Research 学术

We examine whether the US Securities Exchange Act of 1934 significantly stabilized the market by introducing mandatory disclosure of information. We argue that mandatory information disclosure can curb stock manipulation by enhancing transparency, thereby reducing excess stock volatility. After a comprehensive assessment of the voluntary disclosure practices of companies listed on the New York Stock Exchange before 1934, we find that those with poor disclosure practices experienced a significantly greater reduction in volatility after the implementation of the act compared with those with good disclosure practices. Further analysis reveals that the liquidity of these companies with poor disclosure practices also improved significantly more than that of companies with better disclosure, and the improvement in liquidity was linked to the decrease in their volatility. Given that one key purpose of the act’s legislators was to reduce excess market volatility, our findings provide empirical support for considering this legislative aim successful.

We propose a simple, linear-regression-based method for prediction of the time series of stock returns. The method achieves out-of-sample performances comparable to machine learning methods while having ignorable computational costs. The key component of the method is to integrate a straightforward cross-market factor screening into the iterated combination method proposed by Lin et al., (2018). Our empirical results on the U.S. stock market show that the method outperforms many state-of-the-art machine learning methods in certain periods. The method also exhibits greater utility gain and investment profits in most periods after considering transaction costs.

We investigate the performances of two methods of complete subset averaging—complete subset linear averaging (CSLA) and complete subset quantile averaging (CSQA)—on the problem of corporate bond return prediction. We find that the two methods are overwhelmingly better than univariate linear regression and simple forecast combination. Meanwhile, CSQA is better than CSLA in most cases. For practical implementation, we also provide discussions on the selection of the hyperparameter k when applying these complete subset averaging methods.

We employ an ensemble learning approach, “stacking”, to refine and combine a variety of linear and nonlinear individual stock return prediction models. In an application of forecasting U.S. market excess return, stacking with a simple structure can outperform the traditional historical mean benchmark, Mallows model averaging, simple combination forecast, complete subset regression, combination elastic net forecast, and several other models in terms of both in- and out-of-sample performance measures on a consistent basis. More importantly, we find that the out-of-sample gains of stacking are especially evident during extreme downside market movements. Overall, stacking can generate substantive improvements in market excess return predictability.

This paper investigates how the COVID-19 pandemic affects the connectedness network of stock market volatility in 19 economies around the world. Our method builds on the Diebold-Yilmaz volatility network model to construct the volatility spillover index, and uses lag sparse group LASSO to accommodate the high-dimensional system. We find that the outbreak of the COVID-19 pandemic strengthens the overall volatility connectedness, and the global connectedness level remains high throughout 2020. In particular, connections across different continents have become stronger during this period. However, China is shown to be disconnected from the global volatility connectedness network until late November 2020. We find evidence that China is not the main source of volatility spillover during the COVID-19 pandemic.

In May 2025, China’s fund regulator mandated a performance-fee schedule for newly issued active equity funds with an unusual geometry: the penalty for trailing the benchmark by three percentage points is twice the reward for beating it by six, and the fee settles lot by lot over each investor’s own holding window. The design moves past the symmetric fulcrum fee, the only performance-fee contract U.S. law has permitted since 1970, to the penalty side. We model such penalty-heavy, vintage-contingent contracts. The schedule taxes tracking variance: managers below a skill threshold optimally closet-index, and the within-year risk response is non-monotone—gambling zones near the downgrade and upgrade lines, an incentive dead zone in deeper losses. Lot-by-lot settlement makes money age a state variable: young lots face a penalty-tilted lottery, while window averaging launders luck out of maturing lots. The mandate requires 60% of new active equity issues at leading companies to adopt the contract, with fixed-fee products continuing alongside; under this partial mandate, fund companies assign floating contracts to their strongest managers, an endorsement reversal visible in the first cohort’s appointment data. The welfare account is correspondingly asymmetric: the reform’s robust effect is protection, not stimulation—investors of mediocre funds are refunded the noise tax as their managers closet, while the incentive effect on genuinely skilled funds is limited and preference-dependent. First-year evidence from the inaugural cohort is suggestive.

近期，经济学研究的方法论问题重新在中文经济学界引起关注。陆铭教授最近 把经济学研究中的一种现象概括为“过度模型化”：模型构建未能很好纳入现实情 境，出现“为了让模型成立而忽略现实”；过度追求方法复杂性和精确性，出现“为 了让方法成立而牺牲问题”(陆铭, 2026)。本文赞同对这一现象的观察，但认为其所 指向的并非中国经济学近期才有的特殊现象，而是经济学长期面对的基本问题。本 文认为，这一问题的关键不在于“模型化”，而在于如何判断“过度”。为此，本文 提出行为公设、制度框架与经验校准构成的三元框架：经济解释需要说明行动者如 何形成目标、信念和选择方式，行动者处于何种规则和收益后果之下，以及相关机 制在经验上具有多大量级和适用边界。所谓“过度”，可以理解为三者之间的失衡 或断裂。本文以李嘉图式抽象及熊彼特对其“李嘉图恶习”的批评、门格尔与历史 学派之争、米塞斯与弗里德曼围绕先验演绎和经验检验的分歧为线索，说明关于模 型、制度语境和经验证据之间关系的争论贯穿经济学方法史。随后，本文讨论现代 因果识别进入传统历史研究时引发的争议，说明经验工具进入既有命题时仍需面对 机制解释、制度语境和知识增量问题。最后，本文以拍卖理论和期权定价为例说明， 模型复杂本身并不必然导致“过度”。

We propose a new method for forecasting stock market returns in a data-rich environment: the factor-augmented sum-of-the-parts (FA-SOP) approach. Rather than predicting returns directly, FA-SOP decomposes them into three components—dividend-price ratio, earnings growth, and price-earnings ratio growth (gm)—and models each separately. We emphasize that gm is a more promising target for forecasting due to its stronger connection with macroeconomic conditions and greater variability over time. FA-SOP forecasts gm using latent macroeconomic factors extracted from high-dimensional data, capturing the underlying state of the economy while avoiding overfitting. Applied to S&P 500 returns from 1960 to 2022, FA-SOP outperforms predictive regressions, factor-augmented regressions, and traditional decomposition approaches, yielding robust out-ofsample gains in both statistical and economic terms. Simulations based on a present-value model further show that FA-SOP’s advantage stems from its ability to track the true data-generating process more closely. Our results highlight the value of decomposing returns and focusing on components that are more predictably linked to economic fundamentals.

