Testing whether shopper voices improve a shop-level model
This study evaluates whether shopper survey information adds predictive value beyond a completed shopkeeper-only model of relative commercial activity. A conservative temporal linkage rule matched shoppers to shops using the same probable interviewer code, same calendar date, an absolute time gap of no more than 90 minutes, and exactly one eligible shopkeeper anchor. The rule linked 1,883 of 2,595 shopper interviews to 1,048 shops. Shopper responses were transformed and aggregated to the shop level. A supplementary comparison first compressed shoppers into four behavioral segments: Deliberate Planners, Impulse Explorers, Loyal Optimizers, and Overspenders. On 841 linked development shops, adding three nonredundant segment shares improved repeated-cross-validation RMSE from 0.7700 to 0.7676 and R² from 0.2634 to 0.2680. The richer 30-feature store-choice and desired-capabilities domain performed better, reaching RMSE 0.7610 and R² 0.2806; adding segments to that domain changed performance only marginally to RMSE 0.7604 and R² 0.2818. In the original governed evaluation, the parsimonious 109-feature Elastic Net—79 shopkeeper predictors plus those 30 store-choice predictors—was locked before the protected test. On 207 matched test shops, it reduced MAE by 4.69%, reduced RMSE by 4.01%, and increased R² from 0.2154 to 0.2771. Shopper segments therefore provide a compact predictive summary, while detailed store-choice information remains the stronger and better-validated shopper augmentation.
Research question and incremental design
Does aggregated shopper information improve prediction beyond shopkeeper information when both models are trained and evaluated on exactly the same shops?
The incremental design matters. Comparing a 1,291-shop shopkeeper model with a smaller linked-shop model would mix feature value with sample composition. This study therefore rebuilt a matched shopkeeper benchmark on the 841 linked development shops and compared every shopper-augmented candidate on shared outer and inner folds.
The shopper domains are also plausible in the regional retail setting. A Colombian neighborhood-store bulletin connects store scale with assortment, shopping frequency, and ticket patterns,[3] while a Brazilian bakery-sector study emphasizes customer experience, portfolio diversification, digital channels, and loyalty as competitive concerns.[4] These external sources provide context only; the incremental-value claim is estimated exclusively from the matched project sample.
Interpretation boundary
The estimand is predictive improvement among shops with accepted shopper links. It is not the causal effect of shopper preferences, a representative estimate for every retailer, or proof that adding a digital capability will raise commercial activity.
Linking shoppers to shops
The apparent store code was identified as a probable interviewer or fieldwork code. Because interviewers could visit several shops in one day, code-only linkage was not credible. Timing and uniqueness were therefore required.
| Linkage outcome | Shopper interviews | Share |
|---|---|---|
| Uniquely linked | 1,883 | 72.56% |
| Ambiguous within 90 minutes | 25 | 0.96% |
| Same-date candidate outside window | 344 | 13.26% |
| No same-interviewer/date candidate | 343 | 13.22% |
Coverage was broad but not complete: 81.18% of shops had at least one linked shopper. Linked shops had a median of two shoppers and a maximum of five, which limits the precision of shop-level shopper averages.
Shopper feature engineering
The shopper questionnaire contained 214 source columns organized into 57 question families. Administrative fields, linkage variables, conditional free text, ambiguous storage fields, and current transaction spending were excluded from the primary model. The spending field was reserved because it could behave as a contemporaneous proxy for the outcome.
| Shopper domain | Candidate features | Eligible features | Role in final model |
|---|---|---|---|
| Current trip and purchase behavior | 76 | 68 | Compared |
| General attitudes and psychographics | 54 | 27 | Not selected |
| Shopper and household composition | 27 | 24 | Not selected |
| Store choice and desired capabilities | 33 | 30 | Selected |
| Store experience and interaction | 29 | 25 | Not selected |
Numeric responses and ratings were averaged within shop. Multiple-response indicators became shop-level selection proportions. Categorical responses became proportions after rare development categories were pooled and one reference category was omitted. Documented refusal codes became analytic missingness and were median-imputed inside training folds.
Selected information
Alternative places considered, reasons for choosing the store, and the importance of payment or digital capabilities.
