Classical regression models are obtained by choosing a model that minimizes an empirical estimation of the Mean Square Error (MSE). Other quality measures are used, in general for robustness reasons. This is the case of the Huber loss

for instance. Another example of regression quality measure is given by the Mean Absolute Percentage Error (MAPE). If x denotes the vector of explanatory variables (the input to the regression model), y denotes the target variable and g is a regression model, the MAPE of g is obtained by averaging the ratio

The MAPE is often used in practice because of its very intuitive interpretation in terms of relative error. The use of the MAPE is relevant in finance, for instance, as gains and losses are often measured in relative values. It is also useful to calibrate prices of products, since customers are sometimes more sensitive to

In real world applications, the MAPE is frequently used when the quantity to predict is known to remain way above zero. It was used for instance as the quality measure in a electricity consumption forecasting contest organized by GdF ecometering on datascience.net. More generally, it has been argued that

where enough data are available, see e.g. [2].

We study in this paper the consequences of using the MAPE as the quality measure for regression models. Section 2 introduces our notations and the general context. It recalls the definition of the MAPE. Section 3 is dedicated to a first

optimal regression model with respect to the MSE is given by the regression function (i.e., the conditional expectation of the target variable knowing the explanatory variables). Section 3 shows that an optimal model can also be defined for the MAPE. Section 4 studies the consequences of replacing MSE/MAE by the

dimension. We show in particular that MAE based measures can be used to upper bound MAPE ones. Section 5 proves a universal consistency result for Empirical Risk Minimization applied to the MAPE, using results from Section 4. Finally, Section 6 shows how to perform MAPE regression in practice. It adapts

obtained model on simulated data.

We use in this paper a standard regression setting in which the data are fully described by a random pair Z = (X, Y ) with values in . We are interested in finding a good model for the pair, that is a (measurable) function g from ) is “close to” Y . In the classical regression setting, the closeness of g(X) to Y is measured via the risk, also called the mean

squared error (MSE), defined by

In this definition, the expectation is computed by respect to the random pair (X, Y ) and might be denoted to make this point explicit.

settings.

Let m denote the regression function of the problem, that is the function

from

It is well known (see e.g. [3]) that the regression function is the best model in the case of the mean squared error in the sense that ) minimizes the set of all measurable functions from

More generally, the quality of a model is measured via a loss function, l, from . The point-wise loss of the model g is l(g(X), Y ) and the risk of

the model is

leads to the risk defined above as

The optimal risk is the infimum of over measurable functions, that is

where ) denotes the set of measurable functions from to R. As

recalled above we have

As explained in the introduction, there are practical situations in which the risk is not a good way of measuring the closeness of g(X) to Y . We focus in this paper on the case of the mean absolute percentage error (MAPE) as an alternative to the MSE. Let us recall that the loss function associated to the

MAPE is given by

with the conventions that for all and that

MAPE-risk of model g is

Notice that according to Fubini’s theorem, implies in particular that and thus that interesting models belong to is the probability measure on

We will also use in this paper the mean absolute error (MAE). It is based on the absolute error loss,

risks, the MAE-risk is given by

A natural theoretical question associated to the MAPE is whether an optimal model exists. More precisely, is there a function such that for all models

Obviously, we have

A natural strategy to study the existence of is therefore to consider a point-wise approximation, i.e. to minimize the conditional expectation introduced above for each value of x. In other words, we want to solve, if possible, the

optimization problem

for all values of x.

introduce necessary and sufficient conditions for the problem to involve finite values, then we show that under those conditions, it has at least one global solution for each x and finally we introduce a simple rule to select one of the solutions.

3.1. Finite values for the point-wise problem60

To simplify the analysis, let us introduce a real valued random variable T

and study the optimization problem

Depending on the distribution of T and of the value of not always a finite value, excepted for m = 0. In this latter case, for any random variable T, J(0) = 1 using the above convention.

Let us consider an example demonstrating problems that might arise for be distributed according to the uniform distribution on [1].

Then

If 1], we have

This example shows that when T is likely to take values close to 0, then

when as a finite expectation, that is when the probability that |T| is smaller than decreases sufficiently quickly when goes to zero.

More formally, we have the following proposition.

if and only if

If any of those conditions is not fulfilled, then

Proof. We have

If P(T = 0) > 0 then for all . Let us therefore consider the case P(T = 0) = 0. We assume m > 0, the case m < 0 is completely identical.

