I'm trying to get a sense check here. When determining "forecastability" for sales data, I tend to use the CV. However, this is highly susceptible to seasonality and outliers. As such, I was wondering:
- Does the CV by itself account for outliers (both SD and mean are susceptible to outliers) or would a loess decomposition lead to a more "real" output?
- Is it better to difference the time series and remove trends before conducting CV analysis so that an increasing sales volume over years is not seen as "bad" for forecasting?