Review Reports - Shadows of Demand: Uncovering Early Warning Signals of Private Consumption Declines in Romania

Round 1

Reviewer 1 Report

Comments and Suggestions for Authors

This paper presents a well-structured and methodologically sound approach to developing an early warning system for private consumption declines in Romania. The hybrid framework combining TVP-SV-VARX with machine learning classifiers is innovative and well-motivated. The study is policy-relevant and addresses a meaningful gap in the literature. However, there are areas where clarity, presentation, and contextualization can be improved.

Areas for Improvement

1.Introduction and Background

The introduction could better situate the study within the existing literature on early warning systems for consumption (as opposed to banking or currency crises).

More explicit motivation for the Romanian case would strengthen the contribution (e.g., data limitations, economic structure, policy context).

2.Results Presentation

Tables 1–5 are informative but could be better integrated into the narrative. For example, a brief interpretation of each table in the main text would aid readability.

Figure references (e.g., Figures 1–4) are missing from the submitted excerpt, making it difficult to assess their clarity and relevance. Please ensure all figures are included and clearly referenced.

3.Clarity and Language

Some sentences are overly complex or jargon-heavy, which may hinder comprehension. For example:

“The TVP-SV-VARX layer is able to isolate economically significant surprises (i.e., innovations measured by time-varying volatility) before they translate into spending weakness.”

This could be rephrased for clarity.

Consider simplifying technical language in the policy-oriented sections to improve accessibility for a broader audience.

4.Tables and Figures

Table 1: The column labels (e.g., “tvp1rf 28,3”) are cryptic. A legend or more descriptive labels would help.

Table 5: The recent 12-quarter probabilities are useful, but a brief summary of what this implies for policy readiness would enhance interpretability.

6.Limitations and Generalizability

The paper acknowledges that the framework is predictive, not causal, but could more explicitly discuss the implications of this for policy use.

A brief discussion on the generalizability of the method to other small open economies would be valuable.

7.Minor Suggestions

Consider adding a short subsection in Section 3 explaining the rationale behind the choice of hyperparameters (e.g., window length, lag order).

Ensure all acronyms (e.g., TVP-SV-VARX, XGBoost) are defined upon first use.

Author Response

Please see the attachment.

Author Response File: Author Response.docx

Reviewer 2 Report

Comments and Suggestions for Authors

I just have a minor suggestion.

A paragraph or two that summarizes the relevant Romanian literature would provide a little bit more context. For example, "monetary policy landscape, the yield curve's predictive power has not diminished, and that short-term forward spreads can, in the short term, outperform traditional multiple-short term spreads. These signals are essential for any consumption-oriented early warning system because they capture expectations about future policy and growth, as well as the financial channels through which private demand is stimulated," this is well established and cited, but citations here especially those done using Romania data would provide context to the manuscript.
Similar contextual citations would also be helpful.

Apart from these minor suggestion, the paper looks good.

Author Response

Please see the attachment.

Author Response File: Author Response.docx

Reviewer 3 Report

Comments and Suggestions for Authors

Thank you for this research. I had the opportunity to read it and found several suggestions that could improve this manuscript. My review report is divided into two sections: one is organized by page, and the second section is organized by section, highlighting the necessary improvements.

Page 2–3 (Introduction)

L55: “but its decline often precedes official statistics signaling a crisis” → “but declines often precede official crisis statistics.”
L58: Replace semicolon reference format [1; 2] → [1,2] throughout.
L65–69: Streamline: “Early warning systems have traditionally targeted currency and banking crises, showing that pre-crisis anomalies can be identified in real time.”
L70–77: Replace repetitive “essential”/“essentially.”
L78–92: Add commas for readability around subordinate clauses.
L89: “dynamic real financial transmission from consumption” → “dynamic real-financial transmission channels affecting consumption.”
L97–101: Sentence starting “The macroeconomic forecasting literature confirms…” is repetitive of L94–97; merge or delete one.
L102–109: Avoid excessive stacking of citations; group by theme.
L110–120: Correct “verifying stability across timeframes and model classes, often using simple ensemble confirmation rules” → “verify stability across timeframes and model classes, often via ensemble confirmation rules.”
L118: Consider shortening: “Therefore, demand-side early-warning systems complement banking-sector models.”

