[MATH-1277] Incorrect Kendall Tau calc due to data type mistmatch

• Type: Bug
• Status: Closed
• Priority: Minor
• Resolution: Fixed
• Affects Version/s: 3.5
• Fix Version/s: 4.0, 3.6
• Labels: Kendall correlation tau
• Assignee: Unassigned
• Reporter: Marc Rosen
• Created: 18/Sep/15 21:39
• Updated: 25/Jan/16 20:27
• Resolved: 20/Sep/15 08:15

Description

The Kendall Tau calculation returns a number from -1.0 to 1.0

due to a mixing of ints and longs, a mistake occurs on large size columns (arrays) passed to the function. an array size of > 50350 triggers the condition in my case - although it may be data dependent

the ver 3.5 library returns 2.6 as a result (outside of the defined range of Kendall Tau)

with the cast to long below - the result reutns to its expected value

commons.math3.stat.correlation.KendallsCorrelation.correlation

here's the sample code I used:
I added the cast to long of swaps in the

int swaps = 1077126315;
final long numPairs = sum(50350 - 1);
long tiedXPairs = 0;
long tiedXYPairs = 0;
long tiedYPairs = 0;

final long concordantMinusDiscordant = numPairs - tiedXPairs - tiedYPairs + tiedXYPairs - 2 * (long) swaps;
final double nonTiedPairsMultiplied = 1.6e18;
double myTest = concordantMinusDiscordant / FastMath.sqrt(nonTiedPairsMultiplied);