A practical predictive model for Latin American micro-trade
This study develops a shopkeeper-based model that successfully predicts differences in commercial activity among Latin American micro-retail businesses. Using 1,291 interviews from Argentina, Brazil, Colombia, Mexico, and Peru, it turns a detailed account of shop operations, practices, and services into a common framework for comparison. Exploratory analysis describes the businesses and their priorities, while a carefully prepared activity measure combines customer traffic and spending relative to shops of comparable country, format, and size.
A 79-feature Elastic Net, selected through nested evaluation of four model families, achieved pooled development R² of 0.2992 and protected-test R² of 0.2540. On the 259 shops reserved for final evaluation, it reduced MAE by 10.18% and RMSE by 13.63% compared with the mean-target baseline, with MAE of 0.6915 and RMSE of 0.8530 on the log scale. Stable predictive patterns linked relative activity with staffing, digital connectivity, financial access, transaction services, product mix, and modernization. The study establishes a tested foundation for peer benchmarking, understandable business profiles, and more focused conversations between shopkeepers and those supporting them. It demonstrates how survey research can produce useful predictive tools for a diverse and underexplored area of regional commerce.
Understanding differences between neighborhood shops
A shopkeeper learns a great deal through daily trading: what regular customers ask for, which products move, when the shop gets busy, and how to make a small team work well together. That experience shapes decisions about stock, service, prices, and opening hours. This study starts with respect for that knowledge and asks how information from other shops can add a useful point of reference. Its purpose is to give Latin American micro-trade closer research attention and turn what shopkeepers share into something useful for them.
A helpful point of reference needs to fit the business. A bakery, a pharmacy, and a minimarket sell different products, serve customers in different ways, and operate at different scales. For an owner of a small grocery shop, a comparison that recognizes its scale and format is more informative than the sales level of a much larger minimarket. Across countries, spending also needs a common basis for comparison. Taking these circumstances into account gives the analysis a practical connection to the owner's situation.
The survey brings these individual experiences together. By describing the shops, finding shared patterns, and testing predictions, the study develops a common reference that can complement an owner's local knowledge. One question connects the work:
Can shopkeeper survey data provide a useful estimate of commercial activity relative to similar shops in Latin America?
The regional literature helps explain why this question matters. Eslava and colleagues describe the large role of tiny firms in Latin American employment and the barriers they face to productivity and growth.[1] A Colombian industry bulletin documents differences in assortment, purchasing patterns, turnover, and average ticket across store sizes.[2] Research on Brazilian bakeries highlights staffing constraints, rising input costs, competition, diversification, and technology integration.[3] Together, these sources place the analysis within the operating challenges of regional micro-trade. The project survey supplies the shop-level evidence used to develop the model.
The study answers this question by following the path from lived business conditions to useful predictions: first understanding the shops, then creating a comparable activity measure, preparing the predictors, and evaluating the model on held-out businesses. Its practical purpose is to give owners and advisers a clearer view of each shop's position among its peers and the operating patterns associated with activity.
Building confidence through independent evaluation
The study follows two connected stages. The first establishes what the survey describes and how its answers should be interpreted. Basic cleaning and semantic checks support exploratory summaries of the full sample. These summaries precede target construction and provide the context for the analysis that follows.
The second stage develops and evaluates predictions. For this stage, 1,032 interviews form the development sample and 259 form the protected test sample. The development sample supports choices about the activity measure, eligible features, model settings, and diagnostic thresholds. The protected test provides the final assessment after those choices have been fixed.
The separation strengthens the evidence for predictive usefulness. Rules for capping unusual values, filling missing ratings, constructing the target, and preparing predictors are learned from training shops and applied to held-out shops. The final test therefore evaluates the model's ability to carry what it has learned to additional businesses. This makes the reported performance a more convincing achievement than a fit to the development data alone.[5]
With that distinction established, the first step is to understand the businesses represented in the survey.
