1 Introduction
This paper focuses on one of the most frequently asked questions in forecasting theory and practice: which technique(s) should I choose to forecast this time series? In literature, this question has been posed many times and has indeed been answered by benchmarks and competitions (Armstrong and Fildes, 2006; Crone et al., 2011; Petropoulos et al., 2014), as forecasting has been an integral part of the business decisionmaking process for decades and is used for this purpose in many industries (Armstrong and Fildes, 2006; Cang and Yu, 2014; Lessmann and Voß, 2017). However, most studies only take one evaluation criterion into account, i.e. the performance of the techniques on a test set, while the final choice of a model in a business context depends on more considerations. Undoubtedly, the costs that are associated with inaccurate forecasts make sure that accuracy will always remain an important evaluation standard (Kahn, 2003). However, from a decisionmaking perspective, other questions immediately arise in the mind of the business expert as well, such as the potential impact of the forecast on the revenue of the company or the maintenance cost of the model. This paper therefore proposes an expected profit function that can be integrated into several steps of the forecasting process, while also taking a closer look at which types of models perform best in a sales forecast on a tactical or strategic level.
Recent publications have shown a large offering of forecasting techniques, ranging from the statistical methods to machine learning techniques. Given all of these theoretical and technological developments, it is becoming increasingly difficult to select the right type of technique for a given use case. Especially the group of Machine Learning (ML) techniques has received a lot of attention recently, as it constitutes one of the most popular topics in forecasting literature (Fildes, 2006). Most articles on ML techniques report favourable results when compared to more traditional methodologies, both for single use cases and more extensive comparisons (Crone et al., 2011), although publications generally have a tendency to only report on positive outcomes (Armstrong, 2006). However, several authors have expressed their reservations concerning these complex techniques (Makridakis and Hibon, 2000). In contrast, (Crone et al., 2011) have shown that machine learning has caught up with statistical modelling and should not be dismissed lightly for forecasting exercises. This paper therefore also aims to investigate whether these more complex ML techniques truly outperform the classical models for a tactical sales forecast.
In this paper, we will focus on the field of sales forecasting, as successful sales forecasts are vital in both short and longterm strategic and financial planning (Ramos et al., 2015). This research specifically deals with highlevel forecasts, which are primarily meant for decisionmaking purposes, as opposed to inventory planning for specific products. In practice, this typically entails a monthly time series, which is nonintermittent and is prone to display a trend, a seasonal pattern or a combination of these characteristics. This type of time series is common in other fields as well, and has therefore frequently been used for benchmarking purposes (Armstrong and Fildes, 2006; Crone et al., 2011; Petropoulos et al., 2014). We therefore take a look at the performance of techniques that model seasonality versus methodologies that do not have this ability, as season is a typical characteristic of sales time series. While trying out nonseasonal models might seems counterintuitive for this data, a lot of the more recently developed techniques do not have a seasonal component and still seem to perform very well for many applications. This paper therefore also investigates whether this type of model can perform well on these seasonal time series, given the necessary preprocessing of the data. Furthermore, this highlevel data also raises the question of the usefulness of incorporating external factors into the forecast. While th addition of variables has obvious benefits, such as the explanatory value, it frequently leads to higher model maintainability costs. Thus, we also compare univariate techniques with and without the ability to add external drivers to one another in this paper.
Our contributions are twofold, as we aim to both benchmark a large set of forecasting techniques and integrate a practical construct into the model building and evaluating process, i.e. profit. Firstly, we propose a new strategy to inject a profitoriented view into the entire forecasting process without explicitly forecasting profit itself. In practice, this constitutes a different way of performing feature selection, tuning hyper parameters and evaluating the forecasting techniques with the goal of achieving the models that yield the highest expected profit. The expected profit function that is used for this purpose, is easy to understand and adaptable to any situation by combining forecasting accuracy with business expertise
(Van Calster et al., 2017). Furthermore, our methodology ensures a completely automated and datadriven model building process. Secondly, we benchmark a large range of forecasting techniques according to three different categorizations. As mentioned above, we contrast a range of complex techniques and traditional techniques, in order to assess whether the ML techniques are truly able to perform equally in regards to tactical sales forecasts. Secondly, we take the seasonal characteristics of sales time series into consideration by distinguishing techniques that model seasonality themselves and methods that require seasonal dummy variables to achieve the same goal. Finally, we contrast techniques with and without variables, as we investigate the value of external factors in a highlevel sales forecast. In terms of evaluating the techniques, we take accuracy, expected profit, model complexity and model interpretability into consideration in order to integrate the business aspect of forecasting into the benchmark. In the end, we aim to quantitatively select the techniques that forecast accurately, lead to the highest expected profit for any business case, and make the most sense from a business perspective. We will address these research questions by means of a total of 35 monthly sales datasets. The datasets were collected from both The CocaCola Company and from publicly available resources in order to add to the generalizability of the study.
The paper is organized as follows. Section 2 deals with the related work that provides a necessary background to the research questions. Section 3 describes the datasets, the forecasting techniques and the general methodology of the experiments. Next, section 4 focuses on the results of the research, while section 5 includes the conclusion.
2 Related work
This section on related work focuses on the necessary background literature for the research questions. We take a closer look at the forecasting literature on benchmarking, while also considering recent literature on profitoriented analytics.
2.1 Benchmarking in forecasting
Forecasts are typically performed by three categories of techniques (Cang and Yu, 2014): traditional time series analysis (AboagyeSarfo et al., 2015; Akın, 2015; Arunraj and Ahrens, 2015; Athanasopoulos et al., 2011; Franses and Van Dijk, 2005; GilAlana et al., 2008; Gunter and Önder, 2015; Petropoulos et al., 2014; Ramos et al., 2015; Santos et al., 2012), causal regression techniques (Akın, 2015; Arunraj and Ahrens, 2015; Bozos and Nikolopoulos, 2011; Cang and Yu, 2014; Lessmann et al., 2015; Ma et al., 2016; Nikolopoulos et al., 2007)
