Something I learned from building alerting and monitoring analytics is if you are doing timeseries smoothing let’s say of daily or hourly data, calculate the avg value NOT of the past continuous x days window, but rather avg the same hours and days of weeks together, meaning avg the values of 7am on Mondays over the past 3 weeks together then take that value. This will often maintain the reality of your data’s daily/hourly volatility or seasonality trend over time. This matters little though if your data does not abide by a daily or seasonal trend.