Excluded from the final augmentation
Broad psychographics, household composition, general store ratings, and many current-purchase indicators did not add stable enough signal for their complexity.
Shopper segments as a compact alternative
A supplementary analysis asked whether four interpretable shopper profiles could summarize useful shopper information with fewer inputs than the detailed questionnaire domains.
Each linked shopper was assigned to one of four mutually exclusive profiles: Deliberate Planners, Impulse Explorers, Loyal Optimizers, or Overspenders. Individual memberships were averaged within shop, producing segment shares such as 0, 0.5, or 1. Because the four shares sum to one, Loyal Optimizers was retained as the reference and the other three shares entered the model. The comparison used the same 841 linked development shops, fixed Elastic Net settings of alpha = 0.1L1 ratio = 0.1, identical 5-fold cross-validation splits repeated five times, and the same preprocessing for every specification.
| Repeated-CV specification | Shopper inputs | RMSE | MAE | R² | Result |
|---|---|---|---|---|---|
| Shopkeeper only | 0 | 0.7700 | 0.6125 | 0.2634 | Reference |
| Shopkeeper + segment shares | 3 | 0.7676 | 0.6087 | 0.2680 | Better than shopkeeper alone |
| Shopkeeper + store choice | 30 | 0.7610 | 0.6051 | 0.2806 | Most of the shopper gain |
| Shopkeeper + store choice + segments | 33 | 0.7604 | 0.6038 | 0.2818 | Small additional gain |
The segment-only model reduced RMSE by 0.31% and MAE by 0.62% relative to shopkeeper information alone, confirming that the profiles retain some predictive content. The store-choice domain reduced RMSE by 1.17% and MAE by 1.21% in the same comparison. The combined specification was numerically best, but its improvement beyond store choice alone was very small: 0.0006 RMSE, 0.0013 MAE, and 0.0012 R².
Matched evaluation design
Five shared outer folds estimated performance, while four inner folds selected Elastic Net settings. Every model saw the same rows. Shopkeeper preprocessing was preserved; all shopper features were numeric shop-level means or proportions and received training-fold median imputation followed by standardization.
The full comparison included a mean baseline, matched shopkeeper model, shopper-only model, five shopkeeper-plus-domain models, and the complete 253-feature augmented model. The protected test outcome remained unavailable until one fixed specification was recorded.
Development evidence and model lock
| Fixed specification | Features | OOF MAE | OOF RMSE | OOF R² | RMSE change vs shopkeeper |
|---|---|---|---|---|---|
| Shopkeeper only | 79 | 0.6007 | 0.7605 | 0.2902 | reference |
| + Current trip and purchase | 147 | 0.5978 | 0.7578 | 0.2952 | +0.35% |
| + Store choice and capabilities | 109 | 0.5913 | 0.7517 | 0.3066 | +1.16% |
| + All shopper domains | 253 | 0.5952 | 0.7532 | 0.3037 | +0.95% |
The fixed store-choice model improved pooled development MAE by 1.57% and RMSE by 1.16%, with lower RMSE in four of five folds. Development bootstrap intervals crossed zero, so the evidence was treated as promising rather than conclusive.
The locked candidate contained 109 predictors and fixed Elastic Net settings of alpha = 0.1L1 ratio = 0.1.
Protected-test incremental value
| Protected-test metric | Matched shopkeeper | Augmented | Change |
|---|---|---|---|
| MAE (log target) | 0.6894 | 0.6571 | 4.69% lower |
| RMSE (log target) | 0.8539 | 0.8197 | 4.01% lower |
| R² | 0.2154 | 0.2771 | +0.0617 |
| Pearson correlation | 0.4645 | 0.5271 | +0.0626 |
| Spearman correlation | 0.4098 | 0.4877 | +0.0779 |
| Median multiplicative error factor | 1.7734 | 1.7300 | 2.45% lower |
The augmented model produced a lower absolute error for 113 of 207 shops and a higher error for 94. Rank correlation increased from 0.410 to 0.488. The result is meaningful but modest: the shopper block improves the existing prediction rather than replacing shopkeeper operational information.
