|
Curve Predictor, Analyze Data |
 |
 |
 |
|
Written by Dr. Willy Gerber
|
|
Sunday, 19 June 2011 19:14 |
The system selects for you the best model. You can change this selection by choising a model on the “Prediction” tab. In case you are playing with other models, please take notice of the following characteristics of the algorithm used:
Models with a higher correlation and worst prediction capacity
There are models that can have a high correlation (1) but a strong divergence of the estimated error (2). This means that the light blue curve will better fit the real data but the forecast is uncertain.
Models with a low correlation and apparent good prediction capacity
There are models with a low correlation (1), showing a huge discrepancy between the real data and the estimated values. In this case the model is useless even if the expected error of the prediction is low.
Finding a model by system knowledge
In some cases the system will select a model that doesn’t fit the real mechanism because the model has a lower error in comparison to the equation of the actual mechanism.
The correlation may be high (1) explaining the curve by a “spring type” (2) reaction element with a negative (restitution) force factor (3):
But if you know that the right model is a oscillation system described by a linear restitution force element (2) and a speed dependent damping element (3) the curve will explain the oscillations (4) but have a lower correlation factor (1):
Understanding the model you are selecting is even possible to find data that are not accured and misleading the algorithm selecting the wrong model.
Keep it simple
In some case the simple’s cases are the best ones. For example in the following almost a perfect correlation (1) is achieved by a first order linear model:
If a more complex model is selected a minimal improvement could be achieved in the correlation (0.1%!!!) but some strange behavior will appear. The problem is that the last point of the real data is slightly lower as the previous and the more complex models use this to predict a strong deviation of the original tendency. But at the same time the error functions show a huge increment reveling that the model is not to be considered.
Playing with the data
The data can be manually adjusted by dragging the white points.
Each time the user releases the dot the system recalculates the model for the new point. This allows the user to test the sensibility of the data on the model selected.
|
|
Last Updated ( Sunday, 19 June 2011 20:02 )
|
Curve Predictor, Analyze Data
