1: \begin{abstract}

2: Quasi-Maximum Likelihood (QML) procedures are theoretically appealing and widely used for statistical inference.

3: While there are extensive references on QML estimation in batch settings, it has attracted little attention in streaming settings until recently.

4: An investigation of the convergence properties of the QML procedure in a general conditionally heteroscedastic time series model is conducted, and the classical batch optimization routines extended to the framework of streaming and large-scale problems.

5: An adaptive recursive estimation routine for GARCH models named AdaVol is presented.

6: The AdaVol procedure relies on stochastic approximations combined with the technique of Variance Targeting Estimation (VTE).

7: This recursive method has computationally efficient properties, while VTE alleviates some convergence difficulties encountered by the usual QML estimation due to a lack of convexity.

8: Empirical results demonstrate a favorable trade-off between AdaVol's stability and the ability to adapt to time-varying estimates for real-life data.

9: \end{abstract}

10: