Robust Transformer-Based One-Step Stock Index Forecasting via Shifted Data Augmentation

Battle Count: 157

The paper directly addresses one-step stock index forecasting, which is fundamental to short-term trading decisions, portfolio rebalancing, and risk management. The emphasis on stability, reproducibility, and parameter efficiency is highly relevant for production trading systems. However, the paper does not translate forecasts into trading signals, does not account for transaction costs, does not evaluate economic profitability (e.g., Sharpe ratio, returns), and focuses on index-level prediction rather than individual securities. The SDA technique's robustness to distributional shifts is particularly valuable for live trading environments where market regimes change.

The modified Transformer architecture is a relatively straightforward adaptation of the standard Transformer with GeLU activation, modified dropout placement, and linear output projection. The SDA technique is extremely simple (adding constant offsets to replicate training data). Learning-rate scheduling (cosine annealing with warmup) is standard in PyTorch. The main complexity lies in hyperparameter tuning across multiple dimensions (L, N, d_model, d_ff, h, p_drop, k) and the need to select appropriate SDA offset constants per dataset. The model is computationally efficient, with compact configurations (31K parameters) achieving competitive results.

The paper provides detailed hyperparameter configurations, learning-rate scheduling formulas, SDA offset constants, and evaluation metrics averaged over 10 independent runs. Datasets are publicly available from Investing.com. Hardware and software environments are specified. However, no GitHub repository or code is provided, and the exact random seeds for the 10 runs are not disclosed. The SDA offset constants (750 for VN30, 4000 for S&P 500) are dataset-specific and require manual selection.