We have

A simple upper bounding gives

and symmetrically

This shows that J(m) is the sum of finite terms and of Because of the symmetry of the problem, we can focus on . It is also obvious that is finite if and only if

more operational conditions in the rest of the proof.

Let us therefore introduce the following functions:

We have obviously for all ). In addition

According to the monotone convergence theorem,

)) are finite, or both are infinite. In addition, we have

therefore is finite if and only if )) is finite. So a sufficient and necessary condition for to be finite is

A symmetric derivation shows that is finite if and only if

The conditions of Proposition 1 can be used to characterize whether ]0]) decreases sufficiently quickly to ensure that J is not (almost) identically

equal to +. For instance, if

and the sum diverges, leading to (for

and thus the sum converges, leading to for all m (provided similar

3.2. Existence of a solution for the point-wise problem

If the conditions of Proposition 1 are not fulfilled, J(m) is infinite excepted in show that J(m) has at least one global minimum. This is done in the following

Proposition 2. Under the conditions of Proposition 1, J is convex and has at least one global minimum.

Proof. We first note that J is convex. Indeed for all is obviously convex. Then the linearity of the expectation allows to conclude

As 0 with either a > 0 or b < 0. Let us assume a > 0, the other case being symmetric. Then for

Then

and therefore lim

Similarly, if

and then

and therefore lim

Therefore, J is a coercive function and has at least a local minimum, which is global by convexity.

3.3. Choosing the minimum95

However, the minimum is not necessary unique, as J is not strictly convex. In general, the set of global minima will be a bounded interval of R. In this case, and by convention, we consider the mean value of the interval as the optimal solution.

As an example of such behavior, we can consider the case where T is a random variable on {1, 2, 3}, such that P(T = 1) = 0.3, P(T = 2) = 0.4 and

P(T = 3) = 0.3. Then the expected loss is

and the figure 1 illustrates that there is an infinity of solutions. Indeed when

More generally, for any random variable T, we have defined a unique value m, which is a global minimum of . Moving back to our problem, it ensures that the MAPE-regression function introduced in 8 is well defined and takes finite values on . As is point-wise optimal, it is

Figure 1: Counterexample with an infinite number of solutions.

One of the most standard learning strategy is the Empirical Risk Minimization (ERM) principle. We assume given a training set (which consists in n i.i.d. copies of the random pair Z = (X, Y ).

from . Given a loss function l, we denote

The empirical estimate of ) (called the empirical risk) is given by

Then the ERM principle consists in choosing in the class G the model that

minimizes the empirical risk, that is

The main theoretical question associated to the ERM principle is how to control ) in such a way that it converges to . An extension of this question

is whether can be reached if G is allowed to depend on n: the ERM is said to

for any distribution of (X, Y ) (see Section 5).

It is well known (see e.g. [3] chapter 9) that ERM consistency is related to uniform laws of large numbers (ULLN). In particular, we need to control

quantities of the following form

This can be done via covering numbers or via the Vapnik-Chervonenkis dimension (VC-dim) of certain classes of functions derived from G. One might think that general results about arbitrary loss functions can be used to handle the case

Lipschitz property of l (see Lemma 17.6 in [4], for instance) that is not fulfilled by the MAPE.

The objective of this section is to analyze the effects over covering numbers (Section 4.2) and VC-dimension (Section 4.3) of using the MAPE as the loss

those analyses (Section 4.4).

4.1. Classes of functions

Given a class of models, G, and a loss function l, we introduce derived classes,

H(G, l) given by

and

4.2. Covering numbers

4.2.1. Notations and definitions

Let F be a class of positive functions from an arbitrary set Z to . The

supremum norm on F is given by

We also define . We have obviously

only in ), hence the absolute value.

Let be a dissimilarity on F, that is a positive and symmetric function from to that measures how two functions from F are dissimilar (in particular can be used to characterize the complexity of F

Definition 1. Let F be a class of positive functions from Z to and a dissimilarity on a positive integer, a size respect to is a finite collection of elements of F such that for all

f

Then the -covering number of F is defined as follow.

Definition 2. Let F be a class of positive functions from Z to be a dissimilarity on F and 0. Then the -covering number of ), is the size of the smallest -cover of F. If such a cover does not exists, the

The behavior of ) with respect to characterizes the complexity of F as seen through . If the growth when 0 is slow enough (for an adapted choice of ), then some uniform law of large numbers applies (see Lemma 1).