Paired uncertainty
A country-stratified paired bootstrap resampled the 207 protected-test shops 5,000 times. Each bootstrap sample compared both models on the same shops.
The test evidence is stronger than the development estimate, but the interval remains conditional on this linked-shop sample and the fitted models. It does not represent uncertainty from collecting an entirely new multi-country survey and repeating every modeling decision.
Interpretation and product implications
The evidence supports a layered use of shopper information: a shopkeeper foundation, a detailed shopper enhancement when linked responses are available, and a compact segment lens for communication.
Core shopkeeper diagnostic
Available wherever a shopkeeper completes the operational questionnaire. It remains the foundation for predicting relative shop performance across the broader shop sample.
Shopper-enhanced diagnostic
Available when recent shopper feedback can be linked and aggregated. Store alternatives, reasons for choosing the shop, and desired payment or digital capabilities provide the strongest tested shopper contribution.
The augmented module should report shopper coverage beside every prediction: how many linked shopper surveys informed the shop-level averages, when they were collected, and whether coverage meets a minimum operational threshold. That keeps a one-shopper estimate from looking as stable as an estimate based on several shopper voices.
The patterns can guide questions for shopkeepers—for example, what alternatives shoppers considered or which payment capabilities they wanted. They should be presented as predictive signals and opportunities for investigation, rather than automatic claims that changing one response will cause higher performance.
Limitations
- Only 72.56% of shopper interviews were uniquely linkable under the conservative rule.
- The matched population excludes shops without accepted shopper links and may differ from the full shopkeeper population.
- Most linked shops have only one or two shopper interviews, creating measurement noise in shop-level averages.
- The segment comparison is supplementary development evidence based on repeated cross-validation. It has not received the same protected-test evaluation as the locked store-choice model.
- The same-day cross-sectional design measures predictive association and does not establish intervention effects.
- The protected test had already been used for the broader shopkeeper study, although it remained unused for shopper linkage, feature selection, domain selection, and parameter choice.
- A production implementation would need a governed process for collecting, linking, and aggregating shopper responses.
Conclusion
Shopper information adds predictive value at two useful levels: broad behavioral segments provide a compact improvement, while detailed store-choice responses deliver the stronger and better-validated gain.
On the shared 841-shop development sample, adding segment shares to the shopkeeper model improved RMSE from 0.7700 to 0.7676, MAE from 0.6125 to 0.6087, and R² from 0.2634 to 0.2680. Store-choice information produced the larger improvement, reaching RMSE 0.7610, MAE 0.6051, and R² 0.2806. Combining store choice and segments was numerically best at RMSE 0.7604, MAE 0.6038, and R² 0.2818, but the extra gain over store choice alone was very small. The segment framework is therefore valuable as an understandable summary of shopper mix, while the detailed domain preserves more predictive detail.
The original governed evaluation provides the strongest evidence for practical use. On 207 protected-test shops, the locked store-choice model reduced MAE by 4.69%, reduced RMSE by 4.01%, and raised R² from 0.2154 to 0.2771 relative to the matched shopkeeper model. This supports a practical architecture for Latin American micro-retail: keep the shopkeeper model as the widely available foundation, add detailed shopper feedback when reliable links exist, and use shopper segments as a clear communication layer for understanding the customers around each shop.
References
- Eslava, M., Meléndez, M., Ulyssea, G., Urdaneta, N., & Flores, I. (2024). Firms and Inequality in Latin America. IDB Working Paper 1587, Inter-American Development Bank.
- Shmueli, G., Bruce, P. C., Gedeck, P., Yahav, I., & Patel, N. R. (2023). Machine Learning for Business Analytics: Concepts, Techniques, and Applications in R (2nd ed.). John Wiley & Sons.
- Tienda Registrada. (2024). Así se comportan las tiendas de barrio según su tamaño. Boletín 11, five-city Colombia summary through October 31, 2024.
- Euromonitor International. (2025). Mercado de Panificação no Brasil: Oportunidades e Desafios. Research commissioned by FIPAN.
- Project shopkeeper and shopper survey instruments, governed analytical artifacts, and group-defined shopper segment comparison (2024–2026). Private course-project materials; only aggregate findings are reported.