Literature Review: Ensemble Methods for Financial Forecasting

1. Time-Series Foundation Models

Recent advances in foundation models have transformed time-series analysis. TimeGPT (Garza & Mergenthaler-Canseco, 2023) demonstrated that pre-trained transformers can achieve zero-shot forecasting competitive with task-specific models.

• Pre-trained models reduce the need for domain-specific feature engineering
• Transfer learning from large-scale temporal datasets improves few-shot performance on financial data

2. Attention Mechanisms in Finance

The application of multi-head attention to financial time-series has shown promising results:

$ python experiments/train.py --model attention_ensemble --epochs 200

• Temporal Fusion Transformers (Lim et al., 2021): Variable selection + interpretable attention weights for multi-horizon forecasting
• Autoformer (Wu et al., 2021): Auto-correlation mechanism replacing self-attention for long-term dependencies
• iTransformer (Liu et al., 2024): Inverted attention across variates instead of time steps — 15% improvement on multivariate benchmarks

3. Ensemble Strategies

Combining heterogeneous model predictions through learned weighting consistently outperforms individual architectures.

4. Research Gaps

While ensemble methods show clear benefits, existing work lacks: (1) adaptive weight learning across market regimes, (2) integration of alternative data sources, and (3) rigorous out-of-distribution evaluation on regime changes.