Moving Average Models: Autocorrelation, Estimation, and Residual Checks
Summary
The document introduces the MA(q) time-series model, in which each observation depends on a finite number of current and past white-noise shocks. It explains that the autocorrelation function should cut off beyond lag q, then illustrates model identification and estimation with simulated MA(1) and MA(3) series. Fitted coefficients and confidence intervals are compared with the simulation inputs, while correlograms show how sampling can produce occasional significant peaks beyond the expected cutoff.
The article then fits low-order MA models to daily log returns for Amazon and the S&P 500 and checks residual autocorrelation. The S&P 500 fits retain significant residual lags, so the tested MA models do not adequately capture its dependence. The examples show why residual diagnostics matter, but they do not establish a profitable forecasting strategy. The source also notes that volatility clustering and long-memory effects remain and motivates considering ARMA and heteroskedastic models.
Key ideas
- An MA(q) process models observations as combinations of a finite number of white-noise shocks.
- For an MA(q) process, theoretical autocorrelation is zero at lags greater than q, though finite samples may show incidental peaks.
- Correlograms can guide selection of q, while residual autocorrelation helps assess whether a fitted model is adequate.
- Simulated MA(1) and MA(3) examples show parameter recovery with uncertainty intervals.
- The tested low-order MA models leave serial dependence in S&P 500 returns, and the article flags volatility clustering and long memory as unresolved.
Tags
This summary was written by Stratmill's research agent from the original; it is not a copy of the source.