Research

This paper studies the consequences of misspecification in structural VARs and shows that conventional low-dimensional specifications are unable to recover the true structural shocks when relevant variables are omitted. In such environments, the reduced-form innovations become contaminated, making structural identification unreliable regardless of the imposed restrictions.

To overcome this informational deficiency, the paper introduces high-dimensional VARs estimated through LASSO regularisation. This approach enables the inclusion of a substantially richer information set while preserving model stability and interpretability. Using both theoretical recoverability measures derived from a DSGE model and extensive simulation exercises, I show that LASSO-VARs dramatically outperform traditional VARs: they recover the structural fiscal shock with markedly higher accuracy and generate impulse responses that remain close to the true underlying dynamics even in severely misspecified settings. Compared with factor-augmented VARs, the LASSO-VAR more effectively captures the key information needed for identification, while its performance is broadly comparable to that of Bayesian VARs in terms of shock recoverability and dynamic accuracy.

Overall, the results demonstrate that expanding the information set through data-driven regularisation provides a robust and effective solution to misspecification, establishing LASSO-based VARs as a powerful tool for credible structural analysis in high-dimensional macroeconomic environments.

In this project, I examine the relationship between fiscal policy and its macroeconomic effects, contributing to the ongoing debate on the size and sign of the fiscal multiplier. Traditional VAR models used for this purpose face substantial limitations when the underlying economic environment is high-dimensional, as they cannot adequately incorporate the large set of variables that influence fiscal dynamics. To address this issue, I propose the use of machine-learning techniques—specifically LASSO-based regularisation—to expand the information set while maintaining model stability and interpretability.

Through a simulation study and a preliminary empirical application, I show that the LASSO-VAR is markedly more effective than conventional VAR, FAVAR and Bayesian approaches in recovering the true responses of macroeconomic variables to government spending shocks. This methodology directly tackles the challenge of deriving reliable impulse responses in high-dimensional systems.

Importantly, when applied to real-world data, the LASSO-VAR yields fiscal multiplier estimates that are systematically higher than those obtained with standard low-dimensional models. This result indicates that once the relevant information set is appropriately enlarged, the dynamic effects of fiscal policy appear more expansionary than typically reported in the empirical literature. These findings underscore the value of regularised high-dimensional VARs for credible fiscal policy analysis and highlight the risk of understating fiscal effects when relying on traditional, more restrictive specifications.

This paper investigates which specific machine-learning features—such as nonlinearity, regularisation, or high-dimensional flexibility—drive forecasting gains in macroeconomic applications. While recent empirical evidence shows that ML methods often outperform traditional econometric models, the underlying sources of these improvements remain insufficiently understood. To address this gap, we design a controlled simulation framework in which synthetic datasets are generated from macroeconomic models exhibiting distinct forms of nonlinearity, including time-varying parameters and asymmetric adjustment frictions. We then compare the forecasting performance of linear benchmarks (AR, factor models, VAR) with nonlinear ML methods such as Random Forests.

Our findings reveal that forecasting gains from ML techniques arise predominantly when the data-generating process features smooth, structural nonlinearities—such as evolving parameters in TVP-VAR environments. In contrast, when nonlinearity stems from asymmetric or state-dependent rigidities of the kind produced by nonlinear DSGE models, ML methods do not outperform traditional linear approaches. Moreover, preliminary evidence indicates that these forecasting advantages are driven primarily by the ability of ML models to capture nonlinear patterns, rather than by their capacity to handle larger cross-sectional dimensionality.

Taken together, the results clarify the conditions under which ML methods provide genuine added value in macroeconomic forecasting, highlighting nonlinearity—not dimensionality—as the key feature behind their superior performance.