Changeset 2827

Timestamp:

02/18/10 13:52:57 (15 years ago)

Author:

gkronber

Message:

Fixed #879 (Calculation of Theil's inequality coefficient U2 should be based on relative errors instead of absolute errors)

File:

• trunk/sources/HeuristicLab.Modeling/3.2/SimpleTheilInequalityCoefficientEvaluator.cs (modified) (2 diffs)

trunk/sources/HeuristicLab.Modeling/3.2/SimpleTheilInequalityCoefficientEvaluator.cs

@@ -34 +34 @@
 
 where P_t is the predicted change of the target variable and A_t is the measured (original) change.
-(P_t = y'_t - y_(t-1), A_t = y_t - y_(t-1)).
+(P_t = (y'_t - y_(t-1)) / y_(t-1), A_t = (y_t - y_(t-1)) / y_(t-1)).
 
 U2 is 0 for a perfect prediction and 1 for the naive model y'_t = y_(t-1). An U2 > 1 means the
@@ -58 +58 @@
       for (int sample = 1; sample < n; sample++) {
         double prevValue = values[sample - 1, ORIGINAL_INDEX];
-        double estimatedChange = values[sample, ESTIMATION_INDEX] - prevValue;
-        double originalChange = values[sample, ORIGINAL_INDEX] - prevValue;
-        if (!double.IsNaN(originalChange) && !double.IsInfinity(originalChange)) {
-          double error = estimatedChange - originalChange;
-          errorsSquaredSum += error * error;
-          originalSquaredSum += originalChange * originalChange;
+        double estimatedValue = values[sample, ESTIMATION_INDEX];
+        double originalValue = values[sample, ORIGINAL_INDEX];
+        if (!double.IsNaN(originalValue) && !double.IsInfinity(originalValue)) {
+          double errorEstimatedChange = (estimatedValue - originalValue) / prevValue; // percentage error of predicted change
+          errorsSquaredSum += errorEstimatedChange * errorEstimatedChange;
+          double errorNoChange = (prevValue - originalValue) / prevValue; // percentage error of naive model y(t+1) = y(t)
+          originalSquaredSum += errorNoChange * errorNoChange;
           nSamples++;
         }