4.2.2. Supremum covering numbers

Supremum covering numbers are based on the supremum norm, that is

For classical loss functions, the supremum norm is generally ill-defined on H(G, l). For instance let be two functions from ), generated by

and a value of ). Then sup

A similar situation arises for the MAPE. Indeed, let be two functions from , generated by

Thus unless G is very restricted there is always and has the general form 0 and thus lim

definition to a subset of Z. This corresponds in practice to support assumptions on the data (X, Y ). Hypotheses on G are also needed in general. In this latter case, one generally assumes . In the former case, assumptions depends on the nature of the loss function.

For instance in the case of the MSE, it is natural to assume that |Y | is upper

with probability one. If (

In the case of the MAPE, a natural hypothesis is that |Y | is lower bounded

by (almost surely). If (

and therefore the supremum norm is well defined.

The case of the MAE is slightly different. Indeed when x is fixed, then for sufficiently large positive values of Similarly, for sufficient large negative values of

. In addition, we have the following proposition.

Proposition 3. Let G be an arbitrary class of models with and let

denote the supremum norm on

Let

Proof. Let 0 and let be a minimal cover of ) (thus )). Let be the functions from G associated to and let be the corresponding functions in

Indeed let h be an arbitrary element of ) associated the corresponding function in ). Then for a given

We have then

For all

Then

and thus

which allows to conclude.

This Proposition shows that the covering numbers associated to a class of functions G under the MAPE are related to the covering numbers of the same class under the MAE, as long as Y stays away from too small values.

4.2.3. Lp covering numbers170

covering numbers are based on a data dependent norm. Based on the training set , we define for

We have a simple proposition:

Proposition 4. Let G be an arbitrary class of models and a data set such

that , then for all

Proof. The proof is similar to the one of Proposition 3.

This Proposition is the adaptation of Proposition 3 to covering numbers.

4.3. VC-dimension

dimension (VC dimension). We recall first the definition of the shattering coeffi-cients of a function class.

Definition 3. Let F be a class of functions from be a positive

integer. Let be a set of

that is the number of different binary vectors of size n that are generated by functions of F when they are applied to

Then the VC-dimension is defined as follows.

Definition 4. Let F be a class of functions from . The VC-dimension

of F is defined by

Interestingly, replacing the MAE by the MAPE does not increase the VC-dim of the relevant class of functions.

Proposition 5. Let G be an arbitrary class of models. We have

Proof. Let us consider a set of k points shattered by

function ) such that . Each corresponds to a

We define a new set of k points, () as follows. If

). For those points and for any

y|.

1 (or will never be possible). In addition when 0 and when ). Notice that min0 (as there is a finite number of binary vectors

and thus , that is . For such that and thus . Then

This shows that for each binary vector , there is a function

shattered by

Therefore we can take )) to get the conclusion.

If then k can be chosen arbitrarily large and

Using theorem 9.4 from [3], we can bound the covering number with a VC-dim based value. If 2, 1, and 0 ,

then

Therefore, in practice, both the covering numbers and the VC-dimension of MAPE based classes can be derived from the VC-dimension of MAE based classes.

4.4. Examples of Uniform Laws of Large Numbers

Rephrased with our notations, Lemme 9.1 from [3] is

Lemma 1 (Lemma 9.1 from [3])be a class of functions from

for all

A direct application of Lemma 1 to H(G, l) gives

provided the support of the supremum norm coincides with the support of (X, Y ) and functions in H(G, l) are bounded.

In order to fulfill this latter condition, we have to resort on the same strategy

As in Section 4.2.2 let and let be such that

almost surely, then

and

Then if ), Lemma 1 applies to (resp. to

Similar results can be obtained for the MAPE. Indeed let us assume that

0 almost surely. Then if

This discussion shows that , the lower bound on |Y |, plays a very similar

MAE and the MSE. A very similar analysis can be made when using the covering numbers, on the basis of Theorem 9.1 from [3]. It can also be combined with the results obtained on the VC-dimension. Rephrased with our notations,

Theorem 9.1 from [3] is

Theorem 1 (Theorem 9.1 from [3]). Let F be a class of functions from Z to

[0

As for Lemma 1, we bound via assumptions on G and on Y . For

instance for the MAE, we have

and for the MAPE

Equation (20) can be combined with results from Propositions 4 or 5 to allow a comparison between the MAE and the MAPE. For instance, using the VCdimension results, the right hand side of equation (19) is bounded above by

24

while the right hand side of equation (20) is bounded above by

In order to obtain almost sure uniform convergence of those right hand side quantities must be summable (this allows one to apply the Borel-Cantelli Lemma). For fixed values of the VC dimension, of this is always the case. If those quantities are allowed to depend on n, then

play symmetric roles for the MAE and the MAPE. Indeed for the MAE, a fast growth of with n might prevent the bounds to be summable. For instance, if grows faster than , then does not converges to zero and the series is not summable. Similarly, if converges too quickly to zero, for

summable. The following Section goes into more details about those conditions in the case of the MAPE.

