Model Reports: Response Chart
The Response Chart shows graphically how well each model predicts the records in the analysis set.
The "Current Model" is highlighted in both the drop panel and on the chart. To change the current model, either click on the model name in the drop panel or click directly on the chart near a point along the model curve.
.png)
The Response Chart shows the proportion of data used across the bottom and the proportion of Analysis Records found up the side. The three lines shown represent:
-
The Random line shows the unguided result where neither skill nor knowledge has been used to select records – there is a direct and linear relationship between the number of records selected and the number of analysis records found.
-
The Hindsight line shows the perfect result using our current knowledge of the Analysis set to plot the best outcome the model could possibly achieve.
-
The Modelled line shows the outcome achieved by the model. The steeper the yellow line and the closer it remains to the red line the better.
The blue shaded area shows the segments that are selected in the gains table. You can change this by clicking on the chart and moving the line.
The chart can be zoomed to give a greater detail of points of interest by using the zoom / reset zoom tools on the toolbar.
The status bar underneath the chart displays:
-
The %Total and %Yes figures corresponding to the point selected in the chart.
- The Power of the Current Model.
Power
This is indicated by a value between 0 and 1 and is calculated by dividing A by B, where
A = the area enclosed by the Random and Model lines
B = the area enclosed by the Random and Hindsight lines
It measures the predictive power of the model compared to the best theoretical predictor (represented by the Hindsight line).
The closer the cumulative model line is to the hindsight line the higher the predictive power. If the model is no better than random then it follows the random line and the power = 0. If the model has the maximum possible predictive power it matches the hindsight line and the power = 1.
Using real data you wouldn't expect the model line to be very close to the hindsight (perfect model). If it is then you might suspect that your model references variable(s) that are only known in past data and therefore will not be suitable for prediction.
If the model line dips below the random line then your model is "worse than random". This is not good! Most likely you have got the model score bands sorted in the wrong direction i.e. selecting the lowest scores first.