VolSense: Explainable Volatility Forecasting & Signal Intelligence

VolSense is an end-to-end, explainable volatility forecasting and signal generation framework — blending econometric foundations (GARCH, EGARCH, GJR-GARCH) with deep neural architectures (LSTM, attention, residual heads) to deliver multi-horizon, interpretable volatility insights across global markets.

---

Overview

VolSense answers a central quant research question:

Can deep learning architectures outperform classical GARCH-style baselines without losing interpretability for traders and risk managers?

To explore this, VolSense introduces a two-tiered architecture:

1. volsense_core — the research and training pipeline for feature engineering, multi-model learning, evaluation, and explainability.
2. volsense_inference — the deployment layer for real-time forecasting, analytics, and signal generation, complete with an interactive Streamlit dashboard.

---

Core Components

1️⃣ VolSense_Core — Research & Model Training

A modular experimentation framework supporting both classical and neural volatility models:

Model	Description
ARCHFamily (GARCH, EGARCH, GJR-GARCH)	Econometric baselines for regime calibration and interpretability.
BaseLSTM	Single-ticker LSTM with residual & attention heads, designed for temporal precision.
GlobalVolForecaster	Multi-ticker shared LSTM with ticker embeddings, horizon-specific heads, and dropout regularization.

Highlights:

• Configurable TrainConfig dataclass for all hyperparameters.
• Robust checkpointing and bundle serialization.
• Multi-horizon support (1d, 5d, 10d).
• Regime-wise validation on calm (2013-14), spike (2020-21), and forward (2023-present) periods.
• Integrated evaluation layer (RMSE, MAE, correlation, Durbin–Watson, feature importance).

Two flagship models — v109 (109 tickers) and v507 (507 tickers) — are trained and released as production-ready baselines.

---

2️⃣ VolSense_Inference — Operational Forecasting & Signals

A portable, production-grade inference layer enabling researchers and traders to run forecasts, analyze analytics, and extract signals on demand.

Modules:

• forecast_engine.py — unified inference entrypoint.
• model_loader.py — loads serialized .pth / .pkl checkpoints.
• predictor.py — manages scaling, batching, and forecast generation.
• analytics.py — computes cross-ticker analytics, realized vs predicted correlation, and feature diagnostics.
• signal_engine.py — transforms forecasts into actionable long/short/neutral signals with sector and regime awareness.
• dashboard.py — a four-tab Streamlit UI for traders and researchers.

Dashboard Tabs:

Tab	Function
Overview	Realized vs predicted volatility snapshot.
Ticker Analytics	Forecast curves, realized overlays, error distributions.
Sector View	Heatmaps and top/bottom sector rankings by Z-score.
Signal Table	Filterable cross-sectional signal matrix by sector, regime, horizon, or position.

---

Data & Feature Engineering

• Automated OHLCV ingestion via Yahoo Finance and fetch.py.
• Unified feature generation with build_features() ensuring identical transformations for training and inference.
• Core features include:
  • vol_3d, vol_20d, vol_60d, vol_ratio, vol_vol, vol_chg
  • return, ret_sq, abs_return, ewma_return_5d
  • beta_20d, market_stress, skew_5d
  • Calendar features: day_of_week, month_sin, month_cos

---

Explainability

VolSense incorporates a multi-layered explainability suite purpose-built for sequence-based volatility forecasting, ensuring transparency across both features and time:

• Attention heatmaps — visualize where the model “looks” across the past 40-day window, revealing temporal focus patterns for each forecast horizon.
• Feature sensitivity analysis — quantifies how changes in key drivers (e.g., abs_return, vol_20d, beta_20d) influence log-volatility forecasts at different horizons (1d, 5d, 10d).
• Temporal sensitivity curves — highlight how recency affects predictive power, showing that recent shocks dominate short-term forecasts while older volatility events fade naturally.
• Cross-horizon explainability — compare how feature importance and attention dynamics evolve from reactive (1-day) to smoothed (10-day) predictions.
• Human-aligned interpretation — findings align with financial intuition: volatility clusters, decays over time, and remains influenced by recent market shocks.

Together, these tools make VolSense not only accurate, but also interpretable — bridging quantitative modeling with human reasoning about market dynamics.

Evaluation

Comprehensive evaluation framework (evaluation.py, metrics.py) featuring:

• RMSE, MAE, correlation, and horizon-specific error tracking.
• Volatility clustering and Durbin–Watson diagnostics.
• Residual distribution and QQ plots for calibration.
• Best/worst ticker identification across horizons.
• Feature importance and prediction–truth correlation analytics.

---

Deployment & CLI

VolSense provides a lightweight CLI for both training and forecasting:

# Train models
volsense-train --config configs/train_v109.yaml

# Run forecasts
volsense-forecast --model v109 --tickers AAPL,MSFT,SPY

Streamlit Dashboard Preview

Launches an intuitive dashboard providing:

• Real-time forecasts and realized vol tracking.
• Sector heatmaps and Z-score distributions.
• Cross-sectional signal visualization with position, regime, and horizon filters.
• One-click CSV export for further analysis.

Documentation

Full technical documentation is generated via Sphinx and Napoleon, including:

• API references for every core/inference module.
• Notebook-integrated tutorials (01_data_pipeline → 06_explainability).
• Architectural diagrams and configuration examples.

Repository Structure

VolSense/
│
├── volsense_core/
│   ├── cli/                  # Training CLI
│   ├── data/                 # Data fetch + feature engineering
│   ├── evaluation/           # Metrics + evaluation
│   ├── explainability/       # SHAP & interpretability notebooks
│   ├── models/               # LSTM, GARCH, GlobalVolForecaster
│   ├── utils/                # Checkpoint utils, scalers
│   └── forecaster_core.py    # Core trainer entrypoint
│
├── volsense_inference/
│   ├── cli/                  # Forecast CLI
│   ├── notebooks/            # Dashboard & explainability demos
│   ├── analytics.py
│   ├── forecast_engine.py
│   ├── model_loader.py
│   ├── predictor.py
│   ├── signal_engine.py
│   ├── dashboard.py
│   └── sector_mapping.py
│
├── models/                   # Trained model bundles (v109, v509)
│
├── notebooks/                # Jupyter notebooks (01–06)
│   ├── 01_data_pipeline.ipynb
│   ├── 02_model_training.ipynb
│   ├── 03_model_inference.ipynb
│   ├── 04_signals_and_dashboard.ipynb
│   ├── 05_v109_training.ipynb
│   └── 06_v109_explainability.ipynb
│
├── docs/                     # Sphinx documentation
├── pyproject.toml
├── README.md
├── requirements.txt
└── .gitignore

Credits & License

Developed by Rahul Karthik as part of the VolSense Research Framework. Inspired by hybrid econometric-deep learning models in quantitative volatility research.

Licensed under the MIT License. Pull requests, collaborations, and discussions are welcome!

Connect

• 📫 LinkedIn: Rahul Karthik
• 💻 GitHub: github.com/rahulmkarthik
• 🧠 Research Interests: Quantitative Finance · Quantitative Systems Design · Volatility Modeling · Explainable AI
• 🧰 Tech Stack: Python · Matplotlib · NumPy · PyTorch · SHAP · Pandas · Seaborn · Streamlit · Sphinx · yFinance