22

I recently designed a time series module where my time series is essentially a SortedDictionnary<DateTime, double>.

Now I would like to create unit tests to make sure that this module is always working and producing the expected result.

A common operation is to compute the performance between the points in the time series.

So what I do is create a time series with, say, {1.0, 2.0, 4.0} (at some dates), and I expect the result to be {100%, 100%}.

The thing is, if I manually create a time series with the values {1.0, 1.0} and I check for equality (by comparing each point), the test would not pass, as there will always be inaccuracies when working with binary representations of real numbers.

Hence, I decided to create the following function:

private static bool isCloseEnough(double expected, double actual, double tolerance=0.002)
{
    return squaredDifference(expected, actual) < Math.Pow(tolerance,2);
}

Is there another common way of dealing with such a case?

PersonalNexus
  • 3,029
SRKX
  • 1,939
  • 1
  • 15
  • 24

3 Answers3

12

I can think of two other ways to deal with this problem:

You can use Is.InRange:

Assert.That(result, Is.InRange(expected-tolerance, expected+tolerance));

You can use Math.Round:

Assert.That(Math.Round(result, sigDigits), Is.EqualTo(expected));

I think that both ways are more expressive than a dedicated function, because the reader can see precisely what's going on with your number before it gets compared to the expected value.

Sergey Kalinichenko
  • 17,591
9

As RichardM suggested, if you do not use NUnit, the best way seems to be using Assert.AreEqual(double,double,double), where the last one is the precision.

SRKX
  • 1,939
  • 1
  • 15
  • 24
4

It depends what you do with the numbers. If you are testing a method which is supposed to e.g. select an appropriate value from an input set based on some criteria, then you should test for strict equality. If you're doing floating-point calculations, usually you will need to test with a non-zero tolerance. How big the tolerance is depends on the calculations, but with double precision a good starting point is to choose 1E-14 relative tolerance for simple calculations and 1E-8 (tolerance) for more complicated ones. YMMV of course, and you need to add some small absolute tolerance if the expected result is 0.

quant_dev
  • 5,227