







































DYNAMIC ECONOMETRIC MODELS


© 2016 Nicolaus Copernicus University. All rights reserved.  

http://www.dem.umk.pl/dem 

D Y N A M I C  E C O N O M E T R I C  M O D E L S  
DOI: http://dx.doi.org/10.12775/DEM.2016.001  Vol. 16 (2016) 5−20 

Submitted Juny 15, 2016  ISSN (online) 2450-7067 

Accepted July 15, 2016 ISSN (print) 1234-3862 

Joanna Bruzda
*
 

Quantile Forecasting in Operational Planning  
and Inventory Management – an Initial  

Empirical Verification 

A b s t r a c t. In the paper we present our initial results of an empirical verification of differ-

ent methodologies of quantile forecasting used in operational management to calculate the re-

order point or order-up-to level as well as the optimal order quantity according to the 

newsvendor model. The comparison encompasses 26 procedures including quantile regres-

sion, the basic bootstrap method and popular textbook formulas. Our results, obtained on the 

base of 30 time series concerning such diversified phenomena as supermarket sales, passenger 

transport and water and gas demand, point to the usefulness of regression medians, regression 

quantiles, bootstrap methods and the procedures available in the SAP ERP system. 

K e y w o r d s: LINLIN loss, quantile forecasting, quantile regression, re-order point, theta 

method.  

J E L Classification: C21; C44; C53. 

Introduction  

 A quantile forecast of a variable tY  is the conditional quantile of tY  

given the information available till time 1t , i.e., the quantile of order : 

                                                 
* Correspondence to: Joanna Bruzda, Nicolaus Copernicus University, Faculty of Eco-

nomic Sciences and Management, Gagarina 13A, 87-100 Toruń, Poland, e-mail: Joan-

na.Bruzda@umk.pl. 
 This work was financed by the Polish National Science Center with a grant obtained 

under decision no. DEC-2013/09/B/HS4/02716. The paper is a continuation of the research 

presented in Bruzda (2016), based on a different set of forecasting procedures. 



Joanna Bruzda  

DYNAMIC ECONOMETRIC MODELS 16 (2016) 5–20 

6 

})(;inf{   yFyy
  (1) 

 

of )(F  being the conditional distribution of tY  given the information set 

1t . The quantity (1) is a solution to the following optimization problem: 

 1|,min t
x

YxLLE , (2) 

where 










yxxy

xyxy
yxLL

dla)1(

dla
),(




              (3) 

is the double linear loss function, often denoted as LINLIN (see Granger, 

1999; Gneiting, 2011; Bruzda, 2014).  

 The loss (cost) function (3) turns out to be very popular in operational 

management applications. For example, in the so-called Newsvendor prob-

lem (the one-period inventory model) the optimal order or production quan-

tity in the next period is given as a quantile forecast of the next period de-

mand D, i.e., assuming D has a continuous distribution, as: 













 

ou

u

D
cc

c
FQ 1 ,  (4) 

( uc  and oc  are the unit costs of under- and overstocking, respectively). In 

a similar way, setting the next period re-order point (ROP) or order-up-to 

level  S for a service level  the following standard textbook formulas are 

used: 

,1)1( DLzLdROP    (5a) 

DTLzTLdS   )(  (5b) 

( d  denotes the mean demand, L is the lead time, T stands for the inventory 

cycle period, D  denotes the standard deviation of D and z  is the  

-quantile of the standard normal distribution). If the demand is forecasted, 

quantiles of a predictive distribution (a conditional distribution with estimat-

ed parameter values) will be used to compute ROP and S, respectively.  

 Further in the paper we discuss in some detail the premises of quantile 

forecasting in operational planning and inventory management, and under-



Quantile Forecasting in Operational Planning and Inventory Management… 

DYNAMIC ECONOMETRIC MODELS 16 (2016) 5–20 

7 

take the task of an empirical verification of different statistical procedures 

leading to quantile predictions. 

1. Quantile Forecasting in Operational Planning and Inventory 
Management 

 There are numerous procedures of computing quantile forecasts, encom-

passing parametric, semiparametric and nonparametric methods, approaches 

utilizing ex ante and ex post forecast errors, and non-simulation- and simula-

tion-based (bootstrap and parametric Monte Carlo) procedures. One can use 

the many existing methods of forecasting the Value at Risk in finance (see, 

e.g., Doman and Doman, 2009) or computing interval forecasts (and fan 

charts) of macroeconomic variables (see, e.g., Chapter IV in Clements, 

2005) as well as procedures and formulas tailored to applications in the area 

of supply chain management (see, e.g., Wagner, 2010; Ciesielski, 2011, and 

SAP Help Portal). A description of chosen methods of quantile forecasting 

with some indications on their potential use in logistics together with a simu-

lation analysis can be found in Bruzda (2014). 

 When considering quantile predictions in operational management, one 

should keep in mind several of its premises. First, to assess the quality of 

quantile predictions in this area, one will normally rely on different measures 

than those utilized to examine interval forecasts, in which case the coverage 

is evaluated, as well as those used to assess forecasts of Value at Risk, in 

which case the most important measures are usually the hit ratio and loss 

functions based exclusively on positive (or, alternatively, negative) forecast 

errors. Table 1 below presents the most important measures of predictive 

accuracy in the case of unbiased forecasts, which are the solution to the min-

imization of the mean squared error (MSE) loss, accompanied by the corre-

sponding measures which should be used in quantile forecasting. 

Table 1.  Basic accuracy measures for unbiased and quantile forecasts 

Unbiased forecasts Quantile forecasts 

 TeME t  TyyIHit tpt /)(   

 TeRMSE t
2    TeeLL tt /)1(     

Note: ty
 – the realized value of tY

; 
tpy

 – the forecasted value of tY
; )(I  – the indicator function 

taking on the value of 1, if the condition in brackets is fulfilled, and 0 otherwise; },0max{ tt ee 
 (posi-

tive forecast errors), },0max{ tt ee 
 (negative forecast errors), where tptt yye  . 



Joanna Bruzda  

DYNAMIC ECONOMETRIC MODELS 16 (2016) 5–20 

8 

The measures in the first row of Table 1 inform whether the evaluated fore-

casts can be considered as unbiased or quantile forecasts, respectively, while 

in the second row there are given mean values of the cost functions for these 

two sorts of predictions. (In the case of logistic forecasting, the quantile cost 

function will often be proportional to the real logistic cost – see, for exam-

ple, the textbook derivation of the formula (4).) In the present paper, the 

LINLIN cost function (3) is considered to be the most important measure of 

forecast accuracy for quantile forecasting in operational management
1
. 

 Among other characteristics of quantile predictions in operational plan-

ning are the following: 

 computation of quantile forecasts for a range of quantiles instead of just 

extreme quantiles, 

 simultaneous forecasting of a large number of univariate time series and, 

due to this, the focus on simplified methods 

 forecasting based on relatively short time series 

 the need to simultaneously model the conditional mean and conditional 

variance of time series 

 the necessity to compute long-term quantile forecasts or, alternatively, 

quantile forecasts for different sampling rates. 

The last observation results, among others, from the practice of freezing the 

master production schedule, the need of long-term optimization of distribu-

tion networks, capacity planning or long lead times. 

2. Empirical Verification of Quantile Forecasting Methodologies 

 The dataset used in the study comprises 30 time series from certain pub-

licly available databases. To simplify matters we concentrate on stationary 

and trend-stationary series (according to the ADF tests with lag lengths cho-

sen with the AIC) and also assume that an examination of short time series 

may constitute an interesting addendum to what is already known about the 

relative performance of different forecasting methodologies. The series are 

presented in Figure 1, whereas below we provide a short description of this 

dataset
2
: 

                                                 
1 Alternative measures of forecasting performance in the inventory management context 

are those based on the concept of expected shortage. However, as the target here is on 

quantile forecasting, we concentrate exclusively on the LINLIN loss, which is the loss mini-

mized with conditional quantiles – see Gneiting (2011). 
2 For further information see http://www.forecastingprinciples.com/index.php/data and 

https://cran.r-project.org/web/packages/bayesm/index.html. 



Quantile Forecasting in Operational Planning and Inventory Management… 

DYNAMIC ECONOMETRIC MODELS 16 (2016) 5–20 

9 

 x1 – x14 – time series from the M3 forecast competition (category ‘other 

series’) ‘gas sent’ (6 series) and ‘water demand’ (8 series), 

 x15 – x23 – weekly data on slice cheese sales in a supermarket chain in 

the US from the R toolbox ‘bayesm’ (the longest point-of-sale data of 

length 68 without visible promotional effects were chosen), 

 x24 – x30 – time series from the T-competition: waste water pipe data 

(4 weekly time series), weekly train passengers in Switzerland (1 series) 

and 2 monthly series: aircraft capacity utilization in the US and 

transborders-rail US-Mex. 

 All the series have been shortened to the length of the shortest time se-

ries of length 68. Then observations from 1 to 64 were used to estimate 

models utilized to compute one-step ahead quantile forecast in the first step. 

In the next steps, the samples were lengthen recursively by one observation 

and all models were re-estimated and used to compute further one-step ahead 

quantile predictions. Continuing in this way, 4 forecast were obtained for 

each time series and each examined quantile
3
. 

 The following methods were compared: 

 M1 – a procedure refering to the textbook formulas (5a)–(5b), based on 

a two-step approach, i.e., a quantile forecast is given as eTp Szy   , 

where Tpy  is a forecast of the conditional mean and eS  is the in-sample 

standard error (the equation for the conditional mean is estimated by 

ols), 

 M2 – as M1 but in the second step the mean absolute deviation (MAD) 

is used, i.e., the quantile forecast is computed as MADzyTp  

2

, 

 M3 – as M1 but with a correction of the conditional mean computed 

nonparametrically, i.e., the forecast is given as eyTp  , where e  de-

notes an estimate of the -quantile of residuals, which we compute ac-

cording to the method implemented in Matlab ver. R2015a, 

 M4 – a basic SAP ERP approach, i.e., a quantile forecast is given as 

MADzyTp  

2

 but the equation for Tpy  is estimated by median 

regression (the least absolute deviation – lad – method), 

 M5 – median regression combined with the nonparametric estimation of 

quantiles as in M3, 

                                                 
3 Considering a larger number of forecasts (and shorter estimation samples) resulted in 

a poor behavior of the ml estimates of our models. 



Joanna Bruzda  

DYNAMIC ECONOMETRIC MODELS 16 (2016) 5–20 

10 

 M6 – (partially iterated) weighted ols, in which an auxiliary regression 

for logarithms of squared ols residuals regressed on the same variables as 

in the equation for the mean is estimated, which gives weights for an ef-

ficient estimation of the equation for the mean, and, in order to forecast 

the variance, the auxiliary equation is re-estimated on residuals obtained 

through the efficient estimation
4
, 

 M7 – ml estimation applied to the model specified in M6 with starting 

values being the (appropriately modified) estimates obtained in M6, 

 M8 – quantile regression (see Koenker, 2005). 

In the procedures M1–M8 the equation for the mean (median, quantile) is an 

autoregression of order p ( 50  p ) with a polynomial trend of order s  

( 20  s ), where the parameters p and s are set automatically with the 

AICC criterion in its versions for the ols, lad and quantile estimation, respec-

tively.  

 M9 – the theta method – a modification of the exponential smoothing 

which was very successful in the M3-forecast competition (see 

Makridakis and Hibon, 2000), with parameters estimated here with lad 

under the initial state set to the first observation in the sample and the 

starting values: )1,0(~U  and 0drift ; 20 randomizations for  were 

assumed; the quantile forecast is then computed as in M4, 

 M10 – the theta model estimated as in M9 combined with smoothing 

MAD as in the SAP ERP system under the default value of the smooth-

ing constant , i.e.,  – the parameter standing by the absolute value of 

the last forecast error – is set to 0.3, 

 M11 – the theta method with smoothing MAD and all parameters (, 

and the drift terms) estimated by quantile regression; the quantile fore-

casts are computed via the following equation
5
: 

  

 ,)ˆ1(ˆˆˆ)ˆ1(ˆˆ
1102110   tttttp MADuzyyy  



 
with

  
                                                 

4 Such an approach produced better forecasts on average (in terms of ratios of the relative 

LINLIN loss) than a procedure based on weighted ols without the re-estimation of the auxilia-

ry equation. In both of these procedures, to measure volatility the standard deviation is used, 

thus the formula for quantile forecasting given in M1 is applied, but this time the standard 

deviation is forecasted.  
5 This specification provided lower values of the relative LINLIN loss as compared with 

an equation without the parameter 0 (except for  = 0.25 and both modelling levels and mod-

elling logarithms) 



