Findings

The COFINFAD project combines confirmatory statistics, behavioural segmentation, and predictive modelling to explain how Colombian fintech users differ in spending, engagement, and customer outcomes.

48,723 customers analysed 3 analytical lenses 1 integrated fintech story

Confirmatory

Most demographic groups show heavily overlapping financial distributions, suggesting that demographics alone do not strongly explain customer activity.

Clustering

Customer segments become much clearer when behaviour and value variables are analysed together rather than demographic labels alone.

Predictive

Machine learning is much stronger for satisfaction score than for transaction behaviour, with Random Forest delivering the best satisfaction model.

Executive Summary

Three major conclusions emerged from the project.

  • First, demographic variables do not create large financial divides on their own. Across education level, income bracket, marital status, and gender, transaction distributions remain strongly right-skewed with substantial overlap across groups.
  • Second, segmentation adds more value than simple demographic slicing. Clustering reveals distinct customer personas with meaningful differences in age, average transaction value, and total transaction volume.
  • Third, behavioural and engagement variables are more useful than demographics for prediction. Predictive modelling shows that satisfaction and transaction behaviour are driven more by usage, value, and account characteristics than by socio-demographics alone.

Confirmatory Analysis

ANOVA And Distributional Evidence

The confirmatory analysis examined whether financial behaviour differed materially across socio-demographic groups. The visual evidence from the Shiny dashboard consistently showed the same structural pattern:

  • transaction measures are highly right-skewed
  • every demographic group contains extreme high-value outliers
  • central tendencies across groups remain close and heavily overlapping

This means demographic categories are helpful for description, but on their own they do not create sharply separated financial profiles.

Education Level

For average transaction value by education level, the non-parametric ANOVA result showed no statistically meaningful difference across groups:

  • Kruskal-Wallis: chi-squared(3) = 3.58
  • p = 0.31
  • sample size: n = 48,723

Median transaction values were nearly identical across categories:

  • Bachelor: about 1.76M COP
  • High School: about 1.77M COP
  • Master: about 1.76M COP
  • PhD: about 1.75M COP

The transaction-count boxplots by education level also showed a common pattern of wide dispersion and heavy overlap, reinforcing that education is not a strong standalone driver of activity.

Income Bracket

For average transaction value by income bracket, the robust ANOVA result also indicated no meaningful difference:

  • Robust test: F(3, 6395.46) = 0.95
  • p = 0.42
  • sample size: n = 48,723

Trimmed means were again very similar:

  • High: about 2.45M COP
  • Low: about 2.51M COP
  • Medium: about 2.49M COP
  • Very High: about 2.52M COP

Although income might be expected to separate customers more strongly, the observed financial distributions remained remarkably similar once the full spread of the data was considered.

Gender And Marital Status

The transaction-count plots by gender and marital status showed the same broad finding:

  • all groups share similar medians
  • all groups contain very large outliers
  • no group displays a clearly dominant distribution

This suggests that customer financial behaviour in the dataset is not easily explained by a single demographic dimension. Behaviour is present in every segment, but it is uneven and highly dispersed.

Association Analysis

Association testing was used to assess whether binned financial behaviour was linked to demographic groups.

For average transaction value versus gender, the result showed effectively no association:

  • Pearson chi-squared: 0.84
  • p = 0.93
  • Cramer’s V: approximately 0.00
  • sample size: n = 48,723

The stacked proportion chart confirmed this visually. Female, male, and other customers were distributed across low, medium, and high spend bands in almost the same proportions.

This matters because it strengthens the earlier ANOVA conclusion: even when transaction value is simplified into broad spending bands, demographic grouping still does not meaningfully separate customers.

Clustering Analysis

Segment Discovery

The clustering module adds the strongest behavioural insight in the project because it groups customers using their observed financial profile rather than demographic labels.

From the dashboard view using age, average transaction value, and total transaction volume, the three-cluster solution can be interpreted as follows:

Cluster Approx. size Average age Average transaction value Total transaction volume Interpretation
3 29,122 36.97 2.42M COP 153.71M COP Younger mainstream users with moderate value and the largest population share
2 4,772 44.32 14.02M COP 1.22B COP Mid-aged high-value customers with exceptionally strong financial contribution
1 14,829 59.50 2.45M COP 150.03M COP Older steady users with moderate average value and lower aggregate contribution than the high-value segment

The main insight is not just that clusters exist, but that behavioural value concentration is highly uneven. A relatively small segment contributes disproportionately high transaction value and total volume.

