SignalFusion is a regime-aware decision intelligence framework that combines unstructured text and structured quantitative signals to infer latent system states and prioritize events under uncertainty.
The system is designed as a general platform, demonstrated across two domains:
- Customer Support Operations
- Financial News & Market Events
- Systems operate in latent regimes that are not directly observable
- Text and quantitative signals jointly reveal those regimes
- Event importance is conditional on regime, not absolute
- Data ingestion (structured + unstructured)
- Text embedding (Sentence Transformers)
- Time aggregation
- Feature alignment
- PCA for stability
- Hidden Markov Model (HMM)
- Post-hoc regime labeling
- Importance scoring
- Explainability & visualization
signalfusion/ ├── config/ │ ├── run.yaml │ ├── importance.yaml │ ├── financial_importance.yaml │ ├── src/ │ ├── ingestion/ │ ├── features/ │ ├── nlp/ │ ├── models/ │ ├── decisioning/ │ ├── visuals/ │ └── utils/ │ ├── customer_ticket_runner.py ├── runner_financial.py ├── outputs/ └── README.md
Latent regimes are inferred using an HMM trained on:
- aggregated structured features
- aggregated text embeddings
The model outputs probabilistic state membership over time.
State semantics are assigned after inference using domain-specific stress indicators.
Each event (ticket or headline) receives an importance score based on weighted components such as:
- recency
- magnitude / urgency
- sentiment
- entity involvement
Scores are bucketed (P0–P3) for operational use.
Within each regime:
- embeddings are clustered (KMeans)
- clusters are labeled using TF-IDF
This reveals sub-themes without modifying the core regime model.
- features_matrix.csv
- state_probabilities.csv
- bi_table.csv
- state_summary.csv
- state_top_terms.csv
- state_topic_clusters_summary.csv
- top_important_events.csv
- regime probability plots
- stress signal overlays
This system is designed for decision support and analysis. It does not constitute investment advice or predictive guarantees.
Built as a platform-oriented data science project emphasizing robustness, interpretability, and cross-domain applicability.