AI 기술 meet-up

Multi seasonal time series analysis: decomposition and forecasting with Python

ARIMA models have some serious drawbacks:

1. the coefficients of the model are not easy to interpret or need detailed explanation
2. efficient for small data sets, it is computationally expensive
3. assumes stationarity of data or else the inputs should be transformed. Consequently, forecasts refer to the transformed data and not to the original time series. Apart from interpretability, this property increases confidence intervals relative to stationary series without transformation

other methods

• Unobserved Components Model (UCM).
  • UCM decompose the original time series to its individual level, trend, cyclic, seasonal components
  • predict future values by modeling and taking the sum of these components
• state space models

statsmodels