SeNA: A Hierarchical Self-Evolving Neural Architecture for Stock Forecasting

This paper introduces SeNA (Self-Evolving Neural Architecture), a hierarchical 3-stage knowledge-distillation framework combining a Transformer teacher, an LSTM parent, and an MLP student with meta-evolution cycles. Unlike traditional hybrid models, SeNA adapts its internal MLP kernel weights through loss-conditioned meta-updates, enabling fast learning and robust generalization.

We evaluate SeNA on 15 years of historical stock market data across AAPL, TSLA, META, MSFT, and NVDA using a strict 100-day fully leak-free holdout. No future information, no overlapping windows, no normalization leakage is allowed in the pipeline.

SeNA achieves strong performance across all tickers, reaching:

- AAPL: R² = 0.964, Directional Accuracy = 67%
- TSLA: R² = 0.936, Directional Accuracy = 63%
- META: R² = 0.827, Directional Accuracy = 61%
- MSFT: R² = 0.796, Directional Accuracy = 56%
- NVDA: R² = 0.831, Directional Accuracy = 54%

These results exceed typical publicly reported benchmarks for OHLC-only, long-horizon blind forecasting models. The combination of cross-architecture distillation and meta-learning provides stability, low latency (under 12 minutes training), and real-world usability.

This work establishes SeNA as a promising foundation for practical forecasting, lightweight financial AI, and future research into self-evolving architectures.