Quantile Forecasting in Operational Planning and Inventory Management… 

DYNAMIC ECONOMETRIC MODELS 16 (2016) 5–20 

11 

  ;0,ˆ,ˆ

,)ˆ1(ˆˆ

,ˆ)ˆ1(ˆˆˆ

111

110

110











MADyyyyu

MADuMAD

yyy

ttt

ttt

ttt





 

the starting values of the smoothing constants were from the )1,0(U  dis-

tribution (20 randomizations were assumed), while the starting values of 

the drift terms were set to 0, 

 M12–M22 – these are M1–M11, respectively, but applied to logarithms 

of the data, 

 M23–M26 – the basic bootstrap method as presented by Clements 

(2005), §4.2.3, based on his formula (4.9) and the percentile method of 

Efron (see the formula (4.4) in Clements, 2005), without bias correction; 

M23 and M24 are performed on levels, while M25 and M26 – on loga-

rithms of the data; in order to better account for estimation errors, in 

M23 and M25 the residuals from the ols estimation of autoregressive 

models with trend functions (as chosen in M1) are resampled according 

to the overlapping block method with block length set to 10, whereas in 

M24 and M26 a simple resampling is used, i.e., the block length is set to 

1; the number of bootstrap replication is set to 250
6
. 

 All computations were performer in Matlab ver. 2015a endowed with the 

newest versions of the optimization and statistics toolboxes, thus they relied 

on many procedures implemented in these packages, including the nonpara-

metric quantile estimation by the function ‘quantile’. It is worth noticing that 

among the different forecasting procedures are 4 methods assuming the use 

of the LINLIN loss function for estimation purposes, i.e., based on quantile 

regression (M8, M11, M19, M22). In the case of the linear and linearized 

models (M8, M19) the estimation relied on linear programming methods (the 

active-set algorithm was used), while on the other two models nonlinear 

                                                 
6 The procedures M23 and M25 are based on a mixed approach in which an equation for 

the conditional mean is explicitly specified, whereas the heteroscedasticity of error terms is 

not modeled. (It is merely accounted for when assessing the variance of estimators in the 

equation for the mean.) This renders forecasting of the residual variance impossible. We 

include these procedures here because, somewhat surprisingly, they appear to often provide 

better forecasts than the other approaches (see Table 2), especially when forecasting quantiles 

near or below the median. This underlines the importance of accounting for estimation errors 

when computing quantile predictions and leaves space for possible improvements through the 

use of model-based bootstrap procedures, i.e., an explicit modeling of the conditional volatili-

ty and possibly also a better specification of the conditional mean through the inclusion of 

nonlinearity and/or longer lag lengths. (Although it is worth adding that our experimentation 

with longer lag lengths, in mean terms, did not improve the results presented here.)  



Joanna Bruzda  

DYNAMIC ECONOMETRIC MODELS 16 (2016) 5–20 

12 

optimization with the function ‘fmincon’ was performed (the interior-point 

algorithm was applied to the nonlinear objective functions). Similar optimi-

zation techniques were also utilized in the case of the linear and nonlinear 

median regression. Furthermore, it is also worth noticing that the bootstrap 

methods M23–M26 (similar to our experiments with the approximation of 

quantiles of predictive distributions through the formula for interval forecast-

ing in classical regression, which provided slightly less satisfactory results 

for our dataset than the method M1 and are not presented here) are an at-

tempt to explicitly account for estimation errors associated with equations 

for the conditional mean in the computation of quantile forecasts.  

 The following quantiles were examined: 0.05; 0.1; 0.25; 0.5; 0.75; 0.9; 

0.95. The evaluation of our forecast procedures was based on mean values of 

the LINLIN loss for each quantile and each time series. Two rankings of the 

methods were constructed – one for the whole set of quantiles and second for 

the two largest quantiles (see Table 2), showing how often a given procedure 

produced the best outcomes. Besides, we also computed aggregate measures 

of the form: 

 
,

)1(11

1 1

 
 

 


m

i

T

t itp

itit

y

ee

Tm
LLrelativemean



 
  (6) 

where m is the number of series (m = 30), T is the number of one-step ahead 

quantile forecasts (T = 4) and 

itpy  is the tth -quantile forecast for time se-

ries i. The measures are also accompanied by the mean values of the hit rati-

os. In accordance with the way of reporting forecasting results for a larger 

dataset, the aggregate measures were also presented in their trimmed ver-

sions based on datasets without 10% of time series with the smallest and 

10% of time series with the largest values of the corresponding mean relative 

LINLIN loss or the mean values of Hit. All the computations are collated in 

Tables 3–4.  

 The conclusions from our study, based on both the presented results and 

an analysis of more detailed statistics, are the following. First, the models we 

use to describe heteroscedasticity suit well the purpose of quantile forecast-

ing for quantiles below the median, as can be seen in Table 3, where we find 

that the ml estimation for variables in levels (the procedure M7) produces 

the best results in terms of the mean values of the relative LL, while the non-

linear quantile regression (M11 and M22) leads to the lowest values of the 

trimmed relative LL. It is worth noting that, according to the rankings pre-

sented in Table 2, the methods based on linear and nonlinear quantile regres-

sion (M8, M11, M19, M22) provide the best quantile forecasts in about 27% 



Quantile Forecasting in Operational Planning and Inventory Management… 

DYNAMIC ECONOMETRIC MODELS 16 (2016) 5–20 

13 

of the examined cases and are ranked high, especially in the ranking for all 

quantiles. Interestingly, in the first ranking the nonlinear quantile models 

have the highest ranks, whereas in the second the linear quantile regression 

is ranked higher than the methods M11 and M22. In fact, even 16 cases out 

of the total of 23 when M8 or M19 lead to the best quantile predictions 

(comp. Table 2) concern quantiles 0.75, 0.9 and 0.95. In mean terms, how-

ever, linear regression quantiles are not that attractive, since for quantiles 0.9 

and 0.95 they are clearly outperformed by median regression with parametric 

or nonparametric correction of the median, as well as ols with a parametric 

computation of quantiles.  

 Generally, it can be noticed that, for quantiles near the median, the theta 

method often produces the best outcomes in terms of the relative LINLIN 

loss. This seems to support the findings from the M3-forecast competition, 

where mean values were exclusively forecasted (see Makridakis, Hibon, 

2000). On the other hand, for the largest two quantiles, the lowest values of 

the mean relative LL are obtained through the lad estimation of simple auto-

regressive models with trend functions (see Table 3). However, depending 

on the quantile, either parametric (M4, M15) or nonparametric (M5, M16) 

estimators in the second step lead to the best mean outcomes. Nevertheless, 

the measures collated in the left-hand side of Table 3 point to the SAP meth-

od applied to logarithms of the data (M15) as that producing the best 

quantile predictions for  = 0.9 and 0.95. 

