Stock Volatility Detection

Real-world example based on the ../examples-notebooks notebook examples/06_real_world_example.ipynb. We use Fast-BOCPD to monitor the US30 (Dow Jones Industrial Average) daily realized volatility from 2008 to 2025 and automatically detect major regime shifts such as the 2008 financial crisis and COVID-19 crash.

Why volatility?

  • Data is naturally positive and skewed, making it a good showcase for the GammaGamma observation model.

  • Regimes last from months to years, which stresses the hazard / run-length configuration.

  • Finance teams care about both historical back-testing and live alerting.

Dataset

  • examples/us30_vol.csv contains normalized daily volatility (already scaled by price) from Jan-2008 to Jan-2025.

  • We split chronologically: 2008–2020 for calibration and 2021–2025 for live simulation—no lookahead.

Training phase (2008–2020)

  1. Estimate Gamma parameters. Method-of-moments gives a shape ≈ 1.06 and rate ≈ 314.3. We wrap those into GammaGamma(alpha0=2, beta0=alpha0/rate, shape=shape)—a weak prior centred on the empirical rate.

  2. Hazard exploration. Compare lambda in {80, 120, 150, 200}. lambda = 150 balances sensitivity and stability; max_run_length = 3 × lambda handles long calm stretches.

  3. Visualization. The notebook plots segment means and MAP run length to verify that large spikes (Lehman, COVID) trigger resets while the long 2010–2019 calm period is kept as a single regime.

Streaming deployment (2021–2025)

  • Fresh detector configured with the chosen hyperparameters. We call fast_bocpd.utils.OnlineChangeDetector with min_cp_prob = 0.3 so alerts fire when P(r_t = 0) crosses 30% or when the MAP run length drops sharply.

  • Stream each day sequentially. Every detection logs the timestamp, posterior probability, and previous segment duration:

    CHANGEPOINT DETECTED
  Date: 2020-03-16
  Index: 271
  P(CP): 85.0%
  Previous segment: 522 days

  • get_segments() and get_map_history() power dashboards that measure how long the market stayed calm vs. turbulent.

How to reuse

  1. Replace us30_vol.csv with your own positive-valued series (e.g. demand, latency, volatility of another symbol).

  2. Redo the Gamma moment-fit and hazard sweep; the notebook exposes helper cells for re-fitting the model.

  3. Plug your alerting backend into OnlineChangeDetector’s stream loop—each fast_bocpd.utils.Changepoint carries probability, segment length, and optional metadata for auditing.

Run the example

jupyter lab examples/06_real_world_example.ipynb

Key takeaways

  • Model choice matters. Positive, heavy-tailed volatility is better served by GammaGamma than Gaussian models.

  • Hazard tuning is crucial. lambda encodes expected regime duration; long calm periods require higher max_run_length to avoid false resets.

  • Online detector = production ready. min_cp_prob, cooldown, and metadata handling make deployments straightforward once the parameters are calibrated.