Relevance
9/10
Highly relevant to quantitative trading. The paper directly addresses uncertainty quantification for financial decision-making, demonstrating practical applications in uncertainty-aware trade filtering and portfolio optimization with uncertainty-based position sizing. The epistemic-aleatoric decomposition enables distinguishing between model uncertainty (reducible) and market noise (irreducible), which is critical for position sizing and risk management. Results show DER achieves superior Sharpe (1.33) and Sortino (2.27) ratios on BTC. However, poor PICP calibration and limited uncertainty-error correlation on some assets temper practical deployment readiness. The work is from JPMorganChase, indicating industry relevance.
Implementation Complexity
6/10
Moderate complexity. The DER head with NIG parameter constraints (Softplus activations, domain restrictions) requires careful implementation. The combined loss function (evidential + coverage + weight decay) and evidence annealing schedule add training complexity. However, the architecture itself is straightforward (standard transformer or LSTM backbone). The controlled evaluation using LSTM is relatively simple to reproduce. Key challenges include: ensuring NIG parameter validity, tuning lambda_coverage, implementing the evidence annealing schedule, and handling the coverage loss computation. Single-pass inference is computationally efficient compared to sampling-based methods.
Reproducibility
4/5
The paper provides detailed hyperparameters (Adam optimizer, LR 0.001, batch size 128, 50 epochs, early stopping patience 10, gradient clipping max norm 1.0, weight decay 0.001), architecture specifications (1-layer LSTM, 32 hidden dims, 0.1 dropout, 50-step lookback), data processing pipeline (outlier truncation at 1st/99th percentiles, per-symbol standardization), temporal split (up to 2023 train/val, 2024+ test), random seed (0), and framework (PyTorch). However, no code repository is provided, and some ProbFM-specific hyperparameters (lambda_coverage, T_anneal) are not fully specified numerically.