Working Papers
Reverse chronological order by first public-release year.
Building Macroeconomically Relevant Climate Indices via the Assemblage VAR
2026 · Conditionally accepted · The Econometrics Journal · with Christophe Barrette and Tim Reinicke
TL;DR: What should a macroeconomically relevant climate index contain? We introduce the Assemblage VAR, which jointly estimates aggregation weights and VAR parameters. Applied to U.S. climate data, the resulting indices emphasize high winds and distributional tails, yielding substantially larger contractionary impulse responses than fixed-weight benchmarks.
[ SSRN ]
LGB+: A Macroeconomic Forecasting Road Test
2026 · Revise and resubmit · Journal of Economic Dynamics and Control
TL;DR: Linear structure remains remarkably pervasive in macroeconomic forecasting, even when nonlinearities matter. LGB+ lets linear and tree updates compete within the same boosting procedure, preserving simple structure where it works and adding nonlinear flexibility where it helps.
Quantifying the Risk-Return Tradeoff in Forecasting
2026
TL;DR: Average forecast accuracy can hide substantial instability over time. By treating forecast loss differentials as returns, this paper brings risk-adjusted performance measures from finance into forecast evaluation and distinguishes persistent gains from fragile ones.
Panel Machine Learning with Mixed-Frequency Data: Monitoring State-Level Fiscal Variables
2025 · with Massimiliano Marcellino and Dalibor Stevanovic
TL;DR: We combine machine learning, mixed-frequency predictors, and panel data to nowcast U.S. state-level fiscal variables. Neural networks learn both shared patterns and differences across states, improving forecasts particularly for volatile expenditures and deficits.
Opening the Black Box of Local Projections
2025 · with Karin Klieber
TL;DR: Which historical episodes drive an estimated impulse response? We decompose local projection estimates into weighted sums of past outcomes, making their historical foundations explicit. The approach also extends to nonlinear machine learning models.
[ SSRN ] [ arXiv ] [ Slides ] [ R Code ] [ Python Code ] [ LinkedIn Discussion ]
Dual Interpretation of Machine Learning Forecasts
2024 · with Maximilian Göbel and Karin Klieber
TL;DR: Machine learning forecasts can be interpreted as combinations of past outcomes. This dual perspective reveals the historical observations behind a prediction, turning an otherwise opaque forecast into a portfolio of historical analogies.
[ SSRN ] [ arXiv ] [ Slides ] [ R Code ] [ Python Code ]
Maximally Forward-Looking Core Inflation
2024 · Revise and resubmit · Journal of Applied Econometrics · with Karin Klieber, Christophe Barrette, and Maximilian Göbel
TL;DR: A useful core inflation measure should tell us where inflation is going. We introduce Assemblage Regression, which learns weights on inflation components to construct an aggregate optimized for predicting future inflation. Applying the same approach to ordered component inflation rates yields a supervised trimmed inflation measure.
[ SSRN ] [ arXiv ] [ Slides ] [ Code ]
The Anatomy of Machine Learning-Based Portfolio Performance
2023 · with Dave Rapach, Erik Christian Montes Schütte, and Sander Schwenk-Nebbe
TL;DR: We introduce a Shapley-value method that decomposes portfolio performance and attributes the economic value of return predictability to individual predictors or groups of predictors, even when forecasts come from black-box machine learning models.
[ SSRN ]
Maximally Machine-Learnable Portfolios
2023 · with Maximilian Göbel
TL;DR: Can we construct a portfolio specifically to make its returns predictable? We introduce MACE, an algorithm that jointly learns portfolio weights and a forecasting model. It extends Alternating Conditional Expectations by combining a random forest with a constrained ridge regression.
The Anatomy of Out-of-Sample Forecasting Accuracy
2022 · Conditionally accepted · Journal of Applied Econometrics · with Daniel Borup, Dave Rapach, Erik Christian Montes Schütte, and Sander Schwenk-Nebbe
TL;DR: Which predictors actually improve a forecast? We introduce Performance-based Shapley Values, which identify how much each predictor increases or decreases out-of-sample RMSE, providing a detailed account of forecasting accuracy.
[ SSRN ] [ Blog by Sander Schwenk-Nebbe ] [ Python Code ]
Work in Progress
Implied Factors: The Linear Skeleton of Machine Learning Forecasts
with Karin Klieber and Gabriel Rodriguez-Rondon
Supervised Detrending
with Josefine Quast and Anne Valder
A Lingua Franca for Macroeconomic Data
with Maximilian Göbel
Publications & Accepted Papers
Reverse chronological order by publication year.
