Educational success rate prediction worldwide still faces fundamental challenge, such as data availability that varies across administrative regions. In this investigation, we address this challenge through a predictive deep learning approach applied to Burkina Faso’s primary education system, where some regions maintain consistent records while others face disruptions from conflict, such as terrorism, and security challenges. During the investigation, we used a dataset that contains 436 communes and collected the data over 12 years (2009–2021). The dataset combines educational metrics with contextual indicators. Rather than requiring uniform temporal depth, we developed an adaptive LSTM architecture that adjusts its lookback window according to each administrative unit’s available data history. We combined this with LightGBM in an ensemble learning manner to strengthen feature representation and capture local contextual factors that influence educational outcomes. This adaptive capability makes the proposed model efficient for educational planning in diverse contexts, from developing countries that are building their data infrastructure to developed countries with historically inconsistent data collection. Our experimental results show that the proposed model yielded an overall prediction \(R^2\) of 0.63 and MAE of 7.38, representing state-of-the-art performance for this prediction task.