Master Time Series Forecasting Accuracy

Key Points on Time Series Forecasting and Accuracy

• Average method forecasts future values as the mean of historical data.
• Naive method forecasts equal to the last observed value.
• Seasonal naive method forecasts equal to the last value from the same season.
• Drift method forecasts equal to the last value plus average change, using Box-Cox transformations for λ close to zero.
• Residuals in forecasting are the difference between observed value and its fitted value.
• Partitioning is the splitting of data into training and test sets for forecasting accuracy.
• A model that fits the training data well will not necessarily forecast well.
• MAE, MSE, and RMSE are all scale-dependent, while MAPE is scale-independent but only sensible if yt≫ 0 for all t.
• The Mean Absolute Scaled Error (MASE) is a scale-independent measure that can handle zero counts.
• Time series cross-validation is a more sophisticated version of training/test sets.
• A good way to choose the best forecasting model is to find the model with the smallest RMSE computed using time series cross-validation.
• Prediction intervals require a stochastic model, and multi-step forecasts for time series require a more sophisticated approach.