Sources and patterns of uncertainty in construction MSMEs: A machine learning study in southwestern Colombia

Uncertainty in construction project management (PM) involves the perceived unpredictability of disruptions that influence project duration, costs and resource availability. This issue is particularly pronounced for micro, small and medium-sized enterprises (MSMEs), especially in regions lacking strong institutional support, digital infrastructure and facing environmental and logistical volatility. This study investigates the internal and external sources of perceived uncertainty among MSMEs in southwestern Colombia. Data from surveys of 25 construction firms were analysed to assess how frequently uncertainties occur, their magnitude, and signalling across 10 domains, both internal and external. Using bootstrapped random forest (RF) models, the most impactful features associated with higher perceived uncertainty were identified. These were complemented by classification trees (CTs) to generate interpretable decision rules. To cope with small sample sizes, a class-preserving data augmentation strategy was validated through Mann-Whitney U Tests. Results indicated that internal sources, such as organisational dynamics and resource estimation, are strongly linked to operational maturity and diversification strategies. External uncertainties, such as logistics, weather and sociopolitical factors, vary notably across different regions. Interestingly, 82.9% of firms with over 29 months of experience followed the most common path for higher perceived market uncertainty, suggesting that experience influences perception. Moreover, high uncertainty was predicted in several domains even without typical signals, implying latent variables, possibly undetected by surveys but captured by models, may be affecting perception. This research offers a practical, data-driven framework employing interpretable machine learning to model uncertainty perception in MSMEs, providing tools for early warning and better decision-making in resource-constrained contexts.