When assessing the quality of a model, being able to accurately measure its prediction error is of key importance. Often, however, techniques of measuring error are used that give grossly misleading results. This can lead to the phenomenon of over-fitting where a model may fit the training data very well, but will do a poor job of predicting results for new data not used in model training. Here is an overview of methods to accurately measure model prediction error.

# Measuring Error

When building prediction models, the primary goal should be to make a model that most accurately predicts the desired target value for *new* data. The measure of model error that is used should be one that achieves this goal. In practice, however, many modelers instead report a measure of model error that is based not on the error for *new* data but instead on the error the *very same* data that was used to train the model. The use of this incorrect error measure can lead to the selection of an inferior and inaccurate model.

Naturally, any model is highly optimized for the data it was trained on. The expected error the model exhibits on new data will *always* be higher than that it exhibits on the training data. As example, we could go out and sample 100 people and create a regression model to predict an individual’s happiness based on their wealth. We can record the squared error for how well our model does on this training set of a hundred people. If we then sampled a different 100 people from the population and applied our model to this new group of people, the squared error will almost always be higher in this second case.

It is helpful to illustrate this fact with an equation. We can develop a relationship between how well a model predicts on new data (its true prediction error and the thing we really care about) and how well it predicts on the training data (which is what many modelers in fact measure).

Here, *Training Optimism* is basically a measure of how much worse our model does on new data compared to the training data. The more optimistic we are, the better our training error will be compared to what the true error is and the worse our training error will be as an approximation of the true error.

## The Danger of Overfitting

In general, we would like to be able to make the claim that the optimism is constant for a given training set. If this were true, we could make the argument that the model that minimizes training error, will also be the model that will minimize the true prediction error for new data. As a consequence, even though our reported training error might be a bit optimistic, using it to compare models will cause us to still select the best model amongst those we have available. So we could in effect ignore the distinction between the true error and training errors for model selection purposes.

Unfortunately, this does not work. It turns out that the optimism is a function of model complexity: as complexity increases so does optimism. Thus we have a our relationship above for true prediction error becomes something like this:

How is the optimism related to model complexity? As model complexity increases (for instance by adding parameters terms in a linear regression) the model will always do a better job fitting the training data. This is a fundamental property of statistical models ^{1}. In our happiness prediction model, we could use people’s middle initials as predictor variables and the training error would go down. We could use stock prices on January 1st, 1990 for a now bankrupt company, and the error would go down. We could even just roll dice to get a data series and the error would still go down. No matter how unrelated the additional factors are to a model, adding them will cause training error to decrease.

But at the same time, as we increase model complexity we can see a change in the true prediction accuracy (what we really care about). If we build a model for happiness that incorporates clearly unrelated factors such as stock ticker prices a century ago, we can say with certainty that such a model must necessarily be worse than the model without the stock ticker prices. Although the stock prices will *decrease our training error* (if very slightly), they conversely must also *increase our prediction error on new data* as they increase the variability of the model’s predictions making new predictions worse. Furthermore, even adding clearly relevant variables to a model can in fact increase the true prediction error if the signal to noise ratio of those variables is weak.

Let’s see what this looks like in practice. We can implement our wealth and happiness model as a linear regression. We can start with the simplest regression possible where Happiness=a+b Wealth+ϵHappiness=a+b Wealth+ϵ and then we can add polynomial terms to model nonlinear effects. Each polynomial term we add increases model complexity. So we could get an intermediate level of complexity with a quadratic model like Happiness=a+b Wealth+c Wealth2+ϵHappiness=a+b Wealth+c Wealth2+ϵ or a high-level of complexity with a higher-order polynomial like

Happiness=a+b Wealth+c Wealth2+d Wealth3+e Wealth4+f Wealth5+g Wealth6+ϵHappiness=a+b Wealth+c Wealth2+d Wealth3+e Wealth4+f Wealth5+g Wealth6+ϵ.

The figure below illustrates the relationship between the training error, the true prediction error, and optimism for a model like this. The scatter plots on top illustrate sample data with regressions lines corresponding to different levels of model complexity.

Increasing the model complexity will always decrease the model training error. At very high levels of complexity, we should be able to in effect perfectly predict every single point in the training data set and the training error should be near 0. Similarly, the true prediction error initially falls. The linear model without polynomial terms seems a little too simple for this data set. However, once we pass a certain point, the true prediction error starts to rise. At these high levels of complexity, the additional complexity we are adding helps us fit our training data, but it causes the model to do a worse job of predicting new data.

This is a case of *overfitting* the training data. In this region the model training algorithm is focusing on precisely matching random chance variability in the training set that is not present in the actual population. We can see this most markedly in the model that fits every point of the training data; clearly this is too tight a fit to the training data.

Preventing overfitting is a key to building robust and accurate prediction models. Overfitting is very easy to miss when only looking at the training error curve. To detect overfitting you need to look at the true prediction error curve. Of course, it is impossible to measure the exact true prediction curve (unless you have the complete data set for your entire population), but there are many different ways that have been developed to attempt to estimate it with great accuracy. The second section of this work will look at a variety of techniques to accurately estimate the model’s true prediction error.

## An Example of the Cost of Poorly Measuring Error

Let’s look at a fairly common modeling workflow and use it to illustrate the pitfalls of using training error in place of the true prediction error ^{2}. We’ll start by generating 100 simulated data points. Each data point has a target value we are trying to predict along with 50 different parameters. For instance, this target value could be the growth rate of a species of tree and the parameters are precipitation, moisture levels, pressure levels, latitude, longitude, etc. In this case however, we are going to generate every single data point completely randomly. Each number in the data set is completely independent of all the others, and there is no relationship between any of them.

For this data set, we create a linear regression model where we predict the target value using the fifty regression variables. Since we know everything is unrelated we would hope to find an *R ^{2}* of 0. Unfortunately, that is not the case and instead we find an

*R*of 0.5. That’s quite impressive given that our data is pure noise! However, we want to confirm this result so we do an

^{2}*F*-test. This test measures the statistical significance of the overall regression to determine if it is better than what would be expected by chance. Using the

*F*-test we find a

*p*-value of 0.53. This indicates our regression is not significant.

If we stopped there, everything would be fine; we would throw out our model which would be the right choice (it is pure noise after all!). However, a common next step would be to throw out only the parameters that were poor predictors, keep the ones that are relatively good predictors and run the regression again. Let’s say we kept the parameters that were significant at the 25% level of which there are 21 in this example case. Then we rerun our regression.

In this second regression we would find:

- An
*R*of 0.36^{2} - A
*p*-value of 5*10^{-4} - 6 parameters significant at the 5% level

Again, this data was pure noise; there was absolutely no relationship in it. But from our data we find a highly significant regression, a respectable *R ^{2}* (which can be very high compared to those found in some fields like the social sciences) and 6 significant parameters!

This is quite a troubling result, and this procedure is not an uncommon one but clearly leads to incredibly misleading results. It shows how easily statistical processes can be heavily biased if care to accurately measure error is not taken.