intraday-vol-mlops

An end-to-end MLOps pipeline for intraday realized-volatility forecasting — synthetic data → feature engineering → walk-forward training with MLflow tracking → conformalised FastAPI inference → Prometheus/Grafana monitoring → drift detection → Rust backtester. The repository is structured to demonstrate the full machine-learning engineering lifecycle on a stock-market problem: turning streaming OHLCV bars into calibrated, uncertainty-aware variance forecasts that a trading service could actually consume.

---

Lifecycle at a glance

                           ┌──────────────────────────────────────────────────────┐
                           │              intraday-vol-mlops pipeline             │
                           └──────────────────────────────────────────────────────┘

   ┌─────────────┐    ┌───────────────┐    ┌────────────────┐    ┌─────────────────┐
   │ Synthetic   │───▶│  Feature      │───▶│  Walk-forward  │───▶│  Model Registry │
   │ OHLCV gen   │    │  pipeline     │    │  training      │    │  (joblib bundle)│
   │ (GBM+GARCH+ │    │  (sklearn)    │    │  + conformal   │    │  + MLflow runs  │
   │  regimes)   │    │               │    │  + metrics     │    │                 │
   └─────────────┘    └───────────────┘    └────────────────┘    └─────────────────┘
          │                   │                    │                      │
          │                   │                    │                      ▼
          │                   │                    │            ┌──────────────────┐
          │                   │                    │            │ FastAPI service  │
          │                   │                    │            │  /score /healthz │
          │                   │                    │            │  /metrics (Prom) │
          │                   │                    │            └──────────────────┘
          │                   │                    │                      │
          │                   │                    ▼                      ▼
          │                   │           ┌──────────────────┐    ┌──────────────────┐
          │                   └──────────▶│  Drift detector  │    │  Rust backtester │
          │                               │  (PSI + KS CLI)  │    │  POSTs /score,   │
          └──────────────────────────────▶│                  │    │  vol-targets,    │
                                          └──────────────────┘    │  PnL vs B&H      │
                                                                  └──────────────────┘
                                                  │
                                                  ▼
                                          ┌────────────────────┐
                                          │ Prometheus +       │
                                          │ Grafana dashboard  │
                                          └────────────────────┘

Quick start

# 1. Install (editable, with dev extras)
pip install -e ".[dev]"

# 2. Generate synthetic bars + labels
intraday-vol generate-data --output data/raw/synth.parquet --n-bars 6500

# 3. Train (writes model + conformal bundles to models/latest/)
intraday-vol train --model lightgbm --data-path data/raw/synth.parquet

# 4. Run the inference service locally
intraday-vol serve --port 8000

# 5. In another terminal: probe it
curl -s localhost:8000/healthz
curl -s -X POST localhost:8000/score \
    -H 'content-type: application/json' \
    --data @examples/sample_score_request.json | jq

# 6. Full stack (API + Prometheus + Grafana) via Docker
docker compose up --build
# Grafana: http://localhost:3000  (anonymous Viewer enabled)

Models

Name	Family	Features	Calibrated [0,1] head	Interval
lightgbm	Gradient boosting	27 tabular features	Platt-scaled logistic on ŷ	Conformal
xgboost	Gradient boosting	27 tabular features	Platt-scaled logistic on ŷ	Conformal
lstm	PyTorch 1-layer LSTM	11-feature sequence	Platt-scaled logistic on ŷ	Conformal

All three implement the same BaseModel interface (fit, predict, predict_proba_high_vol, save, load), share the joblib bundle format ({model, scaler, feature_order, calibrator, metadata}), and are swappable via the --model flag in the training CLI.

Components

Layer	Lives in	Notes
Schemas	src/intraday_vol/schemas.py	Pydantic v2; OHLC + payload validation; shared with the Rust client.
Data	src/intraday_vol/data/	GBM + GARCH(1,1) + Markov regimes + Poisson jumps + diurnal U-shape.
Features	src/intraday_vol/features/	Sklearn transformers; tabular + sequence feature columns.
Models	src/intraday_vol/models/	LightGBM, XGBoost, PyTorch LSTM behind a single interface.
Training	src/intraday_vol/training/	Chrono splits, walk-forward, metrics, split-conformal.
Inference	src/intraday_vol/inference/	FastAPI + Prometheus instrumentation.
Drift	src/intraday_vol/drift/	PSI + two-sample KS, with a CLI for baseline-vs-current comparisons.
CLI	src/intraday_vol/cli.py	`generate-data
Observability	prometheus/, grafana/	Provisioned datasource + dashboard JSON.
Rust backtester	rust-backtester/	Vol-targeted backtest replaying bars through /score.
Tests	tests/	≥20 pytest tests covering data, features, models, drift, API, CLI.
CI	.github/workflows/ci.yml	Ruff + AST parse + pytest + cargo build/test + docker build.

Metrics evaluated

• Point forecast: MAE and RMSE on log-variance, QLIKE on raw variance.
• High-vol classifier head: PR-AUC, ROC-AUC, precision@K (top-decile threshold).
• Intervals: coverage at 90% nominal, mean interval width (log scale).

See docs/MODEL_CARD.md for intended use, out-of-scope claims, and known limitations. See docs/ARCHITECTURE.md for the deeper system design write-up.

License

MIT.