Realised volatility comes in calm and stressed stretches, the clustering that a GARCH model is built to track
Volatility is not constant. A large move makes the next large move more likely, and a calm stretch tends to be followed by another calm stretch. A GARCH model writes that pattern down. The variance of tomorrow return is a weighted sum of a constant, yesterday squared return and yesterday variance. The parameters are estimated from the return history, usually by maximum likelihood.
The chart shows what a forecast is judged against, not a forecast: twenty day realised volatility of the Nifty close from 1 Jul 2025 to 15 Sep 2026. It is high in some stretches and low in others, and the peak came on 8 Apr 2026. The model itself is fitted to the daily returns, and a series like this one is the check.
What the model adds to a simple average is the weighting. A recent shock counts for more than an old one, and the weights decay over time. That makes the forecast jump after a stressed day and fade over the following weeks, which is the shape a volatility index shows after a sharp fall in the index.
A GARCH forecast depends on the sample, the distribution assumed for returns and the frequency of the data. Parameters fitted on a calm year can understate the next stressed one. A fit is tested against realised volatility that the model was not shown, and a model that fits the past well can still miss the next regime.
GARCH is a volatility model, so it says nothing about the direction of the index. Its forecast can also be set beside option implied volatility. The gap between the two is the subject of the variance risk premium page.
Before you trade. Derivatives trading carries a high risk of loss. SEBI's study "Profitability of Individual Traders in the Equity Derivatives Segment (FY25 to FY26)" (August 2026) found that 87.7% of individual traders in equity F&O made net losses in FY26, with aggregate net losses of Rs 91,685 crore. Read the SEBI study