## Reviewer 1 **Overall recommendation:** 1 (weak accept) **Quality of Presentation:** 4 (good) ### Detailed Comments to the Author(s) Strengths: [+] The presentation is easy to follow. Most figures are well-designed and deliver concrete information. [+] The limitations and future studies are discussed at the end of this manuscript. [+] The experimental section is solid, with 6 research questions and concrete evaluations. Major weaknesses: [-] In Table 4, the proposed model outperforms SoAs significantly, which is great. However, the baselines seem too easy and old. There are many advanced models for forecasting and nowcasting, with time series data or unstructured data. The authors need to justify why they are not compared. [-] In Line 512, authors mention there is a theoretical analysis of the proposed method, which can be linked to other methods. However, I didn’t find corresponding sections in the manuscript. Maybe authors mean the content around Line 362? [-] In Table 6, the proposed model did well during COVID-19. Explanations are needed. [-] The dataset statistics needs to be introduced for better understanding. For example, the partition of training, testing datasets. Differences and similarities between two datasets. [-] The writing can be further polished. For example, “want to” is typically not often used in scientific papers. [-] Also, there are some typos. For example, a period is missing in the caption of Table 5, Table 7, and Figure 5. ### Summary to support your recommendation This study nowcasts GDP under normal and abnormal periods. The problem is important and has great social impact. The authors design NCDENow, which integrates the advantage of DFM and NCDE. NCDENow outperforms baseline models significantly on datasets from South Korea and the UK. The main concerns are regarding the experimental design and baseline comparison. Also some important justification and explanation. Another concern is the missing of deployment or system implementation section, which are encourages by the Applied Research Track. ## Reviewer 2 **Overall recommendation:** 0 (borderline paper) **Quality of Presentation:** 5 (excellent) ### Detailed Comments to the Author(s) The author shows the model's performance in the UK and South Korea and releases the codes. Why not show more countries that are impacted by irregular events like COVID-19 as well? This would help to check if the model runs robustly. ### Summary to support your recommendation The paper introduces a novel nowcasting framework, NCDENow, that combines Dynamic Factor Models (DFM) with Neural Controlled Differential Equations (NCDE) to enhance GDP prediction. This hybrid approach aims to address the shortcomings of traditional DFMs by effectively capturing the dynamics of irregular time series data, providing significant improvements in nowcasting accuracy and interpretability compared to existing methods. The method's effectiveness is validated through empirical results on real-world GDP datasets from South Korea and the United Kingdom. ## Reviewer 3 **Overall recommendation:** 1 (weak accept) **Quality of Presentation:** 5 (excellent) ### Detailed Comments to the Author(s) Strong points: - The paper clearly identifies the challenges in GDP nowcasting, particularly the limitations of Dynamic Factor Models (DFMs). This sets a solid foundation for understanding the proposed solution. - The paper presents a novel approach to address the identified issues as the first research to integrate Neural Controlled Differential Equations (NCDEs) with DFMs for economic indicators. The architecture to integrate DFMs with NCDEs is simple to understand and computationally efficient for an initial attempt. - The paper is generally well-written and includes step-by-step numerical experiments to justify the authors' claims. Weak points: - While the performance is superior in scenarios with irregular data and volatile economic conditions, the datasets used for validation seem limited, as the authors also stated. For instance, the ability in the volatile conditions was demonstrated using only one instance: COVID-19 in the Korean economy. Including datasets from a more diverse set of countries could provide a better assessment of the model's generalizability across different economic conditions. - The selection of baselines is limited to variations of RNN-based models, and all RNN-based baselines rely on DFM for data preprocessing. Including a broader range of baseline models would provide a more comprehensive comparison. Question: Could the proposed methodology be applied to other macroeconomic metrics in similar situations, such as inflation forecasting, unemployment rate prediction, or industrial production nowcasting? Has this been verified? ### Summary to support your recommendation The paper offers a valuable contribution to the field of GDP nowcasting by identifying current challenges and proposing a novel framework that integrates Neural Controlled Differential Equations with Dynamic Factor Models. The paper is well-written, proposes initiating solution, and includes detailed numerical experiments. However, due to the limitations regarding dataset validation and the selection of baselines, I recommend a weak acceptance. ## Reviewer 4 **Overall recommendation:** 0 (borderline paper) **Quality of Presentation:** 3 (fair) ### Detailed Comments to the Author(s) This paper proposes a novel framework for GDP nowcasting using irregularly