A Confidence-Calibrated Adaptive Ensemble Voting Framework for Stock Price Direction Prediction

Predicting stock price direction is crucial for investment decisions, yet most machine learning models optimize classification metrics while ignoring probability calibration. Uncalibrated probabilities often produce overconfident predictions, increasing financial risk. To address this gap, we introduce Confidence-Calibrated Adaptive Ensemble Voting (CAEV-C), a framework combining temperature scaling calibration with dynamic confidence-weighted aggregation. We evaluate CAEV-C using XGBoost, LightGBM, and Bidirectional Long Short-Term Memory (BiLSTM) base learners across seven technology stocks (AAPL, AMZN, GOOGL, META, MSFT, NVDA, and TSLA) under a temporal walk-forward validation scheme. CAEV-C achieves the highest average F1-score of 0.5885, outperforming LightGBM (0.4399), Hard Voting (0.5391), and Soft Voting (0.4451). Furthermore, CAEV-C yields the lowest Expected Calibration Error (ECE) of 0.0347, improving calibration quality over LightGBM (0.2053), Hard Voting (0.2629), and Soft Voting (0.1062). Reliability diagrams confirm that CAEV-C confidence estimates track ideal calibration closely, while Wilcoxon Signed-Rank Tests demonstrate statistically significant ECE improvements over Hard and Soft Voting, as well as significant F1-score gains over Soft Voting (p < 0.05). By improving probability reliability while maintaining competitive predictive performance, CAEV-C provides a robust ensemble framework for quantitative stock forecasting.

@article{Lukito2026,
  title     = {A Confidence-Calibrated Adaptive Ensemble Voting Framework for Stock Price Direction Prediction},
  journal   = {International Journal of Advanced Computer Science and Applications},
  volume    = {17},
  number    = {9},
  year      = {2026},
  publisher = {The Science and Information Organization},
  author    = {Yuan Lukito and Nugroho Agus Haryono and Raden Gunawan Santosa},
  doi       = {10.14569/IJACSA.2026.0170936},
  url       = {https://doi.org/10.14569/IJACSA.2026.0170936}
}