, and more complex artificial intelligence techniques
(Akın, 2015; Bozos and Nikolopoulos, 2011; Cang and Yu, 2014; Crone et al., 2011; Fagiani et al., 2015; Lessmann et al., 2015; Taylor et al., 2006). The emergence of new techniques often requires a comparison with former methods, which leads to an extensive literature on benchmarking, both for individual use cases (AboagyeSarfo et al., 2015; Arunraj and Ahrens, 2015; Bozos and Nikolopoulos, 2011; GilAlana et al., 2008; Gunter and Önder, 2015; Lessmann et al., 2015) and for larger sets of time series (Athanasopoulos et al., 2011; Cang and Yu, 2014; Crone et al., 2011; Franses and Van Dijk, 2005; Ma et al., 2016; Makridakis and Hibon, 2000; Petropoulos et al., 2014; Weller and Crone, 2012). This research consists of both fieldspecific (AboagyeSarfo et al., 2015; Akın, 2015; Athanasopoulos et al., 2011; Bozos and Nikolopoulos, 2011; Cang and Yu, 2014; Fagiani et al., 2015; Lessmann et al., 2015; Ma et al., 2016; Weller and Crone, 2012) and industryneutral benchmarks, which are oriented towards general conclusions (Crone et al., 2011; Makridakis and Hibon, 2000; Petropoulos et al., 2014). While some studies use a combination of generated data and industry data (Petropoulos et al., 2014), most use reallife datasets to answer their research questions (Bozos and Nikolopoulos, 2011; Cang and Yu, 2014; Lessmann et al., 2015; Weller and Crone, 2012).In terms of the conclusions that have come out of the larger studies, some discrepancies arise. While several studies point out that the newer ML techniques do not perform as well as the more traditional methods for classical time series (Makridakis and Hibon, 2000), others claim that these complex techniques have caught up in recent years (Crone et al., 2011). In this paper, we therefore take a look at a wider range of techniques from all three categories that were mentioned above. Furthermore, we also contrast techniques with and without external factors, which adds another factor that has not been part of many larger benchmarking studies, except for (Athanasopoulos et al., 2011). Our paper combines these elements in an extensive benchmark that is based on publicly available data and recent sales time series.
2.2 Profitdriven analytics
Profitdriven analytics has recently become a hot topic in analytics, as businesses are interested in the actual value that predictive models generate or the influence that they have on their eventual net profits. Integrating this valuecentric view into analytics, has led to a growing number of profitdriven methodologies, techniques and metrics (Verbeke et al., 2017)
. These profit functions can be used in different steps of the model building and model selection process. For example, profit has been used as an evaluation metric for benchmarks in different fields
(Óskarsdóttir et al., 2017; Verbraken et al., 2013), while it has inspired entire profitdriven algorithms as well (Stripling et al., 2015; Verbeke et al., 2012). In this paper, we aim to integrate this profitoriented view into multiple steps of the forecasting process instead of only using it as an additional evaluation criterion.In forecasting, research on the profit aspect is scarcer than in other fields. While the monetary value of classification models has been extensively reviewed, the same cannot be said for regression models. However, the impact of forecasting accuracy on net profit is an interesting subject, as under and overforecasting both lead to completely different costs. The former might lead to a loss in sales and outofstock products, while the latter can lead to overstock and storage costs. While both directions for the error inevitably bring about a loss of profit, they are often not equal. Completely symmetric profit loss functions that are solely based on accuracy measures are therefore not representative of the real world. The ultimate goal of profitoriented analytics is to find the model with the best balance between costs and accuracy. While these two concepts are inevitably linked in a forecasting exercise, we cannot state that they are exactly the same. Therefore, profitoriented benchmarking should take into account both traditional accuracy metrics and metrics that point to the costs of the forecast, such as expected profit functions or model complexity. So far, two different views on the integration of profit into forecasting exercises have been proposed in recent literature. The first perspective optimizes an asymmetric loss function during model training to model the imbalance between over and underforecasting.
(Crone et al., 2005)applies this methodology to neural networks, while
(Yang et al., 2002)take a closer look at support vector regression models. The second way of integrating profit into a forecasting function takes place after the training process.
(Bansal et al., 2008) propose a tuning procedure that modifies the predictions so they are costoptimal. (Zhao et al., 2011) further finetune this procedure. (Bozos and Nikolopoulos, 2011) also take a monetary value into account when evaluating their forecasts, but do not modify the models in any way.In this paper, we take profit into account in all of the steps that are mentioned above. We optimize the parameters of our models, select features when necessary and evaluate our forecasts based on an asymmetric expected profit function that can easily be adjusted to any business case.
3 Methodology
This methodology section is divided into five parts. We begin by describing the datasets and by explaining the profit function that was used for both optimization and evaluation purposes in this paper. Next, the general experimental setup is introduced, which also includes the description of the feature selection procedure. The fourth subsection is dedicated to an overview of the forecasting techniques, while the last subsection focuses on evaluation metrics.
3.1 Data
The data sets in this paper stem from two sources. Firstly, The CocaCola Company has given us a total of 20 time series, which represent two of their product categories in ten different countries. These monthly time series all range from January 2004 until September 2016. The external variables that correspond with these datasets, were collected by means of incompany data sources and are all based on information about the location of the data. Concretely, they consist of 20 variables that contain information on weather, macroeconomic indicators, holidays and pricing information. As weather information, 4 variables were included, such as temperature and precipitation, while 9 variables allude to macroeconomic information, such as GDP and CPI. Additionally, 3 factors refer to calendar effects of public holidays, while the final 4 variables relate to both incompany and competitor pricing. An overview of these variables can be found in Table 1. These external factors were selected according to data availability, but also take into account the literature on the interesting types of variables for sales forecasting. Several types of information have proven to be useful in this field, although this generally depends on the aggregation level of the time series (Syntetos et al., 2016) and the volatility of the time series (Currie and Rowley, 2010). Research has shown that factors such as weather (Bertrand et al., 2015), macroeconomic influences (Sagaert et al., 2017) and pricing and promotional information (Huang et al., 2014; Ma et al., 2016) all have an impact on sales.
Variable name  Explanation 

Weather  
Maximum temperature  Average daily maximum temperature weighted by population 
Maximum temperature squared  Square of average daily maximum temperature weighted by population 
Precipitation  Average daily precipitation volume 
Sunshine hours  Average daily number of sunshine hours 
Macroeconomic indicators  
Consumer Price Index  Seasonally adjusted percentage change of CPI with regards to the previous month 
Unemployment rate  Percentage of unemployment for entire population 
Exchange rate  Exchange rate with US dollar 
Shortterm interest rate  Shortterm interest rate in percentage per annum 
Industrial production  Seasonally adjusted percentage change of industrial production with regards to the previous month 
Merchandise import  Seasonally adjusted percentage change of Merchandise import with regards to the previous month 
Merchandise export  Seasonally adjusted percentage change of Merchandise export with regards to the previous month 
Gross Domestic Product  Seasonally adjusted annual rate, percentage change of GDP with regards to the previous month 
Private Consumption  Seasonally adjusted annual rate, percentage change of PC with regards to the previous month 
Holidays  
Public holiday  Number of public holidays per month 
Weekend  Number of public holidays in the weekend per month (possibility of long weekend) 
Tuesday/Thursday  Number of public holidays on Tuesday or Thursday per month (possibility of long weekend) 
Pricing  
Company price  Average product category price in US dollars 
Company price deflated  Average product category price in US dollars deflated by CPI 
Competitor price  Average product category price of the main competitor in US dollars 
Competitor price deflated  Average product category price of the main competitor in US dollars deflated by CPI 
Secondly, we include a total of 15 publicly available datasets with similar characteristics in the analyses, in order to increase the generalizability of our findings, which can mostly be found in The Time Series Data Library^{1}^{1}1https://datamarket.com/data/list/?q=provider:tsdl^{,}^{2}^{2}2https://opendata.socrata.com/Business/CarSalesData/da8msmts. The general features of these monthly time series are summarized in Table 2. As all of these datasets also include information on location, we collected twelve external variables that contain information on weather, macroeconomic indicators and holidays as well. Concretely, we include four weather variables, seven macroeconomic indicators and one holiday variable. The weather variables consist of the same features as defined in Table 1, while the macroeconomic information includes all features in Table 1 except Merchandise Import and Merchandise Export. Finally, the models with external factors also contain the number of public holidays for each month. Pricing information was not available for these datasets. The sources for these three categories are publicly available^{3}^{3}3https://crudata.uea.ac.uk/cru/data/hrg/cru_ts_3.23/crucy.1506241137.v3.23/^{,}^{4}^{4}4https://data.oecd.org/^{,}^{5}^{5}5https://pypi.python.org/pypi/holidays.
Name 

