Model confidence sets and forecast combination: an application to age-specific mortality.

Area of Science:

• Demography
• Statistical Modeling
• Forecasting Science

Background:

• Model averaging enhances forecast accuracy by integrating predictions from multiple statistical models.
• Determining optimal weights for model averaging is critical for forecast precision.
• Sub-optimal weight selection can negatively impact the accuracy of averaged forecasts.

Purpose of the Study:

• To introduce and evaluate a novel model averaging procedure using a model confidence set for improved mortality forecasts.
• To assess the efficacy of trimming models based on statistical significance rather than optimal weight determination.
• To apply and validate the proposed method using Japanese national and sub-national mortality data.

Main Methods:

• The study employs a model averaging approach inspired by the model confidence set (MCS) procedure.
• The MCS procedure incorporates statistical significance testing to identify superior forecasting models.
• The method involves trimming models and then averaging forecasts from the remaining superior models.

Main Results:

• The proposed model-averaged procedure demonstrated superior performance in reducing interval forecast errors for Japanese mortality data.
• The method showed particular effectiveness for male mortality forecasts.
• The trimming approach yielded the smallest forecast errors compared to other methods.

Conclusions:

• The trimming method provides robust out-of-sample point and interval forecasts.
• Robustness is achieved through enhanced resistance to model misspecification.
• This approach offers a reliable alternative for accurate mortality forecasting.