17. Ordinary Least Squares as an Attention Mechanism
2026 · NeurIPS · Accepted, Main Track
TL;DR: OLS is a similarity-based estimator in disguise: it optimizes the space in which observations are compared through inner products. Seen this way, its connection to attention—the mechanism underlying many large language models—becomes straightforward.
[ SSRN ] [ arXiv ] [ LinkedIn Discussion ]
16. From Reactive to Proactive Volatility Modeling with Hemisphere Neural Networks
2026 · Journal of Applied Econometrics · with Mikael Frenette and Karin Klieber
TL;DR: Hemisphere Neural Networks provide proactive volatility forecasts based on leading indicators when they can, and react to past prediction errors when they must. Their architecture separates the conditional mean and variance while allowing both to learn from shared information.
[ Journal ] [ SSRN ] [ arXiv ] [ Accepted Manuscript ] [ Slides ] [ Code ]
15. An Adaptive Moving Average for Macroeconomic Monitoring
2026 · Economics Letters · with Karin Klieber
TL;DR: Moving averages balance timeliness and stability, but the appropriate window changes with economic conditions. We introduce a simple adaptive estimator based on a random forest and show how it offers new insights into post-pandemic inflation trends.
[ Journal ] [ SSRN ] [ arXiv ] [ Slides ] [ Code ]
14. A Neural Phillips Curve and a Deep Output Gap
2025 · Journal of Business & Economic Statistics
TL;DR: Economic slack and inflation expectations can be estimated within an interpretable deep neural network. Using data through 2019Q4, the model captures the subsequent inflation upswing and attributes an important role to a large positive output gap, offering a different reading of the episode from standard filtering methods.
[ Journal ] [ SSRN ] [ Slides ] [ TimeWorld Slides ] [ SUERF Slides ]
13. Time-Varying Parameters as Ridge Regressions
2025 · International Journal of Forecasting
TL;DR: Time-varying parameter models can be estimated as ridge regressions, making computation, implementation, and tuning much simpler. I develop extensions and apply the approach to large local projections and VARs studying the evolution of Canadian monetary policy.
12. To Bag Is to Prune
2025 · Studies in Nonlinear Dynamics & Econometrics
TL;DR: How can random forests fit the training sample so closely without paying the usual price out of sample? I show how bagging and randomization implicitly prune a latent tree, providing a form of self-regularization. The same insight suggests ways to improve other greedy learning algorithms.
[ Journal ] [ arXiv ] [ Slides ]
11. The Macroeconomy as a Random Forest
2024 · Journal of Applied Econometrics
TL;DR: Small linear macroeconomic equations are useful, but their coefficients often change over time. Macroeconomic Random Forests let the data determine how those coefficients evolve, accommodating structural change and nonlinearities while preserving an interpretable economic equation.
[ Journal ] [ SSRN ] [ Slides ] [ SOFIE Seminar ] [ General Audience Penn Talk ]
10. Predicting September Arctic Sea Ice: A Multimodel Seasonal Skill Comparison
2024 · Bulletin of the American Meteorological Society · with Mitchell Bushuk and coauthors
TL;DR: How well can we predict September Arctic sea ice several months in advance? This collaborative study compares statistical and dynamical forecasting systems across the Arctic, documenting their skill at different horizons and geographic scales.
[ Journal ]
9. Slow-Growing Trees
2024 · Machine Learning for Econometrics and Related Topics, Springer
TL;DR: A single tree can match the predictive performance of a random forest if it grows slowly enough. I introduce a learning rate that tempers the tree’s greedy fitting procedure.
8. Assessing and Comparing Fixed-Target Forecasts of Arctic Sea Ice: Glide Charts for Feature-Engineered Linear Regression and Machine Learning Models
2023 · Energy Economics · with Francis X. Diebold and Maximilian Göbel
TL;DR: How far ahead can we accurately predict Arctic sea ice? We use “glide charts” to track forecast accuracy as the target date approaches and compare linear regression with machine learning. Predictability often improves only after a threshold is crossed, and machine learning offers particular gains around turning points in the annual sea ice cycle.
