Null Hypothesis

The null hypothesis is the “nothing special is happening” starting point in a statistical test. It’s the claim you assume is true unless the data give you strong enough reason to doubt it.

What it means (without the jargon)
In hypothesis testing, you set up two competing statements:

Null hypothesis (H0): a default claim, usually “no effect,” “no difference,” or “no relationship.”
Alternative hypothesis (H1): what you’d believe if the null looks unlikely (there is an effect/difference/relationship).

Example: If you change a website layout, H0 might be “the conversion rate is the same as before.” You’re not claiming the layout is useless—you’re defining a baseline to test against.

How it’s used in practice
You collect a sample (e.g., a week of conversions), compute a test statistic, and ask: “If H0 were true, how surprising would results like this be?” That “surprise level” is summarized by a p-value. If the p-value is below a chosen threshold (often α = 0.05), you reject H0; otherwise you fail to reject H0 (which is not the same as proving H0 true).

Why it matters in AI/ML
Null hypotheses show up whenever you need to separate real improvements from noise:

A/B testing for product changes: H0 = “no lift in click-through rate.”
Model comparison: H0 = “new model’s accuracy is the same as the old model’s,” accounting for randomness in test samples.
Feature significance in linear/logistic regression: H0 = “this feature’s coefficient is 0.” (Common in statsmodels; related ideas appear in some interpretability workflows.)

Without a clear null, it’s easy to mistake random fluctuation for progress and ship changes that don’t truly help.

Null Hypothesis (H₀) is the default statistical claim that there is no effect, no difference, or no association in the population, used as a baseline for hypothesis testing. Data are summarized into a test statistic and p-value to judge whether observed results are unlikely under H₀. It underpins A/B tests and model comparisons in ML, e.g., H₀: a new recommender does not improve click-through rate.

Imagine you’re testing whether a new recipe really tastes better than the old one. A fair starting point is to assume “nothing has changed” until you have strong evidence.

That starting assumption is the Null Hypothesis. In statistics and AI/ML experiments, it usually means “there is no real effect” or “no difference” (for example, a new model is not actually more accurate than the current model; any improvement is just random luck in the data).

Hypothesis testing checks whether the data looks so unlikely under the Null Hypothesis that you decide to reject it and believe something really changed.