Relevance
4/10
The paper is primarily focused on bank profitability and causal inference rather than direct trading strategies. However, the sentiment analysis methodology (FinancialBERT on quarterly reports) and causal framework could inform event-driven trading strategies. The findings on how leverage and asset composition moderate sentiment effects are relevant for risk-adjusted portfolio construction. The Nepal-specific context and small sample (10 firms) limit direct applicability to quantitative trading in major markets.
Implementation Complexity
8/10
The pipeline requires multiple integrated components: (1) pre-trained FinancialBERT for sentiment classification, (2) XGBoost regression for feature importance, (3) SHAP analysis for interpretability, (4) hierarchical clustering and correlation analysis for feature selection, and (5) Causal Forest with AIPW for heterogeneous treatment effect estimation. Each component requires careful tuning, and the integration between stages (e.g., using SHAP-selected features as inputs to CF) adds complexity. The causal inference framework requires understanding of potential outcomes, propensity scores, and doubly robust estimation.
Reproducibility
2/5
The paper describes methodology in detail with equations and references, but does not provide a public code repository or exact hyperparameters for the Causal Forest or XGBoost models. Data sources (NEPSE, NRB) are identified but access procedures are not fully specified. The pre-trained FinancialBERT model is referenced but the specific version and fine-tuning details are limited. Only 10 firms are analyzed, limiting generalizability. No GitHub repository with code is explicitly provided.