 Interestingly, the rankings presented in Table 2 often point to different 

methods than the mean values of the relative LINLIN loss as the most valua-

ble solutions to the task of quantile forecasting. For example, they single out 

the bootstrap methods, especially the procedures M25 and M23. Analyzing 

more detailed statistics shows, however, that the excellent performance of 

the bootstrap in its different forms (the methods M23–M26) usually takes 

place for quantiles not very far from the median. This suggests that it may be 

worth designing a bootstrap procedure which will make it possible to also 

forecast the variance or, alternatively, one can be interested in performing 

bootstrapping based on median regression (taking into consideration its ex-

cellent performance at forecasting high quantiles). Furthermore, it also calls 

for the inclusion of bootstrap methods in any application of the so-called 

focus forecasting – forecasting with methods producing the lowest values of 

chosen loss functions based on ex post forecast errors. Among other proce-

dures that certainly should also be consider in focus forecasting are those 

based on lad estimation (such as M4 or M5) as well as linear and nonlinear 

quantile regression. 



Joanna Bruzda  

DYNAMIC ECONOMETRIC MODELS 16 (2016) 5–20 

14 

 The most striking finding from our study is probably the performance of 

the textbook formula M1 in relation to the other methods examined here as 

far as the largest quantile 0.95 is concerned. It turns out that M1 produces 

then the best results in terms of the trimmed mean relative LL, while at the 

same time it is ranked among the worst according to the number of the best 

outcomes produced for different time series. We treat this finding as an indi-

cation that this formula may often constitute a ‘safe choice’, which should 

not lead to highly suboptimal results, although usually it will be outper-

formed by other methods used in focus forecasting. 

 Analyzing the mean values of Hit presented in Table 4 it can be noticed 

that the bootstrap and the smoothing methods perform relatively well in 

terms of this measure. In particular, the smallest discrepancies between Hit 

and 1  for quantiles 0.25 and 0.5 are for the smoothing procedures. We 

had to resigned from testing the significance of deviations of the hit ratios 

for individual time series due to the fact that we operate on short time series 

and are able to analyze only a small number of forecasts. It can be noticed, 

however, that the Kupiec test performed on the whole set of the analyzed 

data does not reject its null hypothesis (stating that, for a given method, Hit 

is equal to 1 ) for all the examined quantiles above the median (see Ta-

ble 4). 

 Finally, it can also be observed that, although taking logarithms prior to 

the estimation of our models can improve the forecasting results, it appears 

that the logarithmic transformation is not always necessary and also not al-

ways enough to remove heteroscedasticity from the analyzed time series. 

Final Remarks 

 As a summary of our findings from the present examination it is worth 

underlining that among the most promising methods of quantile forecasting 

in operational planning and inventory management are procedures based on 

median and quantile estimation as well as bootstrap techniques. In particular, 

it seems interesting to join the theta method and the quantile regression 

methodology since, as it takes place in this study for quantiles 0.05, 0.1, 0.25 

and 0.5, such an approach may be competitive to other popular methods of 

quantile forecasting.
7
 

                                                 
7 However, in order to properly assess the performance of this approach in relation to oth-

er methods, further studies are required, possibly based on a larger dataset, since in the pre-

sent study we do not formally test the significance of the differences in predictive abilities of 

the different approaches.  



Quantile Forecasting in Operational Planning and Inventory Management… 

DYNAMIC ECONOMETRIC MODELS 16 (2016) 5–20 

15 

 The above remark stays in accordance with the opinion that regression 

quantiles should be included in the inventory management toolbox
8
. As was 

shown in the present study, they can be used both to optimize the value of 

the smoothing constant  in the MAD equation as well as to compute 

quantile forecasts based on autoregressive specifications with trend func-

tions. 

 Some promising areas of further studies are the design of bootstrap pro-

cedures based on an explicit modelling of volatility with the aim of a use in 

inventory planning and an assessment of simulation- and non-simulation-

based procedures of longer-term quantile forecasting in operational man-

agement applications. Some of these problems will be discussed further in 

the present project. 

References  

Applied Forecasting, webpage sponsored by Route Optimization and Heiloo, 

http://www.appliedforecasting.com/quantiles-forecasts-for-inventory-optimization/ 

(26.09.2015). 

Bruzda, J. (2014), Prognozy kwantylowe w zastosowaniach logistycznych. Wprowadzenie do 

problematyki (Quantile Forecasts in Logistic Applications – An Introduction), in 

Chaberek M and Reszka L. (eds.) Zeszyty Naukowe Uniwersytetu Gdańskiego, 

Ekonomika Transportu i Logistyki (Scientific Papers of the University of Gdańsk, 

Transport and Logistics Economics), 51, 175–195. 

Bruzda, J. (2016), Metody wyznaczania prognoz kwantylowych w logistyce – weryfikacja 

empiryczna (Methods of Quantile Forecasting in Logistic Applications – an Empirical 

Verification), Logistyka (Logistics), 5/2016, Logistyka – nauka (Logistic research),  

9–13. 

Ciesielski, M. (2011), Zarządzanie Łańcuchami Dostaw (Supply Chain Management), Polskie 

Wydawnictwo Ekonomiczne, Warszawa. 

Clements, M. P. (2005), Evaluating Econometric Forecasts of Economic and Financial Vari-

ables, Palgrave Macmillan, New York. 

Doman, M., Doman, R. (2009), Modelowanie Zmienności i Ryzyka. Metody Ekonometrii 

Finansowej (Risk and Volatility Modelling. Methods of Financial Econometrics), 

Wolters Kluwer Business, Kraków. 

Forecasting Principles, webpage sponsored by the International Institute of Forecasters, 

http://www.forecastingprinciples.com/index.php/data (26.09.2015). 

Gneiting, T. (2011), Quantiles as Optimal Point Forecasts, International Journal of Forecast-

ing, 27, 197–207, DOI: http://dx.doi.org/10.1016/j.ijforecast.2009.12.015. 

Granger, C. W. J. (1999), Outline of Forecast Theory Using Generalized Cost Functions, 

Spanish Economic Review, 1, 161–173,   

DOI: http://dx.doi.org/10.1007/s101080050007. 

Koenker, R. (2005), Quantile Regression, Cambridge University Press, Cambridge. 

                                                 
8 See Applied Forecasting, http://www.appliedforecasting.com/quantiles-forecasts-for-

inventory-optimization/. 

http://www.appliedforecasting.com/quantiles-forecasts-for-inventory-optimization/
http://www.forecastingprinciples.com/index.php/data
http://ideas.repec.org/a/eee/intfor/v27yi2p197-207.html
http://ideas.repec.org/s/eee/intfor.html
http://ideas.repec.org/s/eee/intfor.html
http://dx.doi.org/10.1016/j.ijforecast.2009.12.015


Joanna Bruzda  

DYNAMIC ECONOMETRIC MODELS 16 (2016) 5–20 

16 

Makridakis, S, Hibon, M (2000), The M3-competition: results, conclusions and implications, 

International Journal of Forecasting, 16, 451–476,  

 DOI: http://dx.doi.org/10.1016/S0169-2070(00)00057-1. 

R Archive Network, https://cran.r-project.org/web/packages/bayesm/index.html (26.09.2015). 

SAP Help Portal, http://help.sap.com (26.09.2015). 

Wagner, B. (2010), Purchasing and Forecasting Using SAP ERP, Operations and Supply 

Chain Management Library, http://scn.sap.com/community/uac/operations-and-scm-

library (26.09.2015). 

Prognozy kwantylowe w planowaniu operacyjnym i zarządzaniu  
zapasami – wstępna weryfikacja empiryczna 

Z a r y s  t r e ś c i. W artykule prezentuje się wyniki wstępnej weryfikacji empirycznej metod 

prognozowania kwantylowego mających zastosowanie w logistyce do ustalania punktu od-

nowienia i granicy uzupełniania zapasów czy optymalnej wielkości zamówienia w modelu 

jednookresowym. Porównaniem objęto 26 procedur, a w tym regresję kwantylową, podsta-

wową metodę bootstrapową i popularne formuły podręcznikowe. Wyniki otrzymane na bazie 

analizy 30 szeregów czasowych dotyczących tak różnorodnych zjawisk jak sprzedaż w su-

permarkecie, przewozy pasażerskie i zużycie gazu i wody wskazują na użyteczność median 

regresyjnych, kwantyli regresyjnych i procedur dostępnych w SAP ERP. 

S ł o w a  k l u c z o w e: funkcja straty LINLIN, prognozy kwantylowe, punkt odnowienia, 

regresja kwantylowa, wyrównywanie wykładnicze.  

 

http://dx.doi.org/10.1016/S0169-2070%2800%2900057-1
https://cran.r-project.org/web/packages/bayesm/index.html
http://help.sap.com/
http://scn.sap.com/community/uac/operations-and-scm-library
http://scn.sap.com/community/uac/operations-and-scm-library
http://scn.sap.com/community/uac/operations-and-scm-library
http://scn.sap.com/community/uac/operations-and-scm-library


Quantile Forecasting in Operational Planning and Inventory Management… 

DYNAMIC ECONOMETRIC MODELS 16 (2016) 5–20 

17 

Table 2. Rankings of forecasting procedures 

Method 
Percentage (number) of 

best results Method 

Percentage (number) of 
best results for   

= 90%, 95% 

M11 8.10% (17) M8 11.67% (7) 
M22 8.10% (17) M10 11.67% (7) 
M21 6.67% (14) M4 8.33% (5) 
M8 6.19% (13) M19 8.33% (5) 
M25 6.19% (13) M5 6.67% (4) 
M10 5.71% (12) M3 5% (3) 
M23 5.71% (12) M6 5% (3) 
M24 5.71% (12) M11 5% (3) 
M4 4.76% (10) M15 5% (3) 
M19 4.76% (10) M25 5% (3) 
M5 4.29% (9) M14 3.33% (2) 
M3 3.81% (8) M17 3.33% (2) 
M7 3.81% (8) M21 3.33% (2) 
M9 3.33% (7) M23 3.33% (2) 
M18 3.33% (7) M1 1.67% (1) 
M6 2.86% (6) M2 1.67% (1) 
M15 3.33% (7) M7 1.67% (1) 
M17 2.86% (6) M9 1.67% (1) 
M14 2.86% (6) M12 1.67% (1) 
M20 2.86% (6) M16 1.67% (1) 
M12 1.90% (4) M20 1.67% (1) 
M16 1.43% (3) M22 1.67% (1) 
M1 0.95% (2) M24 1.67% (1) 
M26 0.95% (2) M13 0% (0) 
M2 0.48% (1) M18 0% (0) 
M13 0.48% (1) M26 0% (0) 

Note: The best result means that a method produced the lowest value of the LINLIN loss (3) for a given 
series and a given quantile. 



 

 

F
ig

u
re

 1
. 

 T
h
e 

ti
m

e 
se

ri
es

 u
se

d
 i

n
 t

h
e 

st
u

d
y

 

2
0

4
0

6
0

0

5
0
0
0

1
0
0
0
0

x
1

2
0

4
0

6
0

0

5
0
0
0

1
0
0
0
0

x
2

2
0

4
0

6
0

2
0
0
0

4
0
0
0

6
0
0
0

x
3

2
0

4
0

6
0

2
0
0
0

4
0
0
0

6
0
0
0

x
4

2
0

4
0

6
0

2
0
0
0

4
0
0
0

6
0
0
0

x
5

2
0

4
0

6
0

0

5
0
0
0

1
0
0
0
0

x
6

2
0

4
0

6
0

1
5
0
0

2
0
0
0

2
5
0
0

x
7

2
0

4
0

6
0

2
0
0
0

2
5
0
0

3
0
0
0

x
8

2
0

4
0

6
0

2
0
0
0

2
2
0
0

2
4
0
0

x
9

2
0

4
0

6
0

2
0
0
0

2
2
0
0

2
4
0
0

x
1
0

2
0

4
0

6
0

2
0
0
0

2
2
0
0

2
4
0
0

x
1
1

2
0

4
0

6
0

1
5
0
0

2
0
0
0

2
5
0
0

x
1
2

2
0

4
0

6
0

2
0
0
0

2
2
0
0

2
4
0
0

x
1
3

2
0

4
0

6
0

1
5
0
0

2
0
0
0

2
5
0
0

x
1
4

2
0

4
0

6
0

0

5
0
0
0

1
0
0
0
0

x
1
5

2
0

4
0

6
0

2
0
0
0

4
0
0
0

6
0
0
0

x
1
6

2
0

4
0

6
0

012
x
 1

0
4

x
1
7

2
0

4
0

6
0

0

5
0
0
0

1
0
0
0
0

x
1
8

2
0

4
0

6
0

0

2
0
0
0

4
0
0
0

x
1
9

2
0

4
0

6
0

0

2
0
0
0

4
0
0
0

x
2
0

2
0

4
0

6
0

0

5
0
0
0

1
0
0
0
0

x
2
1

2
0

4
0

6
0

1
0
0
0

2
0
0
0

3
0
0
0

x
2
2

2
0

4
0

6
0

012
x
 1

0
4

x
2
3

2
0

4
0

6
0

0
.51

1
.5

x
 1

0
4

x
2
4

2
0

4
0

6
0

0

5
0
0
0

1
0
0
0
0

x
2
5

2
0

4
0

6
0

012
x
 1

0
4

x
2
6

2
0

4
0

6
0

024
x
 1

0
4

x
2
7

2
0

4
0

6
0

4

4
.55

x
 1

0
4

x
2
8

2
0

4
0

6
0

2
0

3
0

4
0

x
2
9

2
0

4
0

6
0

0

5
0
0

1
0
0
0

x
3
0



 T
ab

le
 3

. 
M

ea
n
 v

al
u
es

 o
f 

th
e 

re
la

ti
v
e 

L
IN

L
IN

 c
o

st
 f

u
n
ct

io
n

 

M
et

h
o

d
s 

M
ea

n
 r

el
at

iv
e 
L
L

 
T

ri
m

m
ed

 m
ea

n
 r

el
at

iv
e 
L
L

 

0.
05

 
0.

1 
0.

25
 

0.
5 

0.
75

 
0.

9 
0.

95
 

0.
05

 
0.

1 
0.

25
 

0.
5 

0.
75

 
0.

9 
0.

95
 

M
1 

0.
01

62
 

0.
02

55
 

0.
03

86
 

0.
04

41
 

0.
03

38
 

0.
01

68
 

0.
00

93
 

0.
01

31
 

0.
02

16
 

0.
03

35
 

0.
03

88
 

0.
03

12
 

0.
01

56
 

0.
00

8 
M

2 
0.

01
59

 
0.

02
5 

0.
03

85
 

0.
04

41
 

0.
03

36
 

0.
01

67
 

0.
00

93
 

0.
01

27
 

0.
02

09
 

0.
03

33
 

0.
03

88
 

0.
03

12
 

0.
01

56
 

0.
00

82
 

M
3 

0.
01

62
 

0.
02

46
 

0.
03

85
 

0.
04

45
 

0.
03

37
 

0.
01

69
 

0.
01

02
 

0.
01

3 
0.

02
21

 
0.

03
4 

0.
03

92
 

0.
03

08
 

0.
01

55
 

0.
00

87
 

M
4 

0.
01

67
 

0.
02

47
 

0.
03

69
 

0.
04

28
 

0.
03

22
 

0.
01

63
 

0.
00

92
 

0.
01

43
 

0.
02

18
 

0.
03

32
 

0.
03

93
 

0.
03

05
 

0.
01

52
 

0.
00

85
 

M
5 

0.
01

64
 

0.
02

36
 

0.
03

62
 

0.
04

28
 

0.
03

24
 

0.
01

58
 

0.
00

96
 

0.
01

36
 

0.
02

1 
0.

03
26

 
0.

03
93

 
0.

03
03

 
0.

01
47

 
0.

00
89

 
M

6 
0.

01
82

 
0.

04
26

 
0.

03
9 

0.
04

38
 

0.
03

38
 

0.
01

75
 

0.
01

01
 

0.
01

38
 

0.
02

16
 

0.
03

37
 

0.
03

84
 

0.
03

13
 

0.
01

57
 

0.
00

86
 

M
7 

0.
01

29
 

0.
02

03
 

0.
03

32
 

0.
04

26
 

0.
03

44
 

0.
01

92
 

0.
01

18
 

0.
01

17
 

0.
01

92
 

0.
03

03
 

0.
03

83
 

0.
03

09
 

0.
01

62
 

0.
00

91
 

M
8 

0.
01

8 
0.

02
41

 
0.

03
78

 
0.

04
28

 
0.

03
05

 
0.

01
76

 
0.

01
11

 
0.

01
51

 
0.

02
06

 
0.

03
24

 
0.

03
93

 
0.

02
94

 
0.

01
68

 
0.

00
99

 
M

9 
0.

01
75

 
0.

02
57

 
0.

03
48

 
0.

03
75

 
0.

03
02

 
0.

01
62

 
0.

00
96

 
0.

01
38

 
0.

02
23

 
0.

03
16

 
0.

03
34

 
0.

02
82

 
0.

01
53

 
0.

00
88

 
M

10
 

0.
01

73
 

0.
02

37
 

0.
03

54
 

0.
03

75
 

0.
03

02
 

0.
01

72
 

0.
00

99
 

0.
01

4 
0.

02
02

 
0.

03
13

 
0.

03
34

 
0.

02
77

 
0.

01
57

 
0.

00
85

 
M

11
 

0.
01

38
 

0.
02

41
 

0.
03

42
 

0.
03

7 
0.

03
06

 
0.

01
73

 
0.

01
07

 
0.

01
11

 
0.

01
92

 
0.

03
05

 
0.

03
28

 
0.

02
82

 
0.

01
63

 
0.

00
91

 
M

12
 

0.
01

44
 

0.
02

28
 

0.
03

6 
0.

04
14

 
0.

03
32

 
0.

01
7 

0.
01

04
 

0.
01

22
 

0.
01

97
 

0.
03

21
 

0.
03

72
 

0.
03

06
 

0.
01

58
 

0.
00

93
 

M
13

 
0.

01
42

 
0.

02
25

 
0.

03
59

 
0.

04
14

 
0.

03
31

 
0.

01
69

 
0.

01
02

 
0.

01
21

 
0.

01
95

 
0.

03
21

 
0.

03
72

 
0.

03
07

 
0.

01
58

 
0.

00
93

 
M

14
 

0.
01

4 
0.

02
3 

0.
03

56
 

0.
04

18
 

0.
03

29
 

0.
01

76
 

0.
01

12
 

0.
01

21
 

0.
02

01
 

0.
03

21
 

0.
03

81
 

0.
02

99
 

0.
01

59
 

0.
00

96
 

M
15

 
0.

01
5 

0.
02

27
 

0.
03

5 
0.

04
11

 
0.

03
11

 
0.

01
56

 
0.

00
89

 
0.

01
24

 
0.

01
98

 
0.

03
19

 
0.

03
75

 
0.

02
93

 
0.

01
53

 
0.