Exploratory data analysis: getting to know the shops
The survey first reveals the setting in which prediction will take place: who was interviewed, how the shops operated, and how respondents described their business conditions. It also establishes how the questionnaire's answers can be translated into useful analytical variables.
Categorical responses were summarized with counts and percentages. Numeric profiles used medians so that a few unusually large reports would not dominate the description. This stage builds the picture of daily business life that informs the activity measure and the later interpretation of predictions.
A regional picture across five countries and shop formats
The 1,291 interviews cover five countries and five shop formats. Brazil, Colombia, and Mexico account for most responses; Argentina and Peru each contribute 76. Store formats are more evenly represented, with minimarkets and bakeries forming the largest groups.
| Dimension | Group | Shops | Percent |
|---|---|---|---|
| Country | Brazil | 381 | 29.5% |
| Country | Mexico | 379 | 29.4% |
| Country | Colombia | 379 | 29.4% |
| Country | Argentina | 76 | 5.9% |
| Country | Peru | 76 | 5.9% |
| Store type | Minimarket | 310 | 24.0% |
| Store type | Bakery | 301 | 23.3% |
| Store type | Pharmacy | 228 | 17.7% |
| Store type | Liquor or tobacco store | 227 | 17.6% |
| Store type | Grocery store | 225 | 17.4% |
The mix of countries and formats gives the study a regional perspective while preserving distinct local contexts. The three largest country samples contribute most to the overall results. Later country-level evaluation makes those differences visible and helps identify opportunities for further local research.
The median respondent was 40 years old, the median business had operated for seven years, and the median respondent had five years of experience in that business. The reported workforce was typically small: the median was three people, including the respondent. Median daily customer traffic was 70. Together, these figures describe established businesses in which relatively small teams sustain an active flow of customers.
Daily operations center on availability and customer contact
Long opening schedules were common. A total of 1,097 shops, or 85.0%, reported opening every day of the week. Hybrid service—customers retrieve some products and request others—was the most frequent format, followed by staff-assisted service. Complete self-service was less common.
For the people running these shops, the opening schedule means arranging cover, replenishing stock, and serving customers across the week. In a staff-assisted shop, someone must also be available to retrieve products and answer requests. Staffing, facilities, and service format therefore belong in the same account of daily operations. Reading headcount alongside service format and opening schedules provides a fuller account of the people and organization behind the business.
Stable businesses, optimistic expectations, and difficult growth
Most respondents described their current business as stable. This group contained 723 shops, compared with 326 reporting growth and 174 reporting that they were doing very well or considering expansion. The two smallest groups each contained 34 shops: businesses still being tested and businesses struggling or considering closure.
The picture changes when respondents look ahead. Recent sales were reported as better or much better by 634 shops (49.1%), but 972 (75.3%) expected better or much better sales in the following year. At the same time, 742 (57.5%) described growth as difficult or very difficult. The responses capture a recognizable tension across the sample: hoping for better trading while seeing growth as difficult. Reading the two findings together gives a fuller picture of the business outlook.
When asked about profitability, respondents most often selected customer service, good prices, and product variety. These concerns connect directly to everyday decisions: taking time to help a customer, setting an affordable price, or choosing which products deserve limited shelf space. The responses bring shopkeepers' own priorities into the study and give the later analysis a clear connection to the choices they make each day.
The questionnaire also has to be understood
The exploratory process also established the meaning of stored responses. A blank could indicate a routed follow-up, while a negative code could record the absence of a facility. Unusually large customer counts were flagged for review. Resolving these distinctions turned the questionnaire into a more faithful representation of shopkeepers' answers.
