Interpretable Systematic Risk around the Clock

Referee: [Methodology] The methodology section provides no details on jump detection thresholds, window sizes, or data exclusion rules (listed as free parameters in the analysis). These choices directly determine which jumps enter the LLM classification step and therefore affect all downstream heterogeneity results and portfolio performance metrics.

Authors: We agree that the lack of specific details on these parameters is a shortcoming that needs to be addressed. In the revised manuscript, we will expand the methodology section to include a comprehensive description of the jump detection algorithm, specifying the exact thresholds used (such as the multiple of standard deviation for identifying jumps), the window sizes for volatility estimation, and all data exclusion rules (e.g., handling of market closures, low-volume periods, or overnight returns). Additionally, we will include robustness analyses showing how variations in these parameters affect the main results on risk premia heterogeneity and portfolio performance. revision: yes

Referee: [LLM classification procedure] No human validation, inter-rater checks, prompt-robustness tests, or alternative-LLM comparisons are reported for the LLM-based cause classification. Because the category-specific betas, risk premia, and the selection of the 'most strongly priced' category for the mimicking portfolio rest entirely on these labels, the lack of validation is load-bearing for the central claims of heterogeneity and out-of-sample Sharpe ratios.

Authors: This is a valid concern, as the reliability of the LLM classifications is central to our findings. Although the original manuscript did not include explicit validation steps, we recognize their importance. In the revision, we will add a new subsection on classification validation. This will include: (1) human annotation of a stratified random sample of 300 jump events by two independent annotators, with inter-rater agreement metrics (e.g., Cohen's kappa); (2) comparison of LLM outputs against these human labels to report accuracy, precision, and recall per category; (3) sensitivity tests to prompt variations and temperature settings; and (4) a comparison with classifications from an alternative model such as Llama-3 or GPT-4o. We believe these additions will substantiate the use of the LLM and bolster confidence in the heterogeneity results. revision: yes

Referee: [Portfolio construction] The annually rebalanced Fama-MacBeth mimicking portfolio is built directly from the same in-sample jump-risk premia estimates used to document heterogeneity; while the out-of-sample test mitigates some circularity, the choice of which category to isolate is informed by the very heterogeneity the paper measures, requiring explicit discussion of potential selection bias.

Authors: We thank the referee for highlighting this potential issue of selection bias. It is true that the choice of the macroeconomic news category for the mimicking portfolio is guided by the full-sample heterogeneity analysis. However, the annual rebalancing and out-of-sample evaluation are performed in a forward-looking manner using only information available at the time of rebalancing. To address the concern explicitly, we will revise the portfolio construction section to discuss the selection process in detail, including why macro news was chosen based on economic rationale and persistence. Furthermore, we will add robustness checks: (i) results for mimicking portfolios based on all categories, (ii) a version where the category is selected using only the first half of the sample and held fixed thereafter, and (iii) a discussion of how this affects the interpretation of the out-of-sample alphas and Sharpe ratios. We maintain that the real-time nature and out-of-sample testing provide substantial protection against overfitting, but agree that explicit discussion is warranted. revision: partial