00
85

 
M

16
 

0.
01

47
 

0.
02

22
 

0.
03

5 
0.

04
11

 
0.

03
19

 
0.

01
56

 
0.

01
01

 
0.

01
29

 
0.

02
 

0.
03

2 
0.

03
75

 
0.

02
99

 
0.

01
52

 
0.

00
98

 
M

17
 

0.
01

47
 

0.
02

21
 

0.
03

54
 

0.
04

15
 

0.
03

2 
0.

01
73

 
0.

01
03

 
0.

01
19

 
0.

01
85

 
0.

03
12

 
0.

03
79

 
0.

03
09

 
0.

01
61

 
0.

00
93

 
M

18
 

0.
01

34
 

0.
02

15
 

0.
03

63
 

0.
04

36
 

0.
03

51
 

0.
01

93
 

0.
01

13
 

0.
01

14
 

0.
01

87
 

0.
03

19
 

0.
03

83
 

0.
03

23
 

0.
01

66
 

0.
00

93
 

M
19

 
0.

01
75

 
0.

02
41

 
0.

03
67

 
0.

04
11

 
0.

03
08

 
0.

01
94

 
0.

01
09

 
0.

01
46

 
0.

02
13

 
0.

03
25

 
0.

03
75

 
0.

02
94

 
0.

01
86

 
0.

00
98

 
M

20
 

0.
01

47
 

0.
02

3 
0.

03
35

 
0.

03
74

 
0.

03
 

0.
01

64
 

0.
00

98
 

0.
01

2 
0.

02
02

 
0.

03
06

 
0.

03
35

 
0.

02
75

 
0.

01
56

 
0.

00
9 

M
21

 
0.

01
38

 
0.

02
09

 
0.

03
42

 
0.

03
74

 
0.

03
03

 
0.

01
73

 
0.

00
99

 
0.

01
17

 
0.

01
84

 
0.

03
09

 
0.

03
35

 
0.

02
77

 
0.

01
56

 
0.

00
87

 
M

22
 

0.
01

4 
0.

02
2 

0.
03

36
 

0.
03

69
 

0.
03

09
 

0.
01

77
 

0.
01

09
 

0.
01

19
 

0.
01

78
 

0.
03

09
 

0.
03

3 
0.

02
9 

0.
01

71
 

0.
00

99
 

M
23

 
0.

01
9 

0.
02

69
 

0.
04

16
 

0.
04

77
 

0.
03

54
 

0.
01

75
 

0.
00

99
 

0.
01

33
 

0.
02

17
 

0.
03

41
 

0.
03

93
 

0.
03

04
 

0.
01

58
 

0.
00

92
 

M
24

 
0.

01
79

 
0.

02
59

 
0.

04
 

0.
04

58
 

0.
03

39
 

0.
01

65
 

0.
00

92
 

0.
01

4 
0.

02
22

 
0.

03
41

 
0.

03
95

 
0.

03
07

 
0.

01
59

 
0.

00
85

 
M

25
 

0.
01

56
 

0.
02

51
 

0.
03

92
 

0.
04

54
 

0.
03

56
 

0.
01

97
 

0.
01

12
 

0.
01

26
 

0.
02

07
 

0.
03

24
 

0.
03

85
 

0.
03

14
 

0.
01

7 
0.

00
98

 
M

26
 

0.
01

51
 

0.
02

34
 

0.
03

8 
0.

04
37

 
0.

03
5 

0.
01

83
 

0.
01

05
 

0.
01

24
 

0.
01

99
 

0.
03

35
 

0.
03

89
 

0.
03

19
 

0.
01

7 
0.

00
96

 

N
o
te

: 
T

h
e 

b
es

t 
o
u
tc

o
m

es
 a

re
 i

n
 b

o
ld

. 
 

 
 



 

 T
ab

le
 4

. 
M

ea
n
 v

al
u
es

 o
f 

th
e 

h
it

 r
at

io
s 

M
et

h
o

d
s 

M
ea

n
 h

it
 r

at
io

s 
T

ri
m

m
ed

 m
ea

n
 h

it
 r

at
io

s 

0.
05

 
0.

1 
0.

25
 

0.
5 

0.
75

 
0.

9 
0.

95
 

0.
05

 
0.

1 
0.

25
 

0.
5 

0.
75

 
0.

9 
0.

95
 

M
1 

0.
93

33
 

0.
81

67
**

 
0.

69
17

 
0.

37
5*

* 
0.

20
83

 
0.

08
33

 
0.

03
33

 
0.

96
88

 
0.

85
42

 
0.

70
83

 
0.

36
46

 
0.

18
75

 
0.

05
21

 
0 

M
2 

0.
92

5 
0.

82
5 

0.
68

33
 

0.
37

5*
* 

0.
20

83
 

0.
08

33
 

0.
03

33
 

0.
95

83
 

0.
86

46
 

0.
69

79
 

0.
36

46
 

0.
18

75
 

0.
05

21
 

0 
M

3 
0.

86
67

**
 

0.
8*

* 
0.

66
67

 
0.

4 
0.

20
83

 
0.

08
33

 
0.

03
33

 
0.

90
63

 
0.

83
33

 
0.

68
75

 
0.

38
54

 
0.

17
71

 
0.

05
21

 
0 

M
4 

0.
89

17
 

0.
85

83
 

0.
69

17
 

0.
40

83
 

0.
22

5 
0.

08
33

 
0.

05
 

0.
92

71
 

0.
88

54
 

0.
71

88
 

0.
39

58
 

0.
21

88
 

0.
05

21
 

0.
02

08
 

M
5 

0.
89

17
 

0.
81

67
**

 
0.

62
5*

* 
0.

40
83

 
0.

21
67

 
0.

06
67

 
0.

03
33

 
0.

92
71

 
0.

84
38

 
0.

63
54

 
0.

39
58

 
0.

19
79

 
0.

03
13

 
0 

M
6 

0.
9 

0.
83

33
 

0.
67

5 
0.

36
67

**
 

0.
20

83
 

0.
07

5 
0.

04
17

 
0.

92
71

 
0.

86
46

 
0.

67
71

 
0.

35
42

 
0.

18
75

 
0.

03
13

 
0.

01
04

 
M

7 
0.

89
17

 
0.

85
 

0.
63

33
**

 
0.

38
33

 
0.

21
67

 
0.

1 
0.

06
67

 
0.

91
67

 
0.

88
54

 
0.

64
58

 
0.

37
5 

0.
19

79
 

0.
05

21
 

0.
04

17
 

M
8 

0.
88

33
**

 
0.

