Volatility Forecast Models

Canonical definition, formula, interpretation, and API reference.

Definition

Five models that forecast a symbol's variance over the next 1, 5 and 21 sessions, using 5-minute realized variance and 30-day implied vol. Each is fitted once a day, after the session reaches the archive.

Formula
RVₜ = overnight return² + Σ r² (5-minute returns)
HAR: mean RV over next h = β₀ + βd RVₜ + βw RV(5 sessions) + βm RV(22 sessions)
Reported vol = 100 × √(252 × variance)

Where RV is realized variance and h is 1, 5 or 21 sessions. The other four models extend this measure: see The Five Models below.

Inputs
5-minute stock midsovernight return30-day ATM IV
Output
har_rv_5mhar_ivharqrealized_garchheavy
Interpretation
  • Each forecast is annualized vol in percent, for the mean daily variance over its horizon.
  • A 1-session forecast above the 21-session forecast: the model expects recent variance to fade.
  • Compare the 21-session forecast with iv30, the 30-day implied vol at the same close.

API Reference

Endpoint
GET /v1/volatility/forecast/{symbol}
Point in time
GET /v1/volatility/forecast/{symbol}?at=... (historical API)
Tier
Alpha+
Response field
realized_models.as_of_session, computed_at, iv30, models[].model, models[].forecast, models[].params

How to Read the Forecasts

Forecast above iv30

The model expects more variance over 21 sessions than 30-day implied vol prices at the same close.

Forecast below iv30

The model expects less variance over 21 sessions than 30-day implied vol prices at the same close.

Models far apart

A wide spread across models shows how much the answer depends on the model.

Example: AAPL at the close on November 6, 2026

Model1 session5 sessions21 sessions
har_rv_5m38.8533.7129.82
har_iv34.2329.1727.02
harq39.4433.4729.99
realized_garch36.3733.1328.22
heavy30.8929.8927.61

Annualized vol in percent. 30-day implied vol at the same close: 25.48%.

The Five Models

har_rv_5mHAR-RV (Corsi 2009) on 5-minute realized variance plus the overnight return. One least-squares regression per horizon on the daily, 5-session and 22-session means.
har_ivHAR-RV with 30-day implied variance added as a regressor.
harqHARQ (Bollerslev, Patton and Quaedvlieg 2016). Lowers the daily weight when realized variance is measured with more noise, using realized quarticity.
realized_garchLog-linear Realized GARCH(1,1) (Hansen, Huang and Shek 2012), fitted by Gaussian maximum likelihood.
heavyHEAVY (Shephard and Sheppard 2010). Two equations driven by intraday realized variance, each fitted by Gaussian quasi-maximum likelihood.

How the forecasts are produced

  • Fit window. Up to 1,000 sessions ending at the forecast session.
  • Schedule. Once a day, after the session reaches the archive. as_of_session and computed_at state what each block is based on.
  • Point in time. The historical API runs the same models on data through any past session's close, with no look-ahead.
  • Compared with har_rv. har_rv uses squared daily returns. har_rv_5m uses 5-minute returns plus the overnight return.
Ship faster

Live for 6,000+ US symbols. One API call, sub-200ms.

Stop scraping chains and coding Black-Scholes from scratch. FlashAlpha computes GEX, DEX, VEX, CHEX, 15 BSM Greeks, SVI surfaces, max pain, VRP and more. Free tier, no credit card required.

Free · 5/day
Single-expiry GEX (equities), key levels, gamma flip, 15 BSM Greeks. No card.
Basic · 250/day
DEX, VEX, CHEX, full max pain, plus ETF & index symbols.
Growth & Alpha
Full-chain GEX, 0DTE & flow (Growth); SVI surfaces, VRP, directional VRP (Alpha).

Comparing providers? GEX and VRP API providers · GEX APIs compared · Why GEX numbers differ · Free GEX tools · VRP APIs compared · Max pain tools