The original Spanish responses were retained as the authority where translated labels omitted meaning. Conditional blanks, inactive response slots, and refusals were handled according to their questions. Unusual values were flagged and preserved for review, so later transformations could be traced to an explicit decision.
| Finding | Observed evidence | Analytical response |
|---|---|---|
| Translation consistency | Three meaningful omissions; 28 numeric fields showed no value differences. | Use original Spanish codes as authoritative and preserve bilingual labels separately. |
| Formatting differences | 217 text or mixed fields and 9,431 cells required normalization. | Normalize formatting without changing substantive response categories. |
| Conditional blanks | 3,774 genuine blanks concentrated in conditional “other” fields. | Retain structural missingness instead of treating it as ordinary nonresponse. |
| Supplier-rating routing | 117 seven-item blocks skipped; 122 straightlined; 84 used all tens. | Preserve applicability indicators and audit response-pattern sensitivity. |
| Unusual numeric reports | Ten daily-customer values of 999 and one shop-age value of 99. | Flag rather than silently overwrite; use robust target scaling learned in training data. |
Defining commercial activity on a comparable scale
A central achievement of the study is turning several familiar shopkeeper reports into a common measure of commercial activity. Days open, daily customer traffic, and customer spending bands describe complementary parts of trading: how often the shop serves customers, how many it serves, and their typical spending. Bringing these elements together creates a model target variable that represents relative commercial activity across the surveyed shops.
Combining these inputs requires explicit choices about unusual values, currencies, and shop structure. The construction follows the OECD/JRC approach of stating the measure's purpose, explaining its components and normalization, and examining sensitivity to the choices made.[4] The calculation proceeds in two stages:
+ [log(spending-band midpoint) − country median log spending]
expected log activity = Ridge(country, store type, store size)
SPI (model target variable) = observed log activity − expected log activity
First, days open and reported customers provide a traffic measure. Customer counts are capped at the training sample's 95th percentile to limit the influence of extreme reports. Spending bands are represented by documented midpoints. Their logged values are centered within country, putting spending on a relative scale suited to comparisons across the five national contexts.
Second, a Ridge regression estimates expected log activity from country, store type, and store size. Subtracting this structural expectation gives the model's target variable, SPI (Store Performance Index). A shop receives a positive value when its reported activity is above the model's expectation for those characteristics, and a negative value when it is below.
Reading the target
Zero represents activity at the structural expectation. A target near 0.693 corresponds to twice that expected activity; a target near −0.693 corresponds to half. For example, a shop can have more customers than another shop yet have a lower relative score if its size and format imply a higher expected level of activity. This gives each owner a reference point grounded in the structural context of their business.
Logarithms also address the long upper tails in reported activity. Development diagnostics showed skewness of 3.71 for traffic and 5.448 for the ordinal activity construction, falling to 0.162 and 0.227 after log transformation. The traffic-based and spending-augmented activity constructions had a Spearman correlation of 0.8863. Traffic therefore strongly influenced their rankings, while spending still changed some shops' positions.
The resulting measure makes the central comparison possible: how a shop's activity stands relative to structurally similar businesses. It provides a transparent bridge between everyday survey responses and a statistical prediction target. The next step is to prepare the shop characteristics that will help predict that position.
Turning survey responses into usable predictors
The data-quality findings now become preparation rules. Responses are translated into analytical fields according to the meaning of each question. Single-response answers remain categories, multiple-response options become separate binary indicators, and counts are converted only after their special codes have been checked.
For example, an absence code in an aisle or refrigerator question becomes a zero count, while a separate presence indicator records whether the facility exists. A skipped supplier rating receives a separate treatment that preserves the questionnaire's routing and the meaning of an applicable response. These distinctions help ensure that the prepared data represent what respondents actually reported.
