# How is the MA (moving average model) useful?

How is the MA model useful in modeling financial data, for example the stock indices?

For example, from what i understand in the AR (auto-regressive) model portion, we can use the ADF test to check for the stationarity of the time series. If it is stationary, it is likely that the new trend will follow the old trend.

However, in the case of the MA model, when we suspect that there is a MA component, how do we make use of the knowledge to analyze and predict whether the market movement?

-

In terms of interpretation, an $MA$ model simply means that the time series is a function of the error from previous periods. You might find it informative to consider plotting simple $AR(1)$ models alongside various $ARMA(1,1)$ to develop a more intuitive understanding. For instance, the $AR(1)$ model (chosen as it is common for financial time series) $$x_{t}=\beta x_{t-1}+\epsilon_{t}$$ versus the $ARMA(1,1)$ $$y_{t}=\beta y_{t-1}+\theta\epsilon_{t-1}+\epsilon_{t}$$ for different values of $\theta$ but the same error for each (you may also consider adjusting the mean to ensure it is zero for all). The resulting time series can look very different depending on whether $\theta$ is near $1$ or $-1$. If $\theta$ is near 1, then $y_{t}$ will tend to exhibit some follow through compared to $x_{t}$. By contrast, if $\theta$ is near $-1$, then the series will look more stationary.
For prediction, you can basically just use the formula of whatever $MA$ model it is. Most statistical packages have this functionality built in as well.
In practice, I don't fit a lot of MA models. The main reason is that it is possible to express an $MA(q)$ model as an $AR(\infty)$ model (and vice-versa for expressing $AR(p)$ models as $MA(\infty)$ models). Further, autoregressive models can be fit by least squares, while moving average models cannot (usually maximum likelihood). As a result, rather than spend a lot of time identifying the correct $ARMA(p, q)$ model, it is usually easier to just increase the number of lags in an $AR(p)$ until any moving average components have disappeared from the autocorrelation function (as in the Box-Jenkins methodology) or they're no longer significant or some other approach based on AIC/BIC as in auto.arima for R.