Why predicting NBA career length is harder than it looks

For my Data Science capstone, my team tried to predict how long an NBA player's career would last using their rookie-year stats. It seemed straightforward. It wasn't.

Our first model was overconfident and wrong. We'd fed it points, rebounds, assists — the obvious stuff — and it kept predicting long careers for high-scorers regardless of anything else. Shaquille O'Neal's rookie stats would've gotten a 20-year projection. That's not a model, that's a fan.

The fix came from actually thinking about why careers end. Injuries. Inefficiency at scale. Reduced role as the league evolves around you. Once we reframed the question — not "how good is this player?" but "how durable and adaptable is this player?" — the features changed completely.

True Shooting % beat raw PPG. Games played per season (an injury-history proxy) was one of our strongest signals. Usage rate balance — whether a player could produce without dominating the ball — predicted longevity better than almost anything.

We used a Random Forest Regressor and got the MSE down to a number we were happy with. But what actually took 80% of our time was data cleaning — specifically, normalizing across eras. A 15 PPG season in 1994 is a completely different thing than 15 PPG in 2024. Pace, three-point volume, defensive rules — the game changed. The data didn't know that until we told it.

The real lesson: the model doesn't know what it doesn't know. Your job as the engineer is to encode the context that doesn't show up in a CSV.