Data Science Myths

Data science has picked up a lot of myths. Many of them are repeated by smart people, in big meetings. Before you finish, let's replace the most common ones with what's actually true.

• A lot of myths
• Repeated in big meetings
• Replace them with reality

Myth: the data speaks for itself

Data doesn't speak. Someone chooses what to collect, what to measure, and what to compare. Your Summer Panic chart was accurate, and it led people to the wrong conclusion. Every number comes with choices. Your job is to make them visible, and to ask what a number leaves out.

Myth: a bigger sample fixes everything

A big sample reduces random error. It does nothing for bias. Two million people answered that famous nineteen thirty six poll, and it was badly wrong. With huge data, you'll also find tiny, meaningless effects that are statistically significant. So ask who's in the data, and how big the effect really is.

Myth: complex models beat simple ones

In real work, simple usually wins. In our churn lesson, plain logistic regression beat boosting. In our forecast, exponential smoothing beat machine learning. Simple models are easier to explain, easier to fix, and they fail in ways that you can understand. Start simple, and only add complexity when it pays for itself.

Myth: correlation is enough for business

People say, we don't need cause, we just need prediction. That's true for some things. If you're forecasting demand, correlation's fine. But if you're choosing what to do, you need cause. Ice cream predicts drownings well. Banning ice cream won't save anyone. Prediction tells you what will happen, if you do nothing different.

Summary