HybridSwingNet for explainable swing trading using multi-encoder deep learning and confidence-calibrated signal execution

The challenge with swing trading is that short to medium term price movements are noisy, fast moving and dominated by transient technical, contextual and sentiment driven signals. Existing forecasting approaches typically treat these information sources independently or fuse them via static fusion, which limits the adaptability, selectivity and interpretability in non-stationary markets. To address this issue, this paper proposes a hybrid multi-encoder deep learning framework that combines multi-source market signals with dynamic fusion and confidence-aware signal execution for explainable swing-trading decisions. The key novelties are horizon-specific feature routing, adaptive multi-encoder fusion and calibrated confidence filtering which enable the model to focus on different representations across forecast horizons in a transparent and execution-aware manner. On a 100-stock Indian equity universe (84 Nifty 100 constituents + 16 large-cap Nifty 500 additions) over 2013–2026, the framework achieves 68.7% H1 smoothed-trend directional accuracy on held-out test stocks (70.2% on a fully blind holdout set; \(p = 3.18\times 10^{-63}\) ), rising to 82.9% (test) / 85.3% (blind) at a 0.75 confidence threshold. The Conservative profile yields +11.51% (Sharpe 1.2) across 2025 over 85 trades, with a profit factor of 1.43 (win rate 54.1%) when combined with a dynamic regime filter and full Indian retail delivery transaction costs ( 41 basis points round-trip). A second out of sample period is negative in 2026 and truthfully reported in the limitations. This implies that the main contribution of the framework is to transform noisy forecasts into selective and risk-aware trading actions.