Range 

Location  

Beer  1  January 1956 – August 1995  476  Australia  
Car sales 1  1  January 1996 – December 2008  156  California  
Car sales 2  1  January 1960 – December 1968  108  Canada  
Champagne  1  January 1964 – September 1972  105  France  
Paper  1  January 1963 – December 1972  120  France  
Petrol  4  January 1971 – December 1991  252  USA  
Wine  6  January 1980 – July 1995  187  Australia 
3.2 Expected profit function
The evaluation of any predictive model is generally focused on the accuracy that it achieves on a test set. In this paper, however, we take both accuracy and a more businessoriented profit measure into account. The profit measure is represented by Equation 2, which is dependent on our definition of the Percentage Error (PE), which can be found in Equation 1. This profit measure was first defined in (Van Calster et al., 2017)
and represents an estimation of the expected profit of the target variable. The formula is very easy to interpret and can easily be adjusted to any business use case. The two fundamental components of the expected profit measure are the volume of the sales, as more sales lead to more profit, and the accuracy of the forecast, as bad forecasts inevitably lead to a loss of profit. Next to these two core elements, we introduce several parameters that integrate expert knowledge into the profit function.
(1) 
1cm1cm
(2) 
Firstly, the business user can influence the impact of the forecasting error on the expected profit by setting two parameters. The first one deals with how the size of the error is used as a penalization, as both over and underforecasting have proven to lead to various costs (Kahn, 2003). This penalization factor can be modified according to a specific circumstance in a datadriven manner by executing a sensitivity analysis on a validation set. In this instance, α is set at , which was determined in (Van Calster et al., 2017). Furthermore, the business expert can set penalization boundaries and , which indicate that any forecast that has a PE within these boundaries, does not lead to a significant impact on the final profit. Note that should always be larger than . For example, we set boundaries of error in both directions for The CocaCola Company use case ( and ). This leads to a preference for models with a small overforecast, as the expected profit is higher because of a larger volume of sales in that case. This tendency is deemed appropriate for the use case because the company considers this an investment in the future. However, the and parameters can also be set unequally, if the forecasting error has a larger impact on profit in one particular direction, or even be completely omitted, if every inaccuracy while forecasting leads to a loss of profit.
Secondly, the weight refers to the profit margin for the product or product category at hand. This weight can be expressed both relatively between different products and in absolute numbers, such as currencies. For The CocaCola Company use case, these weights were determined by the profit that the product actually generated in the last year of the original training set. It is important to note that these weights remain constant throughout the analyses once they are set by the training set of the first prediction. The actual profit of a product will fluctuate over time and is driven by many external factors that are not captured in the function. We have chosen to keep this parameter constant because of two reasons: ease of use and availability of profit data. While the first reason is selfexplanatory, the second one is tied to the particular use case of this paper. If data about the actual profit of a product is more readily available, this parameter can be used dynamically by updating it during the testing process. The profit in the analyses of this paper can therefore be viewed as the profit that the product will generate if business stays the same and must truly be interpreted as the expected profit. The weights for the publicly available datasets were chosen randomly with values between 0 and 3, and are displayed in Table 3.
Name  weights 

