Yuying Sun

1 IntroductionExtensive empirical evidence suggests that the structure of economic relationships evolves over state variables, driven by factors such as policy reforms, business cycle fluctuations,...

This paper examines how firm‐level exposure to tax policy uncertainty affects corporate tax avoidance. Using US public firms from 2002 to 2020, we measure exposure with the tax policy risk (TPR) index of Hassan et al., which captures the share of quarterly earnings‐call discussion devoted to risks surrounding tax policy. Firms with greater exposure to tax policy uncertainty exhibit lower effective tax rates: a one‐standard‐deviation increase in TPR raises tax avoidance by 2%–3% relative to the sample mean, and the relationship survives instrumental‐variable estimation, propensity score matching, and lead‐lag tests. Consistent with real options reasoning, in which uncertainty about domestic tax policy raises the value of relocating income to less exposed jurisdictions, high‐TPR firms shift more income abroad and operate more tax‐haven subsidiaries. The response is concentrated among financially flexible firms, firms with earnings‐management expertise, and firms with market power, and shareholders value tax avoidance more highly when tax policy uncertainty is high. Our findings identify tax policy uncertainty as a distinct driver of corporate tax strategies.

Forecasting macroeconomic and financial variables under structural changes increasingly relies on a large pool of predictors, raising a fundamental question: should one impose sparsity, selecting a few key drivers, or density, pooling information across many? This paper develops a dynamic ensemble learning (DEL) framework that addresses this challenge by adaptively combining sparse and dense models through data-driven and time-varying combination weights, without committing to either structure ex ante. We establish the asymptotic optimality of the DEL estimator and the consistency of the combination weights for out-of-sample prediction, even when all candidate models are misspecified. Its effectiveness is demonstrated through extensive simulations and an application to U.S. inflation forecasting. Empirically, neither the sparse nor the dense specification dominates uniformly. By adapting to changes in their relative performance, DEL reduces forecast error by more than 16% relative to either constituent model. These improvements in predictive accuracy are accompanied by economic gains in the portfolio exercise. A long--short strategy based on stocks' exposures to DEL-based inflation innovations delivers the highest annualized return and Sharpe ratio among the competing methods.

Empiricists frequently encounter multiple plausible forecasts for high-frequency data, arising from disparate information sets, alternative statistical approaches, and substantial cross-sectional heterogeneity. We propose a transfer learning approach through high frequency ensemble learning (Transfer-HFEL) that optimally combines forecasts from heterogeneous populations within a state-space framework using a novel local L-fold cross validation criterion. We establish the asymptotic optimality and consistency of the Transfer-HFEL under regularity conditions, and evaluate its performance through simulation studies and an empirical application to ETF data. Empirically, Transfer-HFEL improves predictive accuracy relative to competing methods, delivering an annualized return and Sharpe ratio exceeding the Fama-French three-factor model by 79% and 65%, respectively. Interestingly, transferring idiosyncratic variation across assets yields greater forecasting improvements than transferring common systematic factors. One possible explanation is that model misspecification, time-varying risk premia and microstructure frictions contain valuable information overlooked by parameter-transferred methods. Rather than introducing noise, this idiosyncratic variation provides new insights into return predictability and portfolio performance in high-frequency markets.

1 IntroductionRegularized techniques have revolutionized high-dimensional data analysis across various domains, providing powerful tools for forecasting economic variables and financial outcomes (B...

To support China’s low-carbon transition and green growth, carbon–neutral bonds have been issued since February 2021. Our survey of carbon–neutral bond issuance by Chinese firms from inception to December 2024 shows that these initial carbon–neutral bonds have significantly lower yield spreads, implying lower financing costs than conventional bonds issued by the same firms and green bonds issued during the same period. An event study analysis shows that stock returns increased after the announcement of carbon–neutral bonds. Stock liquidity is identified as the possible mechanism through which carbon–neutral bond issuance may improve stock returns, although stock liquidity could be reduced if a firm with good environmental, social, and governance (ESG) ratings issues carbon–neutral bonds. Our findings that a negative carbon–neutral premium exists highlight the role that future carbon–neutral bonds could play in benefiting Chinese firms issuing such bonds.

We propose a novel interval-based transformer approach called DKformer to forecast quarterly carbon futures prices by treating interval-valued data as inseparable random sets. Our DKformer model leverages the attention mechanism to capture long-term dependencies and complex dynamics in the carbon market. The model integrates the European economic policy uncertainty (EEPU) index and the climate policy uncertainty (CPU) index, utilizing only lags of price values and EPU indices as explanatory variables to effectively capture the impact of policy shifts on carbon futures prices. Empirical tests demonstrate the strong performance of the model in predicting market turning points and responding to various policy change scenarios, including the implementation of the EU ETS Phase 4, the RePowerEU Plan, and the carbon border adjustment mechanism (CBAM). Furthermore, our findings indicate the outstanding performance of our method remains robust to various forecasting period, forecasting horizons, and sets of explanatory variables.

Most deep learning techniques are designed for point‐valued data, while interval‐valued data contain richer information over the same time period, especially during extreme events. To leverage this, we propose DKformer, a novel transformer architecture that models interval data as an inseparable set using the loss function. By combining transformers and loss, DKformer learns complex nonstationary and nonlinear patterns in interval data while accounting for uncertainties and dependencies. Compared to existing interval and point‐based benchmark models, DKformer significantly enhances crude oil price forecast accuracy. Remarkably, it demonstrates robust out‐of‐sample prediction performance across various forecast horizons. Furthermore, the effectiveness of DKformer is preserved even under extreme conditions, including Black Swan events. Our work highlights the potential for DKformer to enhance decision‐making in energy and finance, where robustness to uncertainty and the ability to handle extreme events hold paramount importance.

The extant literature on forecasting interval‐valued crude oil prices has predominantly focused on estimating the conditional mean, while little attention has been paid to the conditional distribution, which may offer more insightful information. This paper proposes a novel model free interval‐based Treeffuser approach to forecast the conditional distribution of interval‐valued crude oil prices. This approach involves estimating the score function in the inverse stochastic differential equation using gradient boosting trees and generating samples through a diffusion process. Empirical results demonstrate that the proposed approach outperforms classical interval‐based machine learning methods, particularly during extreme events. The superior performance is robust to various forecast horizons and estimation periods. Furthermore, we propose an interval‐based trading strategy that can effectively mitigate volatility and boost returns. Notably, our approach captures the significant impact of extreme events on minimum prices, revealing a dynamic pattern where the range of prices initially widens, and later the distribution of minimum prices flattens out.