The prepared table contained 118 documented features: 99 numeric and 19 categorical. Seven supplier-rating columns retained 616 structural missing cells in development and 203 in test; all other prepared cells passed the missingness checks. A development-only eligibility screen then removed three rare binary features and one constant indicator, leaving 114 candidates.
| Feature family | Prepared features | Examples |
|---|---|---|
| Direct business profile | 22 | staffing, work hours, orientation to technology |
| Physical storefront | 7 | barriers, storage, shelves, aisles, refrigerators |
| Technology + financial services | 12 | internet, cash register, bank account, credit |
| Products + services | 34 | product assortment, utility payment, cash withdrawal |
| Payment + sales strategies | 14 | accepted payments and selling practices |
| Modernization | 12 | implemented and competitor modernization |
| Supplier relationships | 17 | applicability, ratings, relationship view |
The feature families cover the business from several angles: people and working practices, physical facilities, technology, products, services, modernization, and supplier relationships. The storefront example below illustrates how a reported physical condition can be represented as a categorical predictor. It also shows that the open storefront pattern seen in the full sample is present in the development sample.
Preparing a credible prediction task
The predictor set focuses on business characteristics that can provide additional information about relative activity. Administrative identifiers, target ingredients such as traffic and spending, and structural-benchmark outputs were kept separate. Outcome-adjacent measures were reserved for validation. This gives the model a substantive task: learn the connection between operating characteristics and the independently constructed activity measure.
For the remaining predictors, preparation continued within each training fold. Categorical encoding, numeric scaling, the log transformation of one highly skewed count, and supplier-rating imputation were fitted on training shops and then applied to validation shops. Feature names were documented in business language, and the development and test data were required to share the same schema. This produced a consistent set of inputs for comparing models.
Choosing a model through repeated held-out comparisons
Model development progressively increased the useful information captured from the survey. The mean training target provided a common baseline, followed by an initial 22-feature direct-question model. Expanding to 114 eligible candidates materially improved Ridge performance. Analysis of feature domains and questionnaire burden then produced a focused set of 79 predictors, balancing survey coverage with a more manageable specification.
Four model families were then compared on those inputs. Ridge and Elastic Net provide regularized linear predictions; Random Forest and histogram gradient boosting allow more flexible relationships. Their performance was assessed on the same held-out shops, using RMSE on the log target to choose model settings.
Separating model tuning from evaluation
The comparison used five outer folds and four inner folds. Each outer fold set aside part of the development sample for evaluation. The remaining shops were divided again into four inner folds to choose hyperparameters, such as regularization strength. Preprocessing and target construction were learned within the appropriate training splits throughout this process.
Once settings had been selected, the pipeline was refitted on the outer-training sample and used to predict its held-out fold. Repeating the procedure produced one out-of-fold prediction for every development shop. Using the same outer folds for all models made the comparison more informative: each algorithm faced the same evaluation cases. The table reports mean performance across those outer folds.
| Model family | Predictors | Development MAE | Development RMSE | Development R² |
|---|---|---|---|---|
| Mean target baseline | 0 | 0.7525 | 0.9418 | −0.0036 |
| Ridge | 79 | 0.6297 | 0.7975 | 0.2794 |
| Elastic Net | 79 | 0.6212 | 0.7884 | 0.2958 |
| Histogram gradient boosting | 79 | 0.6340 | 0.8051 | 0.2652 |
| Random Forest | 79 | 0.6321 | 0.8025 | 0.2705 |
Why Elastic Net was selected
Elastic Net achieved the lowest average error among the compared families while retaining a transparent prediction rule. Its combination of coefficient shrinkage and sparse selection fits a survey with many related business characteristics: several features can contribute together, while selected coefficients are reduced to zero. Comparable performance across the trained families also shows that useful predictive information was available to several modeling approaches.
The final specification fixed alpha = 0.1 and L1 ratio = 0.1. Across development folds, approximately 52–58 prepared predictors were represented by nonzero coefficients after encoding and regularization. The input specification still contains 79 features; the smaller count describes which inputs contributed through nonzero fitted coefficients.
The pooled development predictions had MAE of 0.6212, RMSE of 0.7889, and R² of 0.2992. These pooled measures are calculated over all out-of-fold predictions and differ slightly from the mean-fold values in the table. With the feature set and model settings fixed, the remaining test was performance on the 259 protected interviews.