Page 4–8 (Literature Review)

General: The section is long and partially repetitive. Suggest splitting into (a) Financial Indicators; (b) Modeling Frameworks; (c) Policy Calibration.
L164–175: Check comma placement and hyphenate “short-term forward spreads.”
L178–184: Change “stand-ard” → “standard.”
L192–197: Too dense; break into two sentences for readability.
L199–208: Phrase “fragile growth perspective implies…”—add one citation context before introducing new term.
L209–217: Clarify difference between TVP-SV and dynamic-factor models (currently blurred).
L227–235: Replace “This helps distinguish ‘expected’ dynamics…” → “This distinction helps separate expected dynamics from risk-related anomalies.”
L236–245: Avoid repeating “tree-based ensembles” already covered earlier. Merge examples (14-16).
L256–273: Very strong paragraph; only edit for grammar: “scar-city” → “scarcity.”
L274–282: Clarify that WATCH–AMBER–RED design follows asymmetric-loss logic.
L283–291: “ex-change rate dynamics” → “exchange-rate dynamics.”
L293–300: Replace “long-term estimates” → “long-horizon estimates.”
L301–308: Add sentence transition linking ensemble post-processing to the hybrid approach.
L309–316: Fix “com-parative data show” → “Comparative data show.”
L317–324: “Bioinformatics and machine learning provide powerful tools…” → “These evaluation tools, adapted from bioinformatics…” (avoid off-topic reference).
L333–334: Remove repetition of “consumer warning problem”—replace with “consumption-monitoring challenge.”
L335–343: Simplify concluding sentence: “This integrated framework balances flexibility and interpretability.”

Page 9–12 (Materials & Methods + Results)

L372–381: Tighten: “tailored to the Romanian quarterly data environment” → “applied to Romanian quarterly macro-financial data.”
L383–389: Avoid “pri-vate”; fix hyphenation.
L394–396: Clarify whether transformation refers to “log-differencing” or “index rebasing.”
Equation block L399-401: Add commas and spacing per MDPI math style; define all symbols explicitly below.
L403-406: Replace �̂� placeholder (likely corrupted sigma/epsilon symbols).
L407: Fix: “Random Forest and XGBoost were selected…” (subject-verb).
L414–418: Replace “purge & embargo” → “purge and embargo periods.”
L418–419: The Brier formula lost clarity; rewrite as BS = (1/T) ∑ (pt – yt)².
L422–424: Fix typographical errors in indicators (𝑆𝑡_RF ≥ 𝜏_RED etc.); ensure consistent subscript formatting.
L487–491: Remove “constant recall of 0.83” repetition; already implied.

Page 13–15 (Discussion)

L537–544: Split into two sentences; very long.
L547–552: Replace “quarter-hour lead time” → “one-quarter lead time.”
L557–563: Simplify: “The anomaly layer (no_TVP) removal led to significant drops in PR-AUC and recall, validating H1.”
L564–571: Clarify dual role: decision engine (Random Forest) vs. calibration (Logit).
L573–580: “Sparse logistic plots” → “sparse logistic regression calibration plots.”
L581–587: Ensure clarity on escalation logic—add a diagram reference if available.
L588–593: Reword: “Confidence intervals for PR-AUC are wide—a common issue in rare-event settings—but stability across robustness checks mitigates concern.”
L594–602: Edit “positive warning of private consumption” → “reliable early warning of private consumption decline.”

Page 16–18 (Conclusions)

L603–606: “A median warning time of one quarter, with an operational point of zero” → “...of one quarter, effectively enabling proactive policymaking.”
L619–627 (MFA paragraph): Add source or explanation; currently unreferenced in earlier text.
L628–656: Reduce excessive operational detail; keep concise (≈ 150 words).

Section by Section review:

Title, Abstract, and Keywords

Abstract:

Remove filler phrases (“In this way,” “explicit data leakage protections”).
Replace colon fragments with full sentences.
Add a quantitative outcome (e.g., PR-AUC = 0.87) to give measurable value.

Introduction

Simplify phrasing; too many semicolons and redundant adjectives (“essential,” “dynamic,” etc.).
Combine repetitive statements about the importance of consumption indicators.
Replace multiple citation chains [1;2;3] with [1–3].
The paragraph on “macro-financial transmission” needs smoother linkage to the early-warning context.
Explicitly state the paper’s contribution in the final paragraph (e.g., “This study develops a quarterly early-warning model for Romania using explainable ensemble learning”).

Literature Review

Reorganize into subsections:

Macroeconomic and financial indicators
Forecasting and early-warning models
Policy and interpretability frameworks

Remove repeated discussion of tree-based ensembles.
Clarify differences among TVP-SV, dynamic-factor, and hybrid models—currently blurred.
Correct hyphen breaks (e.g., “stand-ard,” “scar-city,” “ex-change”).
Some citations appear without explanation (e.g., [29] – [31]); each must be contextualized.
Shorten concluding paragraph: one concise summary sentence plus a lead-in to Methods.

Materials and Methods

Early paragraph: define clearly sample period, frequency, and sources of each dataset.
Describe data transformations precisely (“log-differenced,” “normalized,” etc.).
Provide explicit definitions for all symbols in equations.
Replace corrupted symbols (σ̂, ε) and restore standard mathematical notation.
Clarify how Random Forest and XGBoost were tuned (grid, random, or Bayesian search).
Explain “purge and embargo” logic briefly for readers outside finance.
Include number of simulation replications and how stochastic variation was handled.
Justify evaluation metrics: why both PR-AUC and F1 were chosen.

Results

Replace comma decimals with periods (28,3 → 28.3).
In text, summarize key numerical differences rather than repeating entire tables.
Verify consistency between table values and those cited in text (some rounding mismatches).
Provide brief interpretation of each figure (“Figure 3 shows NSGA-III’s superior convergence”).

Discussion

Split long paragraphs into smaller analytic blocks (cause → implication).
Clarify difference between decision layer and calibration layer.
Discuss computational cost: how many hours or CPU cores were needed.
Acknowledge limitations (data scarcity, Romanian specificity, limited cross-validation).
Remove vague statements like “remarkably stable” and replace with quantifiable phrases (“variance < 0.02 across folds”).
Add short comparison with related early-warning studies (cite IMF or OECD if possible).

Conclusions

Streamline to 4–6 sentences.
Rephrase “median warning time of one quarter with operational point zero” → “median warning lead of one quarter, enabling proactive policy response.”
Avoid adding new methods or references here.
Insert one sentence on future work (e.g., expanding to multi-country panels or high-frequency data).

The following paragraph (studies) can strengthen your argumentation and can provide support in finding a research gap: Recent research underscores the growing role of artificial intelligence and data-driven modeling in developing robust forecasting and decision-support systems. Furthermore, research on digital supply chains and risk management confirms that adaptive, ensemble-driven architectures can enhance systemic resilience in dynamic environments .These studies collectively provide a strong methodological foundation for the present work’s focus on explainable, ensemble-based early-warning systems for private consumption forecasting.

Author Response

Please see the attachment.

Author Response File: Author Response.docx

Round 2

Reviewer 3 Report

Comments and Suggestions for Authors

The revised manuscript shows substantial improvement and presents a methodologically rigorous, clearly structured, and policy-relevant contribution. The authors integrate a TVP-SV-VARX structural filter with machine learning classifiers to build an early warning system for private consumption downturns in Romania. The revision demonstrates considerable refinement. The aims are clearer, the model pipeline is well-justified, and the evaluation methods follow best practices for rare-event prediction.