sampled macroeconomic indicators, based on the Neural Controlled Differential Equations framework for modeling latent factors. The authors compare their approach against DFM and its combination with different flavors of sequential models. Several research questions are explored by empirical evaluation on two datasets, including prediction performance, effects of missing data and ODE solvers used. Strong points: * Quite fresh and interesting idea to explore new generalization of both sequential and state-space models in the problem of nowcasting (or filtering) the GDP * Based on the error metrics the proposed method outperforms the alternatives (although a weak ones) by a high margin (though there are some unexpected patterns - more about it in the weak points section) * Easy to follow manuscript and detailed description of algorithm/modeling. Also contributed code and the dataset Weak points: * Weak comparison against the alternative approaches. Mostly using general sequence approaches in combination with DFM, instead of comparing against contemporary specialized GDP-nowcasting approaches mentioned in references like [26] and [40], or others non cited works like: - Richardson, Adam, Thomas van Florenstein Mulder, and Tuğrul Vehbi. "Nowcasting GDP using machine-learning algorithms: A real-time assessment." International Journal of Forecasting 37, no. 2 (2021): 941-948. - Eraslan, Sercan, and Maximilian Schröder. "Nowcasting GDP with a pool of factor models and a fast estimation algorithm." International Journal of Forecasting 39, no. 3 (2023): 1460-1476. * In the Table 4 for South Korea GDP nowcasting, it is a bit unexpected to see that among the baselines DFM has the smallest MSE, but has the highest MAPE. Some comments/discussion on that pattern would be appreciated, especially since that pattern is not present in the results for UK GDP. * The RQ2 is narrowed to just comparing DFM and NCDENow, while it would be interesting to see the results for all the alternative approaches. Also the conjecture/comment was that NCDENow seems “more robust to missing data” but from Table 5 it seems that 20% missing data is actually improving DFM performance for both South Korea and UK tasks (reduced MSE and MAPE compared to no missing data in Table 4.). And there was also no comments/discussion on that quite surprising pattern. ### Summary to support your recommendation Even though the idea seems novel and the results look interesting, it seems that the model is not deployed in the real world. ## Meta-review ### Meta-review The paper presents NCDENow, a novel nowcasting framework that integrates Dynamic Factor Models (DFM) with Neural Controlled Differential Equations (NCDE) to enhance GDP prediction. Despite some areas for improvement, the paper's strengths justify its acceptance. 1. Practical Relevance and Novelty: The paper addresses the critical and timely issue of GDP nowcasting, particularly under irregular and volatile economic conditions. The integration of NCDE with DFM is an innovative approach that significantly improves nowcasting accuracy and interpretability compared to existing methods. This novelty is well-recognized by multiple reviewers, emphasizing its potential impact on economic forecasting. 2. Comprehensive Experimental Design: The authors conducted thorough experiments using real-world GDP datasets from South Korea and the United Kingdom. The experimental design is robust, with six research questions and concrete evaluations. The results show that NCDENow significantly outperforms baseline models, demonstrating its effectiveness in both normal and abnormal periods, such as during COVID-19. 3. Clear Presentation and Methodology: The paper is generally well-written and easy to follow, with well-designed figures that effectively convey information. The detailed description of the algorithm and modeling approach further enhances the paper's clarity and accessibility. The inclusion of limitations and future studies at the end of the manuscript shows the authors' awareness of the study's boundaries and the potential for further research. Areas for Improvement: 1. Baseline Comparisons: Reviewers noted that the selection of baselines could be improved. The current baselines are somewhat outdated, and including more advanced models for forecasting and nowcasting would provide a more comprehensive comparison and validate the effectiveness of NCDENow more robustly. 2. Dataset Statistics and Generalization: The paper would benefit from a more detailed introduction of dataset statistics, including the partitioning of training and testing datasets and the differences and similarities between the datasets used. Expanding the validation to include more countries and diverse conditions affected by irregular events would provide a better assessment of the model's robustness and generalizability. 3. Writing and Formatting: Minor improvements in writing and formatting are needed. Polishing the language, correcting typos, and ensuring adherence to scientific writing standards would improve the overall presentation quality. Specific sections, such as captions for tables and figures, should be revised for clarity and completeness. 4. Broader Validation: While the paper demonstrates the method's effectiveness in South Korea and the UK, expanding the validation to include more countries and diverse economic conditions would provide stronger evidence of the model's generalizability and robustness.