Out-of-Fold Stacking for Reliable Transformer Ensembles in Indonesian Depression/Anxiety Indicator Detection

Transformer-based models are widely used for depression/anxiety indicator detection from social media text, often further enhanced through ensemble learning. While stacking ensembles are commonly adopted, the validation strategy used to train the meta-learner is frequently overlooked and may lead to unreliable performance gains. In particular, conventional stacking approaches that train the meta-learner on predictions from base models trained on the same data risk meta-level overfitting. This paper presents a controlled empirical comparison between conventional stacking and stacking with strict Out-of-Fold (OOF) meta-training for Indonesian depression/anxiety indicator detection from Twitter data. Results show that non-OOF stacking fails to consistently outperform the strongest individual model, whereas OOF stacking consistently produces the best-performing ensemble across multiple runs. Although the absolute gains are modest, the improvement is systematic and reproducible. These findings highlight the importance of validation discipline in stacking ensembles and demonstrate that OOF meta-training yields more reliable ensemble performance for low-resource mental health text classification.