Six transformations, and why they are not variations of one idea
The word curve covers several different operations that behave nothing alike. What separates them is not how much they raise the mean — you can tune any of them to hit the same mean — but what they do to the distance between students.
| Method | The operation | Effect on the gaps between students |
|---|---|---|
| Flat addition | every score plus a constant | None. A student five points ahead stays five points ahead. |
| Shift the mean | plus a constant chosen to land the mean on a target | None. Identical to the above with the constant computed for you. |
| Scale to the top score | multiply by target over highest raw | Multiplies the gaps too. A class where the top score was 80 gets every gap widened by a quarter. |
| Square root | maximum times the square root of the fraction earned | Compresses. Large gains low down, almost none at the top. |
| Standard deviation rescale | target mean plus z score times target deviation | Sets the spread to whatever you asked for, regardless of what it was. |
The two additive methods are the only ones that preserve rank distances exactly. Everything else redistributes, and the redistribution is the part that gets argued about, because it is where a curve stops being generosity and starts being a re-ranking of the class.
Scaling to the top score has a failure mode
Multiplying every score so the highest becomes 100 is intuitive and widely used, and it hands its largest gift to the student who needed it least. If the top raw score was 80, the multiplier is 1.25. The top student gains 20 points, a student at 60 gains 15, and a student at 40 gains 10. The gap between the top two students widens along with everything else, so the ranking is unchanged but the differences all inflate.
The other exposure is that the whole class result now depends on one paper. One unusually strong student compresses the curve for everybody, and one absent student who would have scored highest changes every other grade in the room. A method whose output is that sensitive to a single observation is worth noticing before it is applied to a permanent record.
The square root curve, worked out
On a test out of 100, the square root curve replaces a raw score of 64 with 80, because the square root of 64 is 8 and 8 times 10 is 80. Written generally for a test out of M points, the new score is M times the square root of x divided by M.
Its shape is fixed and worth memorising because it removes any guessing about how generous it is. The largest gain lands exactly a quarter of the way up the scale, and that gain is exactly a quarter of the total points. Out of 100, a raw 25 becomes a 50. A raw 0 stays 0 and a raw 100 stays 100 — the curve cannot move either end. In the middle it is still substantial: a 49 becomes a 70, and an 81 becomes a 90.
That profile is why it is chosen when a test came out harder than intended for the weaker half of the class while the top of the class performed as expected. It is also why it is objected to: it changes the distance between a student who understood most of the material and one who understood a quarter of it, and it changes it a lot.
What a curve does to the record
An uncurved percentage claims to be a statement about material. Seventy-two percent says roughly that seventy-two percent of what was asked was answered. After any curve, that claim no longer holds, and the number is instead a statement about where a student sat relative to this class on this test. Both are legitimate things to record and they are not the same thing, which is the whole of the pedagogical disagreement in a sentence.
This matters practically because grades travel. A transcript does not carry a note about the curve. A student who moves districts, or a committee reading a file, sees the number and reads it as the first kind of claim. None of that makes curving wrong — a badly calibrated test produces a misleading record too — but it is the reason many schools have a written position on it, and some prohibit it. Find out what yours says. This page will compare methods for you and will not tell you which to use.
Related arithmetic on the site: the weighted gradebook calculator produces the course grades that a curve would be applied to, class rank places a single score inside a distribution with ties handled, and the test average calculator works the same weighted arithmetic from the student side.
Questions people ask
Which curve method should I use?
This page will not answer that, and it is not being coy. The methods differ in what they do to the spread rather than in how much they help, and which of those effects is appropriate depends on why the raw scores came out where they did — a badly worded question, material that was not covered, a test that was too long for the period, or a class that genuinely did not learn it. Those causes call for different responses, and one of the available responses is not curving at all. What the page does is show the consequence of each choice on your actual numbers so the decision is made with the distribution in front of you.
Is curving allowed?
It depends entirely on where you teach. Some districts have written grading policy that permits it, some require department or administrator approval, some prohibit it, and some say nothing and leave it to custom. There are also assessment types where it is clearly inappropriate regardless of local policy, such as anything reported against a fixed standard. Check your handbook and ask before applying a curve to grades that go on a record. Nothing on this page constitutes a statement about what is permitted anywhere.
What is the difference between the flat method and shifting the mean?
Nothing, except who chooses the constant. Adding five points to everyone and shifting the mean to a target both add the same number to every score, so both leave the spread untouched and both preserve every gap between students exactly. The tool shows them as separate rows because teachers arrive at the same operation from two directions — sometimes you know you want to give five points back on a question that was misprinted, and sometimes you know the class mean has to land near a particular figure. If both rows show in your results, compare the constants: they tell you how far off the target the flat number you had in mind actually was.
Why is the standard deviation given as a population figure?
Because the class is the whole group being described rather than a sample drawn from something larger. The population deviation divides by the number of scores; the sample version divides by one less than that, which is the right correction when you are estimating a wider population from a subset. For a set of thirty test papers where the thirty papers are the entire thing you care about, the population figure is the correct one. The difference is small at class sizes anyway — under two percent at thirty scores — and it does not change any comparison between methods here, since every method is measured the same way.
Can I curve to a fixed distribution, like ten percent A grades?
That is a different operation again, usually called norm referencing or grading on a strict bell, and it is deliberately not offered here. It assigns grades by rank position regardless of what was scored, which means a fixed number of students receive each grade whether the class did well or badly. It is far more strongly restricted than the methods on this page and is prohibited in many places. If you want to see how a given method would fall out across bands, the band count table in the results does that without imposing a target distribution — enter your scale and read the counts.