Hacker News Reader

Error bars on what? The ratings themselves can be measured with perfect accuracy, and the results were averaged over 5000 merchants, so any error in the mean would be too small to see.

View on HN

It's true that the average clearly moved, but the author was trying to make a generalized inference from the average, that there was a structural break in the "true average" insofar as the measured average is just a sample/noisy measurement of the "true average". Therefore, error bars/confidence bars would indicate how likely the structural break in measured outcomes is to reflect a structural break in the underlying "true average."

The volatility/variance comes from the fact that 1. not everyone rates their experience, 2. there's variability in the experience, 3. different people have different expectations or likelihood to rate high/low, etc.

I hope that clarified what I believe was the parent's meaning.

That break at day 0 was because they reset the moving average calculation on that date.

With error bars/lines on that plot, this discussion wouldn't have been needed. There are nice ways of doing error contours - showing the 1% quantile lines on top and on the bottom.

Reply on news.ycombinator.com