diff --git a/CLAUDE.md b/CLAUDE.md
new file mode 100644
index 0000000..bd617d6
--- /dev/null
+++ b/CLAUDE.md
@@ -0,0 +1,112 @@
+# CLAUDE.md
+
+## Project Overview
+
+**volatility-forecast** is a Python library for time-series volatility forecasting combining statistical and deep learning models. It provides a scikit-learn compatible API and integrates with **alphaforge** for scalable dataset construction.
+
+## Repository Structure
+
+```
+volatility_forecast/           # Main package
+  model/                       # Forecasting models (ES, STES, XGBoost-STES, VolGRU)
+  features/                    # Feature engineering (returns, signatures, range estimators)
+  targets/                     # Target variable definitions (squared returns)
+  sources/                     # Data source adapters (Tiingo EOD, simulated GARCH)
+  data/                        # Data handling (PriceVolume dataclass)
+  evaluation/                  # Metrics (RMSE, MAE, QLIKE), Diebold-Mariano test
+  benchmark/                   # Benchmarking utilities
+  reporting/                   # Reporting and visualization
+  api/                         # Flask REST API
+  forecast/                    # Forecast generation
+  pipeline.py                  # Dataset building pipeline (VolDatasetSpec)
+  storage.py                   # DuckDB/filesystem storage
+  main.py                      # Entry point
+tests/                         # Pytest test suite (24+ files)
+examples/                      # Runnable example scripts
+scripts/                       # Utility scripts
+config/                        # YAML configs (model, data, app)
+docs/                          # MkDocs documentation
+notebooks/                     # Jupyter notebooks
+```
+
+## Development Environment
+
+- **Python**: 3.11 or 3.12 (conda env `py312` on macOS)
+- **Install**: `pip install -e .[dev]`
+- **Environment variables**: stored in `.env` at repo root
+  - `TIINGO_API` — Tiingo API key (required for live data tests)
+  - `TIINGO_CACHE_DIR` — cache directory
+  - `TIINGO_CACHE_MODE` — cache mode
+
+## Commands
+
+```bash
+# Run all tests
+pytest tests/ -v
+
+# Run tests (conda)
+conda run -n py312 pytest -q
+
+# Run a specific test file
+pytest tests/test_signature_features.py -v
+
+# Run examples (some require TIINGO_API)
+python examples/demo_forecast.py
+```
+
+## CI/CD
+
+GitHub Actions (`.github/workflows/ci.yml`):
+- Triggers on push/PR to `main`
+- Tests on Python 3.11 and 3.12 (Ubuntu)
+- Installs with `pip install -e .[dev]` then runs `pytest tests/ -v`
+
+## Architecture & Key Patterns
+
+### Model hierarchy
+All models inherit from `BaseVolatilityModel` (in `model/base_model.py`) which extends scikit-learn's `BaseEstimator`/`RegressorMixin`. Models implement `fit()` and `predict()`. Fitted attributes use trailing underscore convention (e.g., `params_`, `is_fitted_`).
+
+### Key models
+| Model | File | Description |
+|-------|------|-------------|
+| ES | `model/es_model.py` | Exponential Smoothing (constant alpha) |
+| STES | `model/stes_model.py` | Smooth Transition ES with logistic smoothing |
+| XGBoost-STES | `model/tree_stes_model.py` | XGBoost hybrid with STES |
+| VolGRU | `model/volgru_model.py` | GRU neural network (PyTorch/JAX backends) |
+
+### Feature system
+Features use alphaforge `FeatureTemplate` patterns. Key feature types:
+- **Return features** (`features/return_features.py`): lagged returns, abs/squared returns
+- **Signature features** (`features/signature_features.py`): rough path theory via sktime
+- **Range features** (`features/range_features.py`): high-low range estimators
+
+### Data pipeline
+- `pipeline.py` provides `VolDatasetSpec` and `build_vol_dataset()` for dataset construction
+- Point-in-time (PIT) alignment via `asof_utc` columns is required throughout
+- `storage.py` provides DuckDB-backed persistence for models and forecasts
+
+### Evaluation
+- Fixed-split evaluation: Train on `[is_index : os_index)`, Test on `[os_index : ]`
+- STES prediction is aligned to next-step variance; compare to shifted target when needed
+- Metrics: RMSE, MAE, QLIKE, hit_rate, correlation (in `evaluation/metrics.py`)
+- Diebold-Mariano test for statistical comparison (`evaluation/dm_test.py`)
+
+## Conventions
+
+- **Naming**: PascalCase for classes, snake_case for functions/methods, leading underscore for private
+- **Minimal edits**: prefer small, targeted changes over refactors
+- **Compatibility**: maintain alphaforge conventions
+- **Tests**: keep tests aligned with agreed conventions; test files mirror source structure
+- **Caching**: Tiingo caching is pluggable via alphaforge cache backends; use CLI flags when available
+
+## Configuration
+
+- `config/model_config.yaml` — model definitions, hyperparameters, feature lists
+- `config/data_config.yaml` — data sources, tickers (SPY), date ranges
+- `config/app_config.yaml` — application settings
+
+## Dependencies
+
+Core: numpy, pandas, scipy, scikit-learn, joblib, requests, sqlalchemy, pandas-market-calendars, python-dotenv, alphaforge, duckdb
+
+Optional: xgboost, sktime (>=0.24.0), torch, jax, flask, apscheduler, matplotlib, seaborn
diff --git a/tests/test_signature_features.py b/tests/test_signature_features.py
index 75d173e..0bb9668 100644
--- a/tests/test_signature_features.py
+++ b/tests/test_signature_features.py
@@ -9,6 +9,7 @@
 from datetime import datetime, timedelta
 
 pytest.importorskip("alphaforge")
+pytest.importorskip("esig")
 
 from volatility_forecast.features.signature_features import SignatureFeaturesTemplate
 from volatility_forecast.sources.simulated_garch import SimulatedGARCHSource