Relevance
7/10
The 65.15% directional accuracy significantly exceeds random (50%) and benchmarks (52-55%), suggesting potential utility for generating trading signals. However, relevance is tempered by: (1) daily frequency only, (2) no transaction cost modeling, (3) single emerging market focus with low liquidity, (4) modest R² indicating limited variance capture, (5) failure during extreme events, and (6) no portfolio-level or risk-adjusted return analysis. The walk-forward validation methodology is highly relevant for realistic backtesting. The feature importance analysis provides interpretable insights useful for strategy design.
Implementation Complexity
4/10
Core XGBoost implementation is straightforward with well-documented libraries. However, the full pipeline adds complexity: Optuna hyperparameter tuning with time-series CV, walk-forward validation with expanding/rolling windows, multiple lag configurations, feature engineering (lags, RSI, rolling volatility), price reconstruction via exponentiation, and comparison against 6+ benchmark models (ARIMA, Ridge, CNN, LSTM, N-BEATS, TFT). Deep learning benchmarks require GPU infrastructure. The statistical testing (DM, PT, bootstrap) adds further implementation overhead. Overall moderate complexity for an experienced ML practitioner.
Reproducibility
5/5
Complete source code, processed datasets, trained models, and results are publicly available on GitHub (https://github.com/sahajrajmalla/nepse-xgboost-forecasting). Fixed random seeds are used. All hyperparameter search spaces, validation protocols, and feature engineering steps are fully documented. Data source (NepseAlpha) is specified. The paper provides detailed tables of all configurations and results.