Multi-Modal Features of Trading Candle Chart Imagery & Volume For Predicting Financial Market Movements, Using the Proposed BLENNs Architecture.
Loading...
Authors
Adeyemo, Emmanuel
Issue Date
2026-04
Type
Dissertation
Language
en
Keywords
Business, Engineering, Science, & Technological Innovation , multimodal learning , financial forecasting , explainable AI , BLENNs , noise robustness
Alternative Title
Abstract
Financial forecasting models face significant challenges due to reliance on single-source data and lack of transparency required by regulators, impacting institutional investors, retail traders, and regulators who need reliable and interpretable AI systems for market decisions. This study developed and evaluated the Blended Neural Networks model, known as BLENNS, a hybrid deep learning framework integrating convolutional neural networks for pattern recognition, long short-term memory networks for sequential data, attention mechanisms for feature weighting, interpretability techniques, and probabilistic uncertainty estimation. Guided by multimodal learning, signal detection, and explainable AI theories, the study investigated whether multimodal fusion improves forecasting accuracy, if the Blended Filtered Candles preprocessing method enhances noise robustness, and whether interpretability aligns with expert trading rules. Using daily financial data from 2010 to 2025 on six diverse assets with over 21,000 observations, the Blended Filtered Candles method applied a three-stage filtering process including exponential smoothing, an enhanced candle transformation, and adaptive Kalman filtering. Walk-forward validation with multiple expanding windows ensured rigorous out-of-sample testing. BLENNs achieved 97.55% directional accuracy, a 113.77% improvement over traditional models, while BFC preprocessing improved the signal-to-noise ratio by 134.8%, outperforming common smoothing techniques by large margins. Interpretability analysis showed statistically significant, though modest, agreement with expert trading principles, emphasizing the value of explainable AI combined with human oversight. Simulated and live trading demonstrated strong returns and win rates, with live performance reflecting realistic trading costs and execution factors. This framework offers a foundation for meeting regulatory transparency requirements, though further compliance testing, extended live validation, scalability assessment, and transaction cost considerations are needed for practical deployment. The consistent noise reduction by BFC across multiple assets supports its broader application. Resources are available for further research exploring additional markets, higher-frequency data, institutional trading, adaptive parameter tuning, and establishing interpretability standards for regulation.