BNK/CGM48 4th General Election Predictive Analysis
An end-to-end machine learning case study — scraping to prediction — forecasting BNK48/CGM48 general election outcomes in the entertainment industry.
- Result
- Predicted each member’s election rank to within about 7 places on average (leave-one-out mean absolute error 7.22, labelled RMSE in the notebook, with preliminary-round results). With no data from the election itself, past popularity and visibility explained about half the variation in rank (R² ≈ 0.50).
- My role
- Solo; my first end-to-end machine learning project.
Problem
Can public statistics on each BNK48 and CGM48 member predict their rank in the groups’ 4th General Election? A fun personal project, and my first end-to-end machine learning work.
Approach
Past election results, fan-app activity, theatre appearances and time as centre were collected from public sources. Variables known only after the result (final tokens, wallets) were dropped, and a second model was built with no data from the election at all. With one row per member, leave-one-out cross-validation replaced 5-fold, whose fold errors swung from 7.2 to 11.2 ranks.
Outcome
Rank was predicted to within about 7 places on average (leave-one-out mean absolute error 7.22, labelled RMSE in the notebook, with preliminary-round results). Without any data from the election, past popularity and visibility explained about half the variation in rank (R² ≈ 0.50); times as centre and previous election rank were the strongest predictors of making the senbatsu.