<![CDATA[This study addresses the challenge of transforming sequential clickstream data into accurate yet interpretable behavioral predictions for operational decision-making in digital systems. While complex machine learning models often achieve high accuracy, their limited transparency hinders practical adoption. Therefore, this research aims to develop and evaluate a probabilistic Markov-based framework for predicting users’ next actions while maintaining interpretability. A quantitative data mining approach is applied to e-commerce clickstream data collected in January 2026. User interactions are sessionized and mapped into eight discrete behavioral states. The study compares a frequency-based baseline with first-order, second-order, and variable-order Markov models using back-off and Laplace/Dirichlet smoothing. Model evaluation employs a time-based train–test split with Accuracy@1, Mean Reciprocal Rank (MRR), and log-loss as performance metrics. Results indicate that the variable-order Markov model achieves the best performance, improving Accuracy@1 from 0.231 to 0.331, increasing MRR from 0.318 to 0.437, and reducing log-loss from 1.74 to 1.39. The findings demonstrate that Markov-based models offer an effective balance between predictive accuracy and interpretability, enabling the identification of dominant transitions, drop-off points, and conversion bottlenecks. Future research may extend this framework with time-aware or hidden-state models to capture latent user intent, while managerial implications include data-driven system optimization, recommendation enhancement, and user retention strategies.]]>
