Quantitative Financial Modeling for Sri Lankan Markets: Approach Combining NLP, Clustering and Time-Series Forecasting

Battle Count: 124

Highly relevant to quantitative trading as it directly addresses market signal generation (Buy/Sell/Hold), directional accuracy, price forecasting, economic regime detection, and risk assessment for emerging markets. The rule-based fusion logic is designed as a Decision Support System for trading. The integration of ESG sentiment with time-series forecasting and global correlations (S&P 500) provides a multi-factor approach. However, the directional accuracy (52.78%) is modest, the fusion logic is rule-based rather than learned, and the evaluation is preliminary. The framework is more suited for risk-averse, disciplined trading in volatile emerging markets rather than high-frequency or aggressive strategies.

High complexity due to the multi-stage pipeline requiring: (1) FinBERT fine-tuning/inference for ESG sentiment extraction, (2) UMAP dimensionality reduction and HDBSCAN clustering, (3) PCA validation, (4) Autoencoder training, (5) Two-stage ensemble classifier (GBM + MLP/DNN), (6) Multiple time-series model training (SRNN, MLP, LSTM, GRU), (7) Correlation analysis with S&P 500, (8) Rule-based fusion logic implementation. Requires expertise in NLP, unsupervised learning, deep learning, time-series analysis, and financial data engineering. Multiple data sources need integration and alignment. The pipeline has many hyperparameters to tune across components.

The methodology is described in detail with equations and model architectures, but no code repository is provided. Data sources (Bloomberg, Yahoo Finance, CBSL, Trading Economics) are publicly accessible but require subscriptions for some. Specific hyperparameters are partially disclosed (e.g., GBM: 100 trees, depth 4, lr 0.1; MLP: 128-64-32-5 neurons, 0.3 dropout, Adam optimizer). The 70/15/15 train-test-validation split is mentioned. However, exact preprocessing steps, random seeds, and full dataset specifications are not provided. The rule-based fusion logic is described but its full evaluation is noted as incomplete.