This study employs an ensemble machine learning method, known as “Stacking," to forecast corporate bond returns. We find that the Stacking method introduces new features into combination forecasts, increasing the predictive model’s power and generating higher statistical and economic gains across bond ratings and maturities. Moreover, the method is efficient for tackling high dimensionality and achieves the best result when using predictors from corporate bond, Treasury, and stock markets jointly. While the overall performance of different Stacking models is satisfactory, simpler Stacking models appear to outperform others and generate optimal forecasts.

We contrast continuous magnitude estimation with discrete directional classification in equity premium prediction. Through a symmetric, single-pass out-of-sample evaluation of traditional econometric and machine learning algorithms, we document a distinct divergence in predictive performance. While continuous models struggle with low signal-to-noise ratios and generally fail to outperform the historical mean, discrete classifiers consistently reveal statistically significant predictability and generate substantial economic utility. We demonstrate that this divergence stems from the inherent sensitivity of continuous models to magnitude noise. Under continuous estimation, algorithms face a structural dilemma: they either overreact to unforecastable extreme shocks or resort to excessive shrinkage, both of which severely limit their market-timing ability during downturns. Conversely, discarding magnitude estimation acts as a form of structural regularization. This transformation frees nonlinear algorithms from extrapolating non-stationary macroeconomic trends, allowing them to utilize persistent high-frequency risk signals to execute timely market exits. Ultimately, our findings suggest that the fundamental choice of the predictive paradigm exerts a first-order impact that outweighs specific algorithmic sophistication.

Without a state, formal law, or police, how do people spontaneously begin to follow a leader (authority) and respect others’ possessions (property rights)? We study the micro-foundations of these two institutions using a controlled computational laboratory. By pre-generating and freezing all random events in advance, we place traditional rule-based agents (ABM) and modern Large Language Model agents (DeepSeek, GPT-6, and Gemini) into identical parallel worlds and directly compare their choices. Our experiments reveal four clear findings. First, shared records affect authority and property in fundamentally different ways: in group decision-making, accurate records of past performance are sufficient on their own to make agents voluntarily follow the most capable leader; however, giving someone an official title without a real track record only creates blind following around a name. Second, in conflicts over scarce resources, shared records are necessary but not sufficient for property rights: knowing who claimed a resource first does not stop others from grabbing it unless agents also hold a behavioral norm to back down or face lasting exclusion for intruding. Third, AI agents display a striking gap between words and actions: in resource conflicts, Large Language Models verbally recognize others’ claims 75%–100% of the time, yet still step in to grab the resource over 95% of the time, triggering costly cycles of mutual punishment unless they are guided to honor their word or allowed to vote after experiencing conflict. Fourth, once established, authority built on accumulated past reputation is slow to change when a leader loses competence, whereas property arrangements with clear rules for challenging and transferring claims adapt quickly. Together, these findings clarify how shared information and behavioral restraint jointly give rise to spontaneous social order.

Can machine learning autonomously discover cross-sectional asset pricing factors without degenerating into an uninterpretable black box? We examine this question by designing a formulaic reinforcement learning framework that searches for explicit return predictors in Reverse Polish Notation (RPN). Operating on 1,712,870 CRSP/Compustat firm-month observations over 1965–2024 under strict formulaic, numerical, unit, and horizon constraints, a Maskable PPO agent is rewarded by panel Fama-MacBeth t-statistics controlling for standard benchmark factors. By benchmarking reinforcement learning (RL) against both uniform random sampling (RS) within the same grammar and canonical two-variable data-mining operators (CLZ), our empirical analysis reveals four main findings. First, representation space dominates search optimization: while the formulaic RPN framework outperforms canonical literature operators by a wide margin (15.75% vs. 11.50% in univariate sort hurdles), the incremental advantage of reinforcement learning over random sampling is narrow (12.75% vs. 12.50% in Fama-MacBeth hurdles), concentrated in earlier decades, and completely vanishes after 2005. Second, machine-learned factors achieve substantial dimension reduction: portfolios constructed from only 200 RL-mined signals via neural networks generate an equal-weighted monthly return of 1.22% with an annual Sharpe ratio of 1.20, matching the predictive performance of brute-force universes containing over 18,000 raw characteristics. Third, structural dissection reveals that the agent gravitates toward ratio-based expressions, scaling profitability, accrual, and momentum metrics by economic scale bases. Fourth, when subjected to a multi-agent Large Language Model debate as an ex-post interpretability hurdle, RL signals survive at two to three times the rate of random expressions (1.50%–3.50% vs. 1.00%), yet lag far behind published human factors (49.64%), demonstrating that statistical significance does not equate to economic plausibility. Together, these results delineate the exact boundaries of AI-driven factor discovery, showing that while symbolic algorithms excel at compressing cross-sectional predictability, human domain knowledge remains essential for defining the economic grammar and enforcing theoretical plausibility.