Predictive achievements on held-out shops
The study achieved its central modeling objective: shopkeeper survey responses predicted meaningful differences in relative commercial activity among shops excluded from model selection. The final Elastic Net reduced MAE by 10.18% and RMSE by 13.63% compared with the mean-target baseline.
Why this result matters for everyday micro-trade
The protected-test R² of 0.254 represents a 25.4% reduction in squared error compared with predicting the test sample's mean target. The pooled development R² was 0.2992. These results show that the model captured a meaningful, transferable pattern across businesses with different formats, practices, and national contexts.
This is an encouraging achievement in the setting of the study. The survey brings together people making daily business decisions, customers with different habits, and shops operating in varied local circumstances. Finding a pattern that captures roughly a quarter of the variation in the log activity measure on held-out shops is a substantive contribution. Together with lower prediction errors and consistent business associations, it shows that the survey contains knowledge that can be turned into useful comparisons for other shops.
The two reported comparisons serve different purposes. R² summarizes squared-error improvement relative to the test sample's own mean. The operational baseline in the figure uses the development mean, showing the benefit over the simple prediction available before the test outcomes were examined. Against that baseline, RMSE fell from approximately 0.988 to 0.853.
Useful ranking and average calibration
Development predictions had Pearson correlation of 0.547 and Spearman correlation of 0.535 with observed targets. The ranking result matters for peer comparison: higher predicted activity generally corresponded to higher observed activity. A near-zero mean residual, calibration intercept of 0.0018, and calibration slope of 1.0209 also indicated good average alignment on the development data.
The model produced a smoother range of predictions than the observed activity distribution, with standard deviations of 0.505 and 0.943 respectively. In the lowest observed quintile, the median prediction was 0.656 against an observed ratio of 0.316; in the highest quintile, the corresponding values were 1.450 and 3.471. This identifies a specific opportunity for further improvement: preserving the useful overall ranking while making estimates more responsive at the ends of the distribution.
Evidence across countries and shop formats
Country-level results connect the model back to the regional diversity described in the exploratory chapter. Every country had positive development out-of-fold R², indicating predictive information across all five settings.
| Subgroup | n | MAE | RMSE | R² | Median error factor |
|---|---|---|---|---|---|
| Brazil | 304 | 0.612 | 0.767 | 0.327 | 1.679 |
| Colombia | 303 | 0.577 | 0.745 | 0.287 | 1.586 |
| Mexico | 303 | 0.710 | 0.884 | 0.273 | 1.909 |
| Argentina | 61 | 0.428 | 0.556 | 0.228 | 1.490 |
| Peru | 61 | 0.639 | 0.805 | 0.383 | 1.640 |
Across shop formats, pooled R² ranged from 0.236 for pharmacies to 0.346 for grocery stores. Together, the country and format results demonstrate the value of examining the model at several levels: a common regional framework can reveal shared predictive patterns while retaining visibility into local differences. The 61-shop development samples in Argentina and Peru provide a basis for further evidence gathering in those countries.
Understanding prediction errors in business terms
MAE is the average absolute distance between an observed target and its prediction on the log relative-activity scale. RMSE gives greater weight to larger errors. The test values of 0.6915 and 0.8530 therefore describe complementary aspects of prediction performance. Test RMSE was about 8% higher than mean-fold development RMSE, alongside the sustained improvement over the baseline on the final sample.
For a concrete example, suppose a shop's observed activity is twice its structural expectation, giving a log target near 0.693, while the model predicts activity at expectation, or zero. That individual prediction has an absolute log error of 0.693 and a factor-of-two gap in the relative-activity ratio. This example translates one error of roughly the reported MAE into a business comparison. The median error factor is a separate summary calculated from all individual prediction errors.
The overall result is a tested ability to distinguish activity patterns across shops. This gives owners, advisers, and researchers more information than a common average and establishes a useful foundation for interpreting the business characteristics behind the predictions.
Business patterns that consistently predict activity
The model's predictive success is accompanied by interpretable, consistently retained features. For the fixed Elastic Net specification, the analysis recorded whether each coefficient was nonzero, its direction, and its magnitude in each of five development folds. Every feature highlighted below was retained with the same direction in all five fits. Staffing, connectivity, financial access, transaction services, product mix, and facility modernization therefore contributed consistently across the changing training samples.
| Stable predictor | Block | Direction | Mean standardized coefficient | Nonzero with same sign (folds) |
|---|---|---|---|---|
| Offers utility-bill payment | Additional services | Positive | 0.1935 | 5 / 5 |
| People working, including respondent | Business profile | Positive | 0.1755 | 5 / 5 |
| Has business internet | Technology | Positive | 0.1520 | 5 / 5 |
| Has business bank account | Financial services | Positive | 0.1428 | 5 / 5 |
| Has cash register | Technology | Positive | 0.1291 | 5 / 5 |
| Sells medicines | Products sold | Positive | 0.1169 | 5 / 5 |
| Offers cash withdrawals | Additional services | Positive | 0.0980 | 5 / 5 |
| Modernized store facilities | Modernization | Positive | 0.0776 | 5 / 5 |
| Has barcode scanner | Technology | Negative | −0.0757 | 5 / 5 |
For example, utility-bill payment had a positive coefficient in all five folds and a mean standardized coefficient of 0.1935. Business internet and a business bank account showed the same five-fold consistency, with mean coefficients of 0.1520 and 0.1428. These results make the feature evidence concrete: the highlighted patterns persisted as the training sample changed.
The fold counts describe observed coefficient stability. The training samples overlap, so 5 / 5 is a consistency measure, not a p-value or five independent significance tests. Formal coefficient significance was not estimated in this predictive analysis.
These findings connect the statistical model to recognizable parts of daily trading. Staffing describes the people available to serve customers and operate the shop. Internet access and banking describe elements of the business's digital and financial connections. Utility-bill payment and cash withdrawal services describe ways in which a shop may serve additional customer needs. Their repeated appearance across folds makes them useful elements of a structured business profile.
A positive coefficient associates a feature with higher predicted log relative activity, conditional on the other modeled inputs. Reading the coefficients together preserves that context. For example, the negative barcode-scanner coefficient is a conditional pattern within the combined feature set; relationships among technology, format, and other characteristics contribute to its interpretation.
For shopkeepers and advisers, the value is a clearer account of the business dimensions associated with activity. The model makes these connections visible and provides a focused basis for discussing operations, comparing profiles, and designing follow-up research on the practices that matter to micro-trade.
Putting predictions to work for shopkeepers
The practical value is simple: give a shopkeeper a clearer sense of how the business compares with similar shops, and make the patterns behind that comparison understandable. The model turns a long questionnaire into a relative-activity estimate and a business profile that an owner or adviser can use in a conversation.
Understanding a shop's position among its peers
A small grocery shop can be considered in relation to its country, format, and size. For an owner reviewing a quiet trading period, this comparison can help organize a discussion of demand, access, services, or recent local changes. For a shop with higher predicted relative activity, its operating profile offers a useful starting point for learning about the owner's experience. The estimate and the owner's knowledge contribute different parts of the picture.
Making business profiles easier to discuss
The model also brings several survey domains into one view. Instead of reviewing staffing, digital access, financial services, product range, and modernization as disconnected answers, an adviser can consider how they contribute together to a prediction. This provides a consistent framework for conversations across shops and for identifying subjects that merit a visit or more detailed discussion.
