, consistent with a supply-side real options channel. When both shocks hit simultaneously, the combined effect produces a stagflationary pattern in which oil prices and inflation rise while production and activity respond weakly. These findings show that oil-market narratives contain macroeconomically relevant variation beyond physical oil shocks and that separating directional tone from ambiguity is essential for understanding how oil news transmits to the broader economy.

We study whether economic narratives improve equity premium forecasts and how forecast combination should adapt to changes in the narrative environment. Using narrative sentiment features extracted from major U.S. newspapers over 1940–2021, we generate machine-learning forecasts via Lasso, Elastic Net, and Random Forest, and combine them using Narrative Attention Shrinkage (NAS). Motivated by rational inattention, NAS uses a real-time Narrative Attention Index (NAI) to modulate shrinkage: when narrative activity is elevated, weights concentrate on the strongest models; when activity is low, weights revert toward equal averaging. Out of sample, NAS delivers economically and statistically significant gains relative to the historical average and standard combination benchmarks, with improvements concentrated in crisis episodes characterized by heightened narrative activity and model disagreement. These findings provide a state-dependent resolution to the forecast combination puzzle and show that narrative information is most effectively exploited through adaptive model selection.

Traditional event studies assume that all announcements of a given type generate price reactions at the same speed, justifying fixed event windows. Using high-frequency data on crude oil futures and 612 weekly EIA inventory announcements from 2012-2024, we show this assumption is violated in economically meaningful ways. We develop a two-stage framework that first predicts reaction timing using machine learning, then estimates regime-conditional price impacts. Applying an incremental returns method to isolate genuine price reactions from drift, we identify three distinct regimes: FAST (0-30 minutes, 24% of announcements), NORMAL (35-90 minutes, 45%), and SLOW (95+ minutes, 31%). The NORMAL regime achieves highly significant surprise-return relationships. FAST reactions are immediate but statistically insignificant, while SLOW reactions show no systematic response to inventory surprises. Our two-stage model nearly replicates oracle performance, demonstrating that accounting for heterogeneous timing substantially improves event study measurement.

This paper uses textual analysis to identify and measure the impact of OPEC news shocks on oil markets and the broader economy, developing novel methods to extract information from OPEC announcements and communications.

This research examines the welfare consequences of asset price bubble cycles, analyzing how boom-bust patterns in asset markets affect economic efficiency and social welfare.

This paper develops a novel forecasting methodology combining principal component analysis with weighted elastic net regularization to improve crude oil price return predictions using high-dimensional predictor sets.