Outlier detection
Which videos are statistical outliers?
Odinstat flags a video as an outlier when its lifetime views sit more than 3.5 modified Z-scores from your channel’s median, using the median absolute deviation instead of the standard deviation.
Because the baseline is the median, one viral video cannot drag “normal” up and make every other upload look like a failure.
Why the average is the wrong baseline
A channel with nineteen videos at 2,000 views and one at 400,000 has an average of about 22,000 views. Judged against that average, nineteen of twenty uploads “underperformed”. That is not a useful reading of the channel. The median of the same channel is 2,000, and the one video that genuinely broke out is the one that stands apart from it.
YouTube Studio ranks videos by raw views. That tells you the order, not whether the gap between second and third place is noise or signal. Odinstat answers the second question.
How the score is computed
For every video we take lifetime views, compute the median across your channel, then the median absolute deviation (MAD) from that median. The MAD is scaled by 1.4826 so it behaves like a standard deviation for normally distributed data. Each video’s modified Z-score is its distance from the median divided by the scaled MAD.
Anything beyond 3.5 in either direction is an outlier: a breakout above, or a video that quietly died below. The 3.5 cutoff is the standard Iglewicz–Hoaglin threshold for this estimator, not a tuning knob we picked to make results look interesting.
What you do with it
Positive outliers are the videos to study: topic, packaging, and the traffic mix that carried them. Odinstat shows each one’s traffic sources beside the score, so you can tell a Shorts-feed hit from a search hit from a paid promotion. Negative outliers are where a fixable problem usually lives.
A video that is an outlier because you ran ads on it is still an outlier; it is just not evidence about your organic content. Tell the Channel Assistant a video was promoted and it excludes that video from organic comparisons for the rest of the conversation.
What you need
At least 5 videos. Fewer than that and the median and MAD are not stable enough to call anything an outlier, so Odinstat says “not enough data yet” instead of guessing.
What this is not
- Comparing your channel to other channels. The baseline is always your own history.
- Daily anomaly alerts. Odinstat has a separate check for that (a video’s latest day against its previous four same weekdays); this page is about lifetime performance.
Odinstat runs this on your own channel, read-only, with a 14-day free trial from $10/month.