Beer  0.1 
Car sales 1  2.1 
Car sales 2  0.6 
Champagne  1.0 
Paper  1.9 
Petrol  0.1, 1.2, 0.1, 2.8 
Wine  2.2, 2.2, 2.1, 1.5, 0.4, 2.7 
3.3 Experimental setup
The general experimental setup consists of holdout sample forecasts for all datasets. Concretely, the time series are split up into training, validation and test sets. The test set includes the final two years of the data, which leads to 24 data points to forecast. The validation set then consists of the year before the date that will be forecast and is only used for feature selection and parameter tuning when necessary for the given technique. Parameter tuning is performed once on the first validation set in the testing procedure, in order to avoid computational issues in the testing procedure. However, the feature selection procedure is repeated every three months, in order to keep the model uptodate. Once the necessary variables and hyper parameters have been selected, the training and validation sets are merged together in order to forecast the test set. Both the training and validation sets change with every forecast, as the setup consists of an expanding window. In the end, we therefore collect 24 onemonth ahead forecasts for each technique and for each dataset. The complete experimental setup is visualized in Figure 1.
The feature selection procedure consists of a hybrid method, which is based on the combination of Minimum Redundancy Maximum Relevance criterion (mRMR) that was created by (Peng et al., 2005) as a filtering technique, and a simple incremental wrapper method. The mRMR method is a mutualinformation based algorithm that ranks the external factors according to their shared information with the target variable, while also taking into account their dependency on the other external factors. This can be achieved by finding the feature set with features that maximizes the relevance with the target class and minimizes the dependency between the independent variables. In short, this filter finds the features that maximize Equation 3, which combines Equation 4 (Dependency) and Equation 5 (Redundancy).
(3) 
(4) 
(5) 
In this paper, the first feature step in this paper selects either the top 15 or the top 10 ranking of features, for the CocaCola Company datasets and the public datasets respectively, and then passes this on to the next step. Next, a simple forward incremental wrapper method starts with the top feature of the ranking and forecasts the validation set. Consecutively, one feature is added at a time into the feature set until the entire top 15 or top 10 ranking is used in the forecasting model. This methodology therefore takes advantage of the initial ranking that was made by the mRMR filter. The feature set that will be used to forecast the test set, is selected out of these 15 or 10 options by maximizing the profit function, which is defined in Section 3.2. This entire procedure is explained in Algorithm 1. In our benchmark, is either 15 or 10, depending on the dataset at hand, and is equal to 12 months.
Feature selection is generally important because of two entirely different reasons. Firstly, some of the variables might be correlated or influenced by the same underlying information, which can lead to less accurate forecasts (Boivin and Ng, 2006). A feature selection procedure is therefore used to determine which set of variables has the highest predictive power, while also eliminating any possible multicollinearity. Secondly, feature selection is equally important from a business perspective, as transparent models also have an explanatory advantage. Business analysts are interested in gaining knowledge on which external factors might influence their target variable, which can be useful for strategic decisions (Athanasopoulos et al., 2011). However, this knowledge then also relates to the maintenance of the model, as the variables that never survive the feature selection procedure during testing, are not needed any longer.
3.4 Forecasting techniques
In order to conduct the necessary experiments, a total of 17 forecasting techniques were selected, which are summarized in Table 4. These techniques are categorized according to three different types of attributes in order to answer our research questions. Firstly, we organize the methods according to the ability to use them as univariate with and/or without external drivers. We define a univariate technique without variables as a technique that only makes use of the sales times series itself to predict the next month. Techniques that are able to include variables, however, also integrate the external drivers, such as the weather, to generate a prediction. 8 techniques can be used in both ways, such as regression models, when past sales values are encoded as independent variables, next to the aforementioned external factors. We therefore benchmark a total of 26 techniques in our final analysis. Secondly, Table 4
displays the ability of a technique to explicitly model the seasonality of a time series, as seasonality is a typical characteristic of the sales time series that we are considering in this paper. Thirdly, the forecasting techniques are classified into Machine Learning (ML) techniques and nonML techniques. Recently, a lot of forecasting literature has focused on these ML techniques and often reports them to be more accurate than traditional techniques. In order to simplify the issue of what is considered an ML technique and what is not, we chose to consider methods ML if they belong to one of the four following categories: decision tree learning, neural networks, support vector machines and knearest neighbours, as this last category is based on a clustering algorithm. These three categorizations will underpin the answer to which type of technique is best used to achieve an accurate sales forecast. Finally, the table also contains the hyper parameters that were selected beforehand, and their possible values. Tuning hyper parameters has proven to be essential for truly assessing how well a certain technique can perform
(Carrizosa et al., 2014), and is therefore an essential part of benchmarking in general. In this paper, the parameter selection was conducted by evaluating model performance on the validation set and by applying an exhaustive grid search methodology. The evaluation metric that was optimized, is again the expected profit function that was defined in Section 3.2. Note that only the parameters that are mentioned in 4 are set in this way.Model  Variables?  Seasonal?  ML?  Hyper parameters  Possible values  