We show in this section that one can build on the ERM principle a strongly

almost universal).

Theorem 2. Let Z = (X, Y ) be a random pair taking values in such that 0 almost surely (is a fixed real number). Let ((be a series of independent copies of Z.

such that:

1. G

2. is dense in the set of ) functions from to R for any

If in addition

and there is

then converges almost surely to

Proof. We use the standard decomposition between estimation error and approx-

imation error. More precisely, for , a class of functions,

We handle first the approximation error. As pointed out in Section 2, implies that ). Therefore we can assume there is a series (

of functions from ) such that

by definition of as an infimum.

Let us consider two models . For arbitrary x and y, we have

and thus

and therefore

and thus

This shows that lim

The estimation error is handled via the complexity control techniques studied

in the previous Section. Indeed, according to Theorem 1, we have (for p = 1)

with

Using the fact that log(

and

As lim

As lim

Therefore, for n sufficiently large, ) is dominated by a term of the form

with 0 and 0 (both depending on ). This allows to conclude that . Then the Borel-Cantelli theorem implies that

The final part of the estimation error is handled in a traditional way. Let

There is

Then . By definition

and thus for all g,

By taking the infimum on , we have therefore

Applying again the hypothesis,

and therefore

As a consequence

The combination of this result with the approximation result allows us to conclude.

Notice that several aspects of this proof are specific to the MAPE. This is the

This is also the case of the estimation part which uses results from Section 4 that are specific to the MAPE.

The previous Sections have been dedicated to the analysis of the theoretical

MAPE regression and we compare it to MSE/MAE regression.

On a practical point of view, building a MAPE regression model consists in minimizing the empirical estimate of the MAPE over a class of models

is to solve

where the (are the realizations of the random variables (

Optimization wise, this is simply a particular case of median regression (which is in turn a particular case of quantile regression). Indeed, the quotient

implementation that supports instance weights can be used to find the optimal model. This is for example the case of quantreg R package [5], among others. Notice that when corresponds to linear models, the optimization problem is a simple linear programming problem that can be solved by e.g. interior point

For some complex models, instance weighting is not immediate. As an example of MAPE-ing a classical model we show in this section how to turn kernel quantile regression into kernel MAPE regression. Notice that kernel regression introduces regularization and thus is not a direct form of ERM.

6.1. From quantile regression to MAPE regression

6.1.1. Quantile regression

Let us assume given a Reproducing Kernel Hilbert Space (RKHS), H, of functions from (notice that could be replaced by an arbitrary space

and . As always, we have

The standard way of building regression models based on a RKHS consists in optimizing a regularized version of an empirical loss, i.e., in solving an

optimization problem of the form

Notice that the reproducing property of H implies that there is

f(x

In particular, quantile regression can be kernelized via an appropriate choice for l. Indeed, let [0; 1] and let , introduced in [7]:

The check-function is also called the pinball loss. Then, the kernel quantile optimization problem, treated in [8, 9], is defined by:

where 0 handles the trade-off between the data fitting term and the regular-

6.2. MAPE primal problem

To consider the case of the MAPE, one can change the equation (24) to (25):

Notice that for the sake of generality, we do not specify the value of in this derivation: thus equation (25) can be seen as a form of “relative quantile”.

standard MAPE, that is to . The practical relevance of the “relative quantile” remains to be assessed.

Using the standard way of handling absolute values and using f(x) = , we can rewrite the regularization problem (25) as a (primal) op-

where

6.2.1. MAPE dual problem

Let us denote ) the vector regrouping all the variables of the primal problem. We denote in addition:

max

where the are the Lagrange multipliers. Some algebraic manipulations show that problem (27) is equivalent to problem (28):

We can simplify the problem by introducing a new parametrisation via the variables . Then the value of w is obtained from constraint (29) as ). Constraints (30) can be rewritten into 1

those equations into account, the objective function becomes

Using constraints (31) and (32), the last two terms simplify as follows:

and thus the objective function is given by

where ) is the kernel matrix. This shows that the objective function can be rewritten so as to depend only on the new variables . The last step of

The cases of constraints (29) and (30) have already been handled.

