ForecastingStocks

ARIMA is the classical framework for forecasting a stationary time series from its own past. The name spells out its three parts: AR, I, MA.

The three components

AR (AutoRegressive) — today's value is a weighted sum of its own recent past values. An AR(1) says today depends on yesterday plus noise.
I (Integrated) — the order of differencing applied to make the series stationary. For prices, differencing once (turning them into returns) is usually enough.
MA (Moving Average) — today's value depends on recent forecast errors, letting the model absorb shocks gracefully.

An ARIMA(p, d, q) model has p autoregressive terms, d differences, and q moving-average terms. Choosing those orders — by inspecting autocorrelation plots and comparing information criteria like AIC — is the craft of the method.

What ARIMA can and cannot do

ARIMA is excellent at capturing linear structure and short-term momentum or mean-reversion in a series. It is transparent, fast, and a sensible baseline that any fancier model should be made to beat.

Its limits are real:

It assumes a linear relationship — it cannot capture the nonlinear interactions that machine-learning models can.
On efficient markets the predictable linear component of returns is tiny, so ARIMA forecasts of price level are often barely better than 'tomorrow equals today'.

Treat ARIMA as a baseline and a diagnostic, not a money machine. Its real cousin GARCH, covered next, is far more useful — because the predictable thing in markets is not the return, it is the volatility.