Three development-defined bands make the continuous estimate easier to communicate: lower relative activity below approximately 0.537, a central band between 0.537 and 1.923, and higher relative activity above 1.923. The table retains the label “no strong model signal” for the central band used in the evaluation.
| Observed band | Shops | Precision | Recall | F1 |
|---|---|---|---|---|
| Lower relative-activity signal | 64 | 0.500 | 0.219 | 0.304 |
| No strong model signal | 135 | 0.549 | 0.837 | 0.663 |
| Higher relative-activity signal | 60 | 0.480 | 0.200 | 0.282 |
On the protected test, exact band accuracy was 53.67%, balanced accuracy was 41.86%, macro F1 was 0.416, and quadratic-weighted kappa was 0.237. Opposite-extreme assignments occurred for 1.93% of shops. Recall was 0.219 in the lower band, 0.837 in the central band, and 0.200 in the higher band. These results favor using the continuous estimate as the main comparison and the bands as a communication aid; identifying extreme cases remains a development priority.
Supporting learning across micro-trade
For an advisory or research program, the same framework could help compare business profiles, select a range of shops for qualitative follow-up, and make regional patterns easier to discuss. These are practical directions for applying the demonstrated predictive capability. An initial version could combine the estimated relative activity, a plain-language profile of the shop, and a short conversation about what the owner wants to improve or learn.
The practical opportunity is to connect a tested analytical framework with the experience of shopkeepers. Better structured information can guide attention, make comparisons more consistent, and give local business knowledge a clear place in the discussion.
Building on the achievements
The study brings together three achievements that can support further work in Latin American micro-trade. It creates a documented representation of shop operations from a complex survey, establishes a comparable measure of commercial activity, and demonstrates predictive power on held-out businesses. The result is a reusable analytical foundation with a clear connection to everyday commerce.
Trying the model with shopkeepers
The next practical step is to test how owners and advisers use the model's outputs. A small pilot could present the relative-activity estimate and business profile, invite the shopkeeper's interpretation, and record which comparisons or explanations they find useful. This would build evidence about communication, relevance, and the decisions the tool can support in practice.
Extending the information and regional coverage
Further work can build on the identified predictive patterns by adding complementary customer perspectives, gathering repeated observations of trading activity, and increasing coverage in smaller country samples. The current tendency to smooth extreme outcomes provides a focused modeling objective. These extensions have a clear starting point: the existing model and its held-out performance provide a benchmark against which added value can be measured.
Reading the model in context
The model estimates relative commercial activity from reported traffic and spending, with accuracy shaped by those reports and the peer benchmark. Its business associations describe predictive patterns; assessing the effect of changing a practice requires a causal study. The current evidence covers surveyed urban shops in five countries, with smaller groups less precisely assessed. Work with shopkeepers and new samples can test how the tool contributes to decisions and outcomes in practice.
Conclusion
This study demonstrates that shopkeeper survey data can become a useful predictive model for Latin American micro-trade. It achieves measurable improvements in held-out prediction, identifies consistent business patterns, and creates a practical basis for comparing shops within their own operating context.
The final 79-feature Elastic Net achieved protected-test R² of 0.254 and reduced RMSE by 13.63% and MAE by 10.18% relative to the mean-target baseline. Its development ranking correlations and stable associations across folds add to the evidence that the model captures meaningful information about the businesses. The contribution is both analytical and practical: a complex set of survey responses becomes an understandable estimate of relative activity and a structured account of the shop characteristics associated with it.
The work also gives closer research attention to the people and decisions behind regional micro-trade. The exploratory analysis documents small teams, frequent customer contact, long opening schedules, and ambitions for growth. The predictive analysis builds on that picture to support peer benchmarking, business-profile interpretation, and more informed conversations with shopkeepers.
The study turns shopkeepers' shared experience into a useful source of business knowledge. It offers a working predictive model, a common basis for comparison, and clear directions for practical application. Bringing these results back to owners and advisers is the next step toward the project's larger purpose: better support for the shops and people who sustain Latin America's neighborhood commerce.
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.
- 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.
- OECD & European Commission Joint Research Centre. (2008). Handbook on Constructing Composite Indicators: Methodology and User Guide. OECD Publishing. https://doi.org/10.1787/9789264043466-en
- 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.
- Project shopkeeper survey instrument and governed analytical artifacts (2024–2026). Private course-project materials; aggregate results reported here.