Predicting Restaurant Survival During COVID-19 Using Yelp Data Open Access
Wang, Mingxi (Spring 2026)
Abstract
The COVID-19 pandemic presents a unique case study in which simultaneous and significant supply and demand shocks structurally undermined the operational viability of local restaurants. Recent methodological advancements and growing data availability now make it possible to identify more relevant leading indicators of business distress. This thesis examines the relationship between business attributes, customer sentiment, and restaurant closure rates in the context of federal subsidies such as the Paycheck Protection Program (PPP) disbursed through the CARES Act. To this end, I construct a predictive index that distinguishes the attributes reflecting a business’s active adaptation amid structural transformation, which I term the Business Resilience Index (BRI).
I select features through qualitative domain analysis and textual processing via transformer-based models, and I compare three modeling approaches: a transparent Composite Weighted Index (CWI) consisting of additive decompositions that capture the dimensions of business resilience for policy recommendations; a Cox proportional hazards model that estimates time-to-closure as a function of the full feature set; and a Model-Derived Probability Index (MDPI) employing machine learning algorithms with nonlinear transformations that output probabilistic estimates of business survival.
The MDPI achieves the strongest predictive performance (AUC = 0.84, Brier Score = 0.05), the Cox model provides moderate discrimination with interpretable hazard ratios (C-index = 0.73), and the CWI fails to achieve meaningful separation (AUC = 0.50). A sub-index logistic regression diagnostic reveals that roughly half the CWI’s performance gap can be attributed to fixed weighting and half to information loss during sub-index aggregation. Both the MDPI and the Cox model converge on the finding that dynamic behavioral signals dominate static operational characteristics in predicting survival.
The findings suggest that while a transparent index can provide interpretability, a more flexible model-driven approach is necessary to capture the complex dynamics of business resilience under crisis conditions. I conclude with a discussion of the tradeoffs between transparency and predictive performance in the context of policy applications and offer recommendations for future research on real-time economic measurement using platform data.
Table of Contents
Table of contents
1 Introduction 1
2 Literature Review 5
2.1 Empirical Foundations . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 5
2.2 Conceptual Framework and Research Gaps . . . . . . . . . . . . . . . . . . . 7
3 Data 10
3.1 Dataset Overview . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 11
3.2 Database Schema and Entity-Relationship Structure . . . . . . . . . . . . . 11
3.3 Geospatial Infrastructure . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 13
3.4 Data Pipeline . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 14
3.5 Analytical Sample and Temporal Design . . . . . . . . . . . . . . . . . . . . 14
4 Methodology 16
4.1 Sub-Index Formulas . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 16
4.1.1 Sentiment Quality Score (SQS) . . . . . . . . . . . . . . . . . . . . . 18
4.1.2 Engagement Resilience Score (ERS) . . . . . . . . . . . . . . . . . . 19
4.1.3 Operational Flexibility Score (OFS) . . . . . . . . . . . . . . . . . . 20
4.1.4 Track Record Score (TRS) . . . . . . . . . . . . . . . . . . . . . . . . 21
4.1.5 Context Adjusted Resilience (CAR) . . . . . . . . . . . . . . . . . . . 22
4.2 Composite Weight Index (CWI) . . . . . . . . . . . . . . . . . . . . . . . . . 23
4.3 Model-Derived Probability Index (MDPI) . . . . . . . . . . . . . . . . . . . 24
4.4 Accuracy Assessment Metrics . . . . . . . . . . . . . . . . . . . . . . . . . . 25
5 Results 26
5.1 Composite Weight Index (CWI) . . . . . . . . . . . . . . . . . . . . . . . . . 26
5.1.1 Weight Derivation Comparison . . . . . . . . . . . . . . . . . . . . . 26
5.1.2 CWI Discrimination and Calibration . . . . . . . . . . . . . . . . . . 27
5.1.3 Diagnostic: Sub-Index Logistic Regression . . . . . . . . . . . . . . . 29
5.2 Model-Derived Probability Index (MDPI) . . . . . . . . . . . . . . . . . . . 32
5.2.1 Model Selection . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 32
5.2.2 MDPI Discrimination and Calibration . . . . . . . . . . . . . . . . . 33
5.2.3 Feature Importance . . . . . . . . . . . . . . . . . . . . . . . . . . . 35
5.2.4 Sensitivity to Class Imbalance . . . . . . . . . . . . . . . . . . . . . . 36
5.3 Cox Proportional Hazards Model . . . . . . . . . . . . . . . . . . . . . . . . 37
5.4 Head-to-Head Comparison . . . . . . . . . . . . . . . . . . . . . . . . . . . . 41
6 Discussion 42
6.1 Limitations and Future Research . . . . . . . . . . . . . . . . . . . . . . . . 45
7 Conclusion 47
8 Appendix 49
8.1 Addressing Data Imbalance . . . . . . . . . . . . . . . . . . . . . . . . . . . 49
8.1.1 Model Comparison Under SMOTE . . . . . . . . . . . . . . . . . . . 49
8.1.2 Feature Importance Shift . . . . . . . . . . . . . . . . . . . . . . . . 50
8.1.3 Test Set Performance . . . . . . . . . . . . . . . . . . . . . . . . . . . 50
8.1.4 Summary . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 50
References 51
About this Honors Thesis
| School | |
|---|---|
| Department | |
| Degree | |
| Submission | |
| Language |
|
| Research Field | |
| Keyword | |
| Committee Chair / Thesis Advisor | |
| Committee Members |
Primary PDF
| Thumbnail | Title | Date Uploaded | Actions |
|---|---|---|---|
|
|
Predicting Restaurant Survival During COVID-19 Using Yelp Data () | 2026-04-08 09:45:51 -0400 |
|
Supplemental Files
| Thumbnail | Title | Date Uploaded | Actions |
|---|