What The Clusters Tell Us

  • Cluster 2 is the most commercially important segment. Even though it is the smallest group, it has by far the highest average transaction value and total transaction volume.
  • Cluster 3 is the operationally dominant segment. It is the largest group and represents the broad mainstream customer base with moderate value behaviour.
  • Cluster 1 appears mature but not high-spending. This older segment behaves more steadily than explosively, suggesting retention and product-fit strategies may matter more than volume expansion.

These clusters show that customer value is better explained by behavioural grouping than by demographics alone. In other words, age contributes to the story, but value intensity is what truly separates the segments.

Predictive Modelling

Satisfaction Score Prediction

For satisfaction score, three regression models were compared: Decision Tree, Linear Regression, and Random Forest.

Model RMSE MAE R2
Decision Tree 0.58 0.43 0.01
Linear Regression 0.58 0.43 0.03
Random Forest 0.43 0.32 0.87

The outcome is very clear: Random Forest is the best validation model by a wide margin.

This tells us two things:

  • satisfaction is influenced by non-linear relationships
  • richer combinations of customer characteristics matter more than a simple linear explanation

The feature-importance chart further supports this. The strongest drivers included:

  • clv_segment_Platinum
  • max_amount_eng
  • avg_amount_eng
  • feature_usage_diversity
  • trans_count_eng
  • customer_tenure

Taken together, these suggest that satisfaction is linked to a mix of customer value tier, transaction intensity, and platform engagement.

Transaction Behaviour Prediction

For transaction behaviour, the target was modelled as log1p(avg_daily_transactions) because the original variable was heavily right-skewed.

Model RMSE MAE R2
Decision Tree 0.193 0.08 0.20
Linear Regression 0.20 0.09 0.04
Random Forest 0.194 0.08 0.20

Here the conclusion is more nuanced:

  • Decision Tree performed best by a very small margin
  • Random Forest was nearly identical
  • Linear Regression underperformed both non-linear models

This means transaction behaviour is harder to predict than satisfaction, and much of its variation is still unexplained by the available features.

The strongest predictors shown in the dashboard were:

  • credit_utilization_ratio
  • customer_tenure
  • base_satisfaction
  • nps_score
  • app_logins_frequency
  • feature_usage_diversity

So while demographics provide context, transaction behaviour is better explained by engagement, account usage, and financial behaviour signals.

Integrated Findings

What The Project Shows Overall

When all three analytical components are read together, a coherent story emerges.

1. Demographics Are Descriptive, Not Deterministic

Education, income, gender, and marital status help describe the customer base, but they do not by themselves create large and consistent financial differences.

2. Behaviour Creates The Real Separation

Clustering reveals the important commercial distinction in the data: a large mainstream base, a smaller older steady group, and a compact but disproportionately valuable high-spend segment.

3. Engagement And Value Signals Matter Most For Prediction

The predictive models repeatedly surfaced variables tied to account usage, transaction magnitude, value tier, and customer tenure. This indicates that fintech behaviour is better understood through how customers use the platform than through static demographic labels.

4. Different Outcomes Have Different Predictability

Customer satisfaction was modelled much more successfully than transaction behaviour. This suggests the available variables align more strongly with customer experience than with the full complexity of day-to-day transaction activity.

Business Implications

Final Conclusion

The project shows that customer behaviour in the Colombian fintech ecosystem is heterogeneous, skewed, and behaviour-driven. Traditional demographic comparisons reveal only limited differences, while clustering and predictive modelling uncover the stronger underlying structure of the customer base.

The most important practical lesson is that customer value emerges more clearly from behaviour, engagement, and transaction patterns than from demographics alone. This makes behavioural analytics the most useful lens for future fintech decision-making.

This findings page synthesises the strongest evidence visible in the project website, Shiny dashboard outputs, and predictive-modelling workflows currently available in the repository and screenshots.