What needed to be understood.
A model can look outstanding while answering the wrong business question. The target was derived from global sales, while regional sales were available as predictors. That created a retrospective branch that was useful for explaining past performance but not suitable for pre-release decision-making.
How the analysis was built.
- Cleaned and prepared the video-game sales dataset.
- Built separate analyst and developer modeling branches.
- Compared logistic regression, decision tree, and random forest models using cross-validation.
- Separated post-release sales variables from variables available before launch.
- Evaluated accuracy, AUC, precision, recall, and F-measure instead of relying on one headline score.
What the evidence showed.
- The analyst branch reached near-perfect performance because regional sales closely determine global sales.
- The developer branch fell to roughly 58.5% accuracy and 0.625 AUC after post-release signals were removed.
- Genre and platform contained some useful signal, but the available data omitted marketing investment, reviews, brand strength, launch timing, competition, and audience demand.
What should happen next.
- Use the analyst branch for retrospective market explanation, not launch prediction.
- Collect pre-release variables such as marketing budget, franchise strength, review forecasts, wishlists, and competitive release windows.
- Define feature availability at the decision date before model training to prevent leakage.
What this project does—and does not—prove.
The project uses a historical public dataset and does not include several commercial variables needed for a deployable launch model. The title-text workflow was exploratory and was not represented as improving the final developer classifier.