1a - Exploratory analysis Flashcards
What is the classical time series decomposition?
Classical time series decomposition describes a univariate time series (Yt) with t≥1 as composed by structural components such as a trend, a seasonal compoment, a cycle, and an erratic component.
How can a classical decomposition be?
- additive: Yt =Tt +St +Ct +Et
- multiplicative: Yt = Tt ∗ St ∗ Ct ∗ Et
Notice that a log transforms a multiplicative decomposition into an additive one: logYt =logTt +logSt +logEt
How do we fit a trend in a series that just shows a trend and no seasonal component? Like Yt = Tt + Et
- Fit a smooth function of t
- Use moving averages
How do we fit a smooth function of t?
We can use for example a linear trend or we can use more complex functions such as a quadratic trend or an exponential trend.
It is important to notice that bein too flexible makes it difficult to identify a trend or a cycle.
How do we use a moving average?
This gives a smoothed version of the observed series, intrpreted as the trend.
- The higher the k, the smoother the fitted trend.
What happens if k is even when we use the moving average?
How do we detrend?
- Having fitted the trend, we can remove it to obtain a detrended time series.
- In this way we get just the erratic component which is possibly stationary
How do we fit a seasonal component in a series that just shows a seasonal component and no trend? Like Yt = St + Et
- Seasonal factors
- Moving averages
How do we use the season factors to fit the seasonal component?
- Suppose you have monthly data y , with mean y ̄ = 0.
- A simple way to proceed is to introduce seasonal factors α(jan), α(feb), …, α(dec) and describe:
yt = αmonth(t) + Et
where, if t corresponds to January, αmonth(t) = α(jan), and so on. - For identifiability, we assume that the sum of the seasonal factors is zero.
- Remeber we are only considering additive decompostition.
How do we use the moving average to fit the seasonal component?
For example, if we have monthly data, we would use a MA(12).
What if we have a time series that has both the trend and the seasonal component? Like yt =Tt +St +Et
We can proceed in two ways:
1. we first detrend, and the we fit the seasonal component
2. we first fit the seasonal component, and then on the deseasonalized series we fit the trend