86
67

 
0.

63
33

**
 

0.
40

83
 

0.
2 

0.
1 

0.
09

17
 

0.
91

67
 

0.
89

58
 

0.
64

58
 

0.
39

58
 

0.
16

67
 

0.
07

29
 

0.
06

25
 

M
9 

0.
90

83
 

0.
85

 
0.

77
5 

0.
43

33
 

0.
17

5 
0.

05
83

 
0.

03
33

 
0.

92
71

 
0.

88
54

 
0.

80
21

 
0.

42
71

 
0.

15
63

 
0.

03
13

 
0.

01
04

 
M

10
 

0.
89

17
 

0.
85

 
0.

72
5 

0.
43

33
 

0.
15

83
 

0.
10

83
 

0.
05

 
0.

90
63

 
0.

85
42

 
0.

73
96

 
0.

42
71

 
0.

13
54

 
0.

09
38

 
0.

03
13

 
M

11
 

0.
90

83
 

0.
86

67
 

0.
72

5 
0.

42
5 

0.
20

83
 

0.
09

17
 

0.
06

67
 

0.
93

75
 

0.
90

63
 

0.
73

96
 

0.
41

67
 

0.
19

79
 

0.
06

25
 

0.
04

17
 

M
12

 
0.

92
5 

0.
81

67
**

 
0.

70
83

 
0.

41
67

 
0.

2 
0.

05
83

 
0.

04
17

 
0.

95
83

 
0.

85
42

 
0.

72
92

 
0.

40
63

 
0.

17
71

 
0.

02
08

 
0.

01
04

 
M

13
 

0.
91

67
 

0.
82

5 
0.

70
83

 
0.

41
67

 
0.

2 
0.

06
67

 
0.

04
17

 
0.

94
79

 
0.

86
46

 
0.

72
92

 
0.

40
63

 
0.

17
71

 
0.

03
13

 
0.

01
04

 
M

14
 

0.
89

17
 

0.
80

83
**

 
0.

68
33

 
0.

40
83

 
0.

2 
0.

1 
0.

05
 

0.
93

75
 

0.
84

38
 

0.
70

83
 

0.
39

58
 

0.
17

71
 

0.
05

21
 

0.
02

08
 

M
15

 
0.

87
5*

* 
0.

83
33

 
0.

7 
0.

4 
0.

21
67

 
0.

07
5 

0.
02

5 
0.

90
63

 
0.

86
46

 
0.

71
88

 
0.

39
58

 
0.

20
83

 
0.

05
21

 
0 

M
16

 
0.

9 
0.

80
83

**
 

0.
65

 
0.

4 
0.

2 
0.

08
33

 
0.

04
17

 
0.

92
71

 
0.

83
33

 
0.

66
67

 
0.

39
58

 
0.

17
71

 
0.

05
21

 
0.

01
04

 
M

17
 

0.
89

17
 

0.
83

33
 

0.
65

83
 

0.
4 

0.
2 

0.
05

83
 

0.
04

17
 

0.
93

75
 

0.
87

5 
0.

67
71

 
0.

39
58

 
0.

17
71

 
0.

02
08

 
0.

01
04

 
M

18
 

0.
89

17
 

0.
82

5 
0.

62
5*

* 
0.

4 
0.

20
83

 
0.

1 
0.

06
67

 
0.

92
71

 
0.

87
5 

0.
63

54
 

0.
39

58
 

0.
18

75
 

0.
06

25
 

0.
04

17
 

M
19

 
0.

87
5*

* 
0.

85
83

 
0.

65
 

0.
4 

0.
2 

0.
14

17
 

0.
09

17
 

0.
90

63
 

0.
88

54
 

0.
65

63
 

0.
39

58
 

0.
17

71
 

0.
11

46
 

0.
06

25
 

M
20

 
0.

88
33

**
 

0.
84

17
 

0.
75

 
0.

43
33

 
0.

17
5 

0.
05

83
 

0.
02

5 
0.

92
71

 
0.

87
5 

0.
77

08
 

0.
43

75
 

0.
15

63
 

0.
03

13
 

0 
M

21
 

0.
88

33
**

 
0.

80
83

**
 

0.
71

67
 

0.
43

33
 

0.
19

17
 

0.
09

17
 

0.
04

17
 

0.
89

58
 

0.
82

29
 

0.
73

96
 

0.
43

75
 

0.
17

71
 

0.
07

29
 

0.
02

08
 

M
22

 
0.

89
17

 
0.

85
 

0.
72

5 
0.

45
 

0.
2 

0.
11

67
 

0.
05

 
0.

91
67

 
0.

88
54

 
0.

75
 

0.
45

83
 

0.
18

75
 

0.
08

33
 

0.
03

13
 

M
23

 
0.

90
83

 
0.

85
 

0.
68

33
 

0.
39

17
 

0.
21

67
 

0.
06

67
 

0.
03

33
 

0.
93

75
 

0.
89

58
 

0.
70

83
 

0.
37

5 
0.

18
75

 
0.

03
13

 
0.

01
04

 
M

24
 

0.
90

83
 

0.
83

33
 

0.
67

5 
0.

38
33

 
0.

19
17

 
0.

07
5 

0.
01

67
 

0.
94

79
 

0.
87

5 
0.

69
79

 
0.

36
46

 
0.

16
67

 
0.

04
17

 
0 

M
25

 
0.

90
83

 
0.

85
83

 
0.

68
33

 
0.

38
33

 
0.

2 
0.

10
83

 
0.

06
67

 
0.

94
79

 
0.

90
63

 
0.

70
83

 
0.

36
46

 
0.

16
67

 
0.

07
29

 
0.

03
13

 
M

26
 

0.
91

67
 

0.
83

33
 

0.
68

33
 

0.
4 

0.
18

33
 

0.
1 

0.
05

 
0.

95
83

 
0.

86
46

 
0.

70
83

 
0.

38
54

 
0.

15
63

 
0.

05
21

 
0.

02
08

 

 N
o

te
: 

In
 t

h
e 

le
ft

-h
an

d
 p

an
el

 o
f 

T
ab

le
 4

 h
it

 r
at

io
s 

si
g
n

if
ic

an
tl

y
 d

ev
ia

ti
n

g
 (

at
 t

h
e 

1
%

 l
ev

el
) 

fr
o
m

 t
h

ei
r 

o
p

ti
m

al
 v

al
u

es
 a

re
 d

en
o
te

d
 b

y
 ‘

*
*
’.

 