Notice that given an arbitrary , there is always that . However, constraints (31) and (32) combined with and 0 show that and (and thus ) cannot be arbitrary, as we

. Conversely, it is easy to see that if satisfies the constraints , then there is for k = 1, . . . , 4 such that and such that the constraints (31), (32) and (33) are satisfied (take

6.2.2. Comparaison to the quantile regression

In the case of quantile regression, [8] shows that the dual problem is equivalent

to

In comparison to problem (34), one can remark that the modification of the

primal optimization problem is equivalent to changing the set of optimization in the dual optimization problem. More precisely, it is equivalent to reducing (resp. increasing) the “size” of the optimization set of

Thus, the smaller is , the larger is the optimization set of . This permits

error is potentially bigger). Moreover, by choosing a very large value of C (or ), one can ensure the same optimal value of each in MAE and MAPE dual problems. This surprising fact can be explained by noticing that a very large value of C corresponds to a very small value of

potential overfitting. When this overfitting appears, regardless of the loss function and thus the different loss functions are equivalent.

6.3. A simulation study

6.3.1. Generation of observations

described in section 6.1 on simulated data, and we compare the results to the ones obtained by kernel median regression. Experiments have been realized using a Gaussian kernel.

As in [8], we have simulated data according to the sinus cardinal function,

defined by

However, to illustrate the variation of the prediction according the proximity to zero, we add a parameter a and we define the translated sinus cardinal function by:

For experiments, we have generated 1000 points to constitute a training set, and 1000 other points to constitute a test set. As in [8], the generation process is the following:

with

estimation, we have computed for several values of a. The value of the regularization parameter C is chosen via a 5-fold cross-validation.

6.3.2. Results

Table 1: Summary of the experimental results: for each value of the translation parameter a, the table gives the MAPE of estimated on the test set. The table also reports the value of the regularization parameter C for both loss function.

Results of experiments are described in the table 1. As expected, in most

especially the case when values of y are close to zero.

6.3.3. Graphical illustration

Some graphical representations of are given on Figure 2. This Figure illustrates several interesting points:

• Up to translation, looks roughly the same for each a, whereas

• Red curves are closer to 0 than blue curves. One can actually show that,

• The red curve seems to converge toward the blue one for high values of a.

We have shown that learning under the Mean Absolute Percentage Error

particularly, we have shown the existence of an optimal model regarding to the MAPE and the consistency of the Empirical Risk Minimization. Experimental results on simulated data illustrate the efficiency of our approach to minimize the MAPE through kernel regressions, what also ensures its efficiency in application

is positive by design and remains quite far away from zero, e.g. in price prediction for expensive goods). Two open theoretical questions can be formulated from this work. A first question is whether the lower bound hypothesis on |Y | can be lifted: in the case of MSE based regression, the upper bound hypothesis on

cannot be adapted immediately to the MAPE because of the importance of the lower bound on |Y | in the approximation part of Theorem 2. A second question is whether the case of empirical regularized risk minimization can be shown to be consistent in the case of the MAPE.

Figure 2: Representation of estimation: in blue and

The authors thank the anonymous reviewers for their valuable comments that helped improving this paper.

Mathematical Statistics 35 (1) (1964) 73–101.

[2] J. S. Armstrong, F. Collopy, Error measures for generalizing about forecasting

(1992) 69 – 80.

Nonparametric Regression, Springer, New York, 2002.

[4] M. Anthony, P. L. Bartlett, Neural Network Learning: Theoretical Founda-

tions, Cambridge University Press, 1999.

[5] R. Koenker, quantreg: Quantile regression. r package version 5.05 (2013).

2004.

[7] R. Koenker, G. Bassett Jr, Regression quantiles, Econometrica: journal of

the Econometric Society (1978) 33–50.

[8] I. Takeuchi, Q. V. Le, T. D. Sears, A. J. Smola, Nonparametric quantile

[9] Y. Li, Y. Liu, J. Zhu, Quantile regression in reproducing kernel hilbert spaces,

Journal of the American Statistical Association 102 (477) (2007) 255–268.