The model that was too perfect
My regression came back with R² = 1.000. That is not a good result — it is a confession. Here is how I found the leak.
When I ran a multiple linear regression on 5,000 student records and R² came back as 1.000, my first feeling was not pride. It was suspicion.
Real data is messy. Students have bad days. Markers are inconsistent. Nothing in the human world lines up perfectly on a regression line. So when a model claims to explain 100% of the variance in an outcome, it usually means one thing: the answer is already hiding in your predictors.
Finding the leak
I went back to the data dictionary and traced how Final Percentage was produced. It wasn't measured. It was calculated — an equal-weighted average of the Math, Science, and English scores, 33.3% each.
So when I asked the model to predict final percentage from those three subject scores, I wasn't building a model. I was rediscovering an arithmetic identity. The regression was solving for weights that already existed in the spreadsheet formula.
This has a name: data leakage. The target variable leaked into the feature set. The model looked brilliant and knew nothing.
Why this matters more than the model
Here is what worries me about this kind of result: it is flattering. A junior analyst under pressure to show impact can screenshot an R² of 1.0 and call it a triumph. Nobody in the room questions a perfect score.
But a perfect score is the loudest possible signal that something is wrong. The habit I brought from physics applies exactly:
Never trust a result you haven't tried to break.
What the data actually said
Once I rebuilt the analysis using genuinely independent predictors, the interesting findings surfaced:
- Internet access was not the divide everyone expected. Students without internet passed at 94.8%. Students with internet passed at 94.6%. A 0.2-point gap — noise, not a barrier.
- Study hours barely moved the needle. The correlation with final grades was r = −0.013. Essentially zero. More hours studied did not mean better outcomes.
Both of those are more useful than a perfect model, because both are surprising and both are actionable. If time invested doesn't predict success but subject mastery does, then the intervention isn't "study more." It's "study the right things, until you actually know them."
The takeaway
A model that fits perfectly is not a finding. It is a question. Ask it.
Next post
Your worst sales day is Tuesday. Here's what to do about it.