[ Journal ] [ arXiv ] [ Web App: September 2022 Forecasts ]
7. When Will Arctic Sea Ice Disappear? Projections of Area, Extent, Thickness, and Volume
2023 · Journal of Econometrics · with Francis X. Diebold, Glenn Rudebusch, Maximilian Göbel, and Boyuan Zhang
TL;DR: We jointly forecast Arctic sea ice area, extent, thickness, and volume, imposing the requirement that all four measures reach an ice-free state simultaneously. This constraint improves the coherence of long-run projections and provides a framework for quantifying uncertainty about when summer sea ice will disappear.
6. How Is Machine Learning Useful for Macroeconomic Forecasting?
2022 · Journal of Applied Econometrics · with Maxime Leroux, Dalibor Stevanovic, and Stéphane Surprenant
TL;DR: Which features of machine learning actually improve macroeconomic forecasts? We disentangle the contributions of nonlinearity, regularization, tuning, and loss functions. Nonparametric nonlinearity emerges as the main source of predictive gains.
5. Optimal Combination of Arctic Sea Ice Extent Measures: A Dynamic Factor Modeling Approach
2021 · International Journal of Forecasting · with Francis X. Diebold, Maximilian Göbel, Glenn Rudebusch, and Boyuan Zhang
TL;DR: Different measures of Arctic sea ice extent combine satellite observations with different processing algorithms, introducing measurement noise. We use a dynamic factor model to combine these measures and extract a common signal while accounting for their differing volatility and correlations.
4. Macroeconomic Data Transformations Matter
2021 · International Journal of Forecasting · with Maxime Leroux, Dalibor Stevanovic, and Stéphane Surprenant
TL;DR: Transforming predictors changes the explicit or implicit regularization of a machine learning model. We examine familiar transformations and propose new ones, showing how thoughtful preprocessing can substantially improve macroeconomic forecasts.
[ Journal ] [ arXiv ] [ Slides by Stéphane Surprenant ] [ Poster by Maxime Leroux ]
3. Arctic Amplification of Anthropogenic Forcing: A Vector Autoregressive Analysis
2021 · Journal of Climate · with Maximilian Göbel
TL;DR: Feedback loops in the Arctic can amplify the effects of CO₂ on sea ice. We introduce VARCTIC, a vector autoregression designed to study these interactions and their implications for future sea ice loss, providing an econometric complement to climate models.
2. Can Machine Learning Catch the COVID-19 Recession?
2021 · National Institute Economic Review · with Massimiliano Marcellino and Dalibor Stevanovic
TL;DR: What happens to machine learning forecasts when economic data move far beyond historical experience? Using a large U.K. macroeconomic dataset, we examine how different methods handle the COVID-19 recession and why the ability to extrapolate matters.
1. On Spurious Causality, CO₂, and Global Temperature
2021 · Econometrics · with Maximilian Göbel
TL;DR: We show how a popular information-flow approach can produce spurious causal conclusions when its underlying assumptions fail. We propose an alternative based on vector autoregressions and revisit the relationship between CO₂ and global temperature.
Policy Documents & Briefs
Reverse chronological order.
Tracing Economic Shocks Through History
May 2026 · SUERF Policy Brief No. 1461 · with Karin Klieber
TL;DR: We explain how local projection estimates can be traced back to the historical episodes that generate them. The approach helps policymakers assess whether an estimated response reflects broad historical evidence or depends heavily on a few exceptional events.
[ Policy Brief ]
Explaining Machine Learning-Based Forecasts Through Historical Analogies
February 2025 · SUERF Policy Brief No. 1089 · with Karin Klieber and Maximilian Göbel
TL;DR: Machine learning forecasts can be explained through the past observations they draw upon. We present a historical-analogy perspective that makes predictions more transparent and connects complex models to familiar economic episodes.
[ Policy Brief ]
Forecasting and Understanding US Inflation with Artificial Intelligence
October 2022 · SUERF Policy Brief No. 455
TL;DR: An interpretable neural network provides a new way to estimate economic slack and inflation expectations. This brief explains what the approach reveals about the inflation surge and how it can inform macroeconomic monitoring.
[ Policy Brief ]
Prévision de l’activité économique au Québec et au Canada à l’aide des méthodes Machine Learning
2020 · CIRANO Report 2020RP-18 · with Maxime Leroux, Dalibor Stevanovic, and Stéphane Surprenant
TL;DR: We apply machine learning methods to forecast key economic indicators for Québec and Canada. Nonlinear models deliver useful gains, and Macroeconomic Random Forests show that their forecasting benefits extend beyond U.S. data.
[ Report ]