Matching Curve Selection to Prediction Loss

Why the Error Measure Matters

Two nonlinear curves can make different kinds of mistakes. One may be close to the observed outcomes for most cases but miss a few by a wide margin. Another may make moderate errors consistently. Which curve is preferable depends not only on how well each fits the data, but on what those prediction mistakes mean in the application.

For example, when forecasting a service time, a large overprediction may waste staff capacity, while a large underprediction may leave people waiting without help. Treating the two errors as equally costly may not suit that task. A curve-selection decision is therefore incomplete until the prediction loss—the numerical penalty assigned to prediction errors—has been chosen and explained.

This tutorial focuses on matching that loss to the prediction task. It does not make one loss universally best. Instead, it shows how to compare candidate curves under a clearly stated measure and how to explain the limits of the resulting choice.

Prediction Errors and Loss

Suppose candidate curve \(m\) gives prediction \(\hat{y}_{mi}\) for case \(i\), whose observed outcome is \(y_i\). Define its prediction error as observed minus predicted:

With this convention, \(e_{mi}>0\) means the curve underpredicts, and \(e_{mi}<0\) means it overpredicts. A loss function \(L(e)\) assigns a nonnegative penalty to an error. A common way to summarize performance over \(n\) validation cases is the mean loss:

A smaller mean loss indicates better performance under that particular loss function and evaluation set. For the comparison to be meaningful, every candidate should be evaluated on the same cases, using the same recorded outcome and loss definition. The validation outcomes should not have been used to fit or tune the candidate curves being compared. The cases should also represent, or be appropriately adjusted to reflect, the situations in which predictions will be used.

The standard summaries below require no normality assumption simply to calculate them. Their practical interpretation does depend on using comparable cases and a sensible outcome scale. As with any out-of-sample comparison, a low score on one validation set does not guarantee the same ranking on every future set.

Absolute and Squared Error

Two widely used loss functions are absolute error and squared error. Their averages are called mean absolute error and mean squared error. Root mean squared error is the square root of mean squared error.

MAE reports a typical absolute error in the same units as the outcome. MSE is in squared outcome units, so it is less directly interpretable in context. RMSE returns to the outcome's units. For a fixed set of predictions, RMSE and MSE always rank candidate curves the same way: taking a square root preserves the ordering of nonnegative numbers.

Squared error gives relatively more influence to a large error than absolute error does. For two positive error magnitudes \(a>b\), their squared-loss ratio is \(a^2/b^2=(a/b)^2\), rather than the absolute-loss ratio \(a/b\). This makes a few large misses more influential in the squared-error average. It does not mean that squared error is always the right choice: that influence is desirable only when large errors should carry especially strong penalties.

A useful algebraic identity is \(a^2-b^2=(a-b)(a+b)\). For positive numerical magnitudes, this difference is greater than \(a-b\) only when \(a+b>1\); it is not a general inequality. Moreover, squared error and absolute error have different units, so directly comparing their numerical values is not a meaningful way to choose between them. The relevant point is the way each loss responds to error magnitudes, not a universal numerical comparison between the two scales.

When the Direction of Error Matters

Absolute and squared loss treat overprediction and underprediction symmetrically: errors of equal magnitude receive equal penalties. In some applications, the consequences are asymmetric. A staffing forecast that is too low might leave a service understaffed, while a forecast that is too high might mainly create modest idle capacity.

One simple asymmetric loss assigns a cost per unit of underprediction and a different cost per unit of overprediction. With \(e=y-\hat{y}\), positive errors are underpredictions and negative errors are overpredictions. If the underprediction cost is \(c_u\) per outcome unit and the overprediction cost is \(c_o\), define:

Both cost rates are nonnegative. When \(e>0\), this gives \(L(e)=c_u e\); when \(e<0\), it gives \(L(e)=c_o|e|\). The values of \(c_u\) and \(c_o\) should reflect the task and be stated explicitly. If they are merely chosen after seeing which curve wins, the comparison can be misleading.

Weighting Cases to Match Intended Use

A validation set may not contain cases in the proportions expected during actual use. For instance, a balanced set of routine and severe cases may be useful for studying both types, even though severe cases are rare in deployment. An unweighted average then gives the two types equal influence, which may not represent the average loss expected in use.

If group \(g\) makes up a proportion \(w_g\) of intended predictions, with \(w_g\geq0\) and \(\sum_g w_g=1\), one can average the mean loss within each group using those proportions:

Here \(\overline{L}_{mg}\) is candidate \(m\)'s mean loss among validation cases in group \(g\). The group proportions must come from a defensible description of intended use, not from whichever weights produce a preferred result. This calculation assumes the validation cases give relevant information about candidate performance within each group and that the chosen deployment proportions are appropriate.

A Practical Selection Process

A defensible comparison connects a clearly defined prediction task to a prespecified loss, then applies that same rule to every candidate curve. The following sequence makes the reasoning explicit.

The loss can be selected using subject knowledge, operational priorities, or explicit costs. For example, MAE may be suitable when a unit of error has roughly proportional consequences and extreme errors should not dominate. Squared error may be appropriate when large errors are disproportionately damaging. An asymmetric loss is useful when error direction changes the consequence. Weighted loss can represent a known target mixture of cases. These are choices to justify, not universal rules.

Common Mistakes and Limits

• Choosing the loss because it makes a preferred curve win. This reverses the logic. Define the prediction consequences and loss first; then compare the candidate curves.
• Calling the smallest score “the best curve” without qualification. Say which loss, validation cases, and target population support the choice. A curve can rank differently under another reasonable loss.
• Ignoring units. MAE and RMSE use the outcome's units, while MSE uses squared units. Values from different loss scales should not be treated as directly comparable.
• Letting an average hide a critical error pattern. A low overall loss can coexist with poor predictions for a rare but important group. Report relevant group performance or encode the cost explicitly.
• Treating a lower validation loss as proof of a true relationship. It is evidence about predictive performance under the evaluation setup, not proof that a curve is scientifically correct or will always predict better.

For example, it would be incorrect to conclude that curve A is simply superior because its MAE is smaller, while ignoring that the task assigns much higher cost to underprediction. The correction is to evaluate the curves with the loss that represents the stated task and to qualify the conclusion accordingly.