Machine learning analysis of subsistence determinants in Peruvian agricultural systems: A multi-regional approach using random forest and gradient boosting

Abstract

Subsistence agriculture persists across Peru with markedly heterogeneous patterns among ecological regions, yet multi-regional analyses using predictive methods remain scarce. This study identified determinants of household subsistence using random forest and gradient boosting classifiers. Microdata from the 2024 Peruvian National Agricultural Survey yielded 34,827 agricultural units after quality control. Subsistence was operationalised as the share of self-consumption in total production value and categorised as low (0% - 20%), medium (20% - 60%) and high (> 60%). On a held-out test set of 10,448 units both algorithms converged (random forest 82.9%, gradient boosting 82.6% accuracy), yet a nested sensitivity analysis showed that this performance depends substantially on predictors that enter the definition of the dependent variable. Once income-derived and market-participation variables were removed, accuracy settled at 65.7% against a 57.9% majority-class baseline, and harvested area and input expenditure emerged as the leading structural determinants. Three findings run against prevailing narratives. First, productive richness does not accompany commercialisation: Andean units combined the highest richness (7.9 versus 5.8 species on the Coast) with the highest subsistence (39.1% versus 15.0%). Second, direct-to-consumer selling was associated with higher subsistence (34.2%) than collector intermediation (14.3%), consistent with intermediaries operating as a pragmatic commercialisation route rather than a barrier. Third, accuracy was lowest precisely in the Andes (77.0%), where subsistence concentrates, and weakest for the transitional medium class (60.7%). Reporting predictive performance without separating definitional from structural predictors overstates what such models explain, a distinction rarely made in this literature.