PPT2 White Noise, Stationarity, AR Class Notes v2.docxVIP

Stationarity, White Noise and AR Models 5.1 Stationarity – Most important concept in time series analysis Strict Stationarity: All moments are independent of time. Rarely seen in real data. Weak Stationarity: First two moments (ie mean and covariance) are constant ? ACF decays rapidly. Models forecasts used in this class require weak stationarity, at a minimum, as a starting point. Step 1) Trends in volatility ? remove by taking log price Step 2) Trends in mean ? remove by taking first order difference Calculating log returns combine steps 1 and 2 above and is generally used for stock returns. Recall: (5.1) Demo Code: ApplePrice.m Observation: Persistent, strong serial correlation ? strong trend in mean and in volatility DemoCode: AppleLogPrice.m Observation: Taking Log Price reduces trend in price from 2x to 1.1x over time period, but does not weaken serial correlation. Demo Code: AppleLogReturn.m Observation: Taking log return (ie differencing the log price) removes trend in volatility drastically reduces serial correlation. Series is now weakly stationary! Lesson: Taking log returns makes sense for most stocks, but may not improve stationarity for interest rates, currencies or futures. (Why?) Need to always first plot the raw data and evaluate systematically, as in the above example. 5.2 Statistical Testing of Weak Stationarity Test for Constant Mean: Step 1: Plot the raw data {xt]. (Always do this with every new time series!) Step 2: Calculate the sample mean from equation (2.1). Remove the mean from the series by subtraction: (5.2) Step 3: Check if most are within the 5% confidence interval: (5.3) Only about 5% of observations should be outside, and have no specific pattern Step 4: Split the sample in two (or more) sub-periods and check if t