## Reviewer 1 **Overall evaluation:** 1 (weak accept) ### Overall evaluation (comments) Summary: The authors focus on solving the GDP nowcasting problem. They notice the drawbacks of the existing DFM model: the lack of capturing economic uncertainties and the limitation of capturing irregular dynamics. To deal with these issues, the authors combine neural controlled differential equations with DFM. They conduct experiments on two real GDP datasets. The results compared with enough competitors demonstrate the effectiveness of the proposed method. Strong Points: S1. The authors clearly demonstrate the solved problem and their motivations. S2. The authors incorporate novel NCDE into classical DFM and propose a new nowcasting model, i.e., NCDENow. S3. The authors conduct experiments on real datasets from SouthKorea and the UK. NCDENow shows better performance compared with 6 baselines in terms of nowcasting accuracy. Weak Points: W1. The authors argue that VAE models are like a black-box, which makes it challenging to maintain DFM's strengths (around in line 289) which seems to have no evidence to support this point from this paper. So, if the authors can cite some related works or add a VAE-based model as baseline, the statement will be more accurate. W2. The combination of NCDE and DFM seems to be a naïve and straight solution to deal with the irregular time series forecasting problem. W3. How is the issue mentioned in introduction, i.e., the lack of capturing economic uncertainties such as sudden recessions or booms, solved in NCDENow? I have no idea about this part. In my understanding about the time series, the emergent event such as Cov-19 cannot be predictable. Thus, it will cause unexpected changes of the observed time series, which is also unpredictable if there is no clear signal to indicate the influence of the event. W4. The ablation study is not conducted. It seems that the factor extraction module can be solely used as predictor to forecast GDP. The influence of this module on the final prediction result is not verified. Moreover, there are lots of methods to obtain the latent factors, such as replacing factor extraction module with VAE. If the authors can conduct extensive experiments to demonstrate which solution is better, the whole manuscript will be more complete. ## Reviewer 2 **Overall evaluation:** 1 (weak accept) ### Overall evaluation (comments) The paper introduces a novel GDP nowcasting model that combines dynamic factor models and neural-controlled differential The paper introduces a novel GDP nowcasting model that combines dynamic factor models and neural-controlled differential equations. An experimental evaluation demonstrates its superior performance over six baselines, and the ability to provide effective insights into economic conditions for informed policy decisions. The paper reads well and the problem is relevant. The main limitation is the application to only two datasets. Since GDP and economic data are usually provided as open data from governments and academics, I would have expected an experimental analysis considering more countries/macro-areas. ## Reviewer 3 **Overall evaluation:** 0 (borderline paper) ### Overall evaluation (comments) This paper proposes a new framework called NCDENow for nowcasting Gross Domestic Product (GDP) growth. In this work, the authors investigate the traditional DFM model and try to fix its issues such as i) the lack of capturing economic uncertainties such as sudden recessions or booms; and ii) the limitation of capturing irregular dynamics from data of mixed frequency. To overcome the issues, authors proposed NCDENow that combines the strengths of Dynamic Factor Model (DFM) in extracting latent factors and Neural Controlled Differential Equations (NCDE) in handling irregular time series data. This paper claims that NCDENow outperforms existing methods in terms of nowcasting accuracy, especially in scenarios involving missing data and volatile economic conditions. Strengths of this work: 1. This paper propose the idea that combines strengths of DFM and NCDE: NCDENow leverages the advantages of both DFM and NCDE, potentially leading to more accurate prediction of irregular time series data for GDP. 2. The proposed method improve handling of irregular data since NCDENow's use of NCDE allows it to better handle missing data and sudden fluctuations in economic indicators, often encountered in real-world data. 3. Form the experiments results, the paper achieves that NCDENow outperforms existing baselines in terms of nowcasting accuracy, particularly in challenging scenarios. 4. Interpretability is a merit. NCDENow provides insights into the relationships between extracted factors and GDP growth, aiding policymakers in making informed decisions. Weaknesses of this work: 1. Some experimental details may be missing. The paper discuss the performance of proposed method but in the experiment setting hyperparameters and some setting details are not introduced. It will be better if authors can add these details. 2. Theoretical underpinnings: While the paper mentions well-posedness of NCDENow, a deeper discussion of the theoretical properties and justifications for the proposed framework would be beneficial. 3. The presentation of this work can be improved. ## Reviewer 4 **Overall evaluation:** -2 (reject) ### Overall evaluation (comments) This paper presents an innovative approach to predicting GDP and demonstrates commendable accuracy in forecasting the GDP of two specific countries. However, despite these strengths, the paper falls short in various critical aspects, leading to the overall recommendation of rejection. The strengths of the paper: It introduces a new method for GDP prediction. The authors successfully implement this methodology to forecast the GDP of two countries, showcasing promising performance. The innovative approach and the accurate prediction of GDP for these countries highlight the potential of the proposed method in economic forecasting. However, several weaknesses hinder the value to the WebConf audience. Firstly, the experiment section appears excessively restricted, failing to cater to the diverse interests of the audience. The paper should encompass broader applications such as predicting other time series in Economical / Financial domain. At the minimum, I suggest predicting GDP for multiple countries, particularly those in developing regions. Secondly, the paper's method description section lacks relevance beyond the realm of GDP prediction. It fails to discuss the potential implications or adaptability of the proposed method to other time series prediction tasks. By extending the discussion to encompass its applicability in diverse forecasting domains, the paper could significantly increase its value to a broader to encompass its applicability in diverse forecasting domains, the paper could significantly increase its value to a broader audience. Lastly, the absence of a rigorous analysis regarding the impact of individual modules within the proposed method is a notable weakness. The paper lacks ablation analysis or insights into the specific roles played by different modules. A deeper understanding of these components and their contributions to the forecasting process is crucial for comprehending the methodology's efficacy and potential enhancements. In conclusion, while the paper introduces a novel method for GDP forecasting and demonstrates promising results for two specific countries, its limitations in scope, applicability, and lack of comprehensive analysis diminish its relevance and appeal to the broader audience of the WebConf. ## Reviewer 5 **Overall evaluation:** 2 (accept) ### Overall evaluation (comments) The paper introduces a novel NCDENow framework for GDP forecasting, merging neural-controlled differential equations with dynamic factor models, an innovative approach particularly effective during economic uncertainties. While the methodology and experimental evaluation are thorough, showing effectiveness against baseline models, the paper could benefit from more detailed technical explanations. Its significance in economic forecasting under uncertain conditions is highlighted, with potential impact on policymaking and economic analysis, though some technical aspects, particularly NCDE integration, warrant further clarity. The paper does not delve deeply into the technical specifics of how Neural Controlled Differential Equations (NCDEs) are integrated with Dynamic Factor Models (DFMs). A more detailed explanation of the NCDEs' role and functioning within the framework would benefit readers, especially those unfamiliar with this approach. While the paper presents results using datasets from South Korea and the United Kingdom, it does not extensively address the generalizability and scalability of the proposed method to other regions or larger datasets. Understanding how the model performs across diverse economic contexts and scales is crucial for its applicability in global economic forecasting. The paper could enhance its analysis of how the proposed model handles sudden economic shocks and irregularities, such as those caused by unforeseen global events. While the COVID-19 pandemic is discussed, a broader examination of the model's robustness in various volatile scenarios would be beneficial.