No  Yes  No  /  /  
Seasonal ARIMA  No  Yes  No 

[0, 5]  

No  Yes  No  /  /  
Seasonal random walk  No  Yes  No  /  /  
ARMAGARCH  No  No  No  AR and MA terms  [0, 5]  
Random walk  No  No  No  /  /  
Seasonal ARIMAX  Yes  Yes  No 

[0, 5]  
Vector Autoregression  Yes  No  No  AR term  [0, 5]  

Both  No  Yes  /  /  
Multiple Linear Regression 
Both  No  No  /  /  

Both  No  No 

[1, 2]  

Both  No  Yes  /  /  
K Nearest Neighbors Regression  Both  No  Yes 



Long Short Term Memory RNN  Both  No  Yes  Number of hidden neurons 
[1, 10]  
Random Forests  Both  No  Yes  /  /  
Simple Multilayer Perceptron 
Both  No  Yes  Number of hidden neurons  [1, 10]  
Support Vector Regression  Both  No  Yes 


It is important to comment on the influence of the type of technique on the data preprocessing aspect of the analyses. Firstly, we normalized all variables to a range between 0 and 1 for all of the analyses in this paper. This step was especially necessary for techniques such as neural networks, as literature reports this as a general practice because they benefit greatly from this step (Sola and Sevilla, 1997). Furthermore, business users can derive insights on the relative importance of variables if the forecasting technique is transparent, in order to identify the most important drivers of their sales. Secondly, the time series that are part of the analyses all display a certain trend and seasonality, which should be incorporated into the forecasting model if possible. The time series analysis techniques that we consider in this paper, explicitly include this seasonality in their model building by, for example, defining seasonal parameters. However, other types of techniques, such as regression models or neural networks, do not have this ability, which can lead to worse forecasts if the trend and season have a strong influence on the sales (Zhang and Qi, 2005). We therefore add two additional data preprocessing steps for this type of models: trend/seasonal differencing and seasonal dummy variables. In the first step, we check whether the time series actually contains either a trend or a season by means of appropriate unit root tests, such i.e. the Augmented DickeyFuller test (Dickey and Fuller, 1979) and the OsbornChuiSmithBirchenhall test respectively (Osborn et al., 1988). If the results thereof show signs of either characteristic, we apply the corresponding differencing. Secondly, if the time series is seasonal, we also include a set of seasonal dummy variables to further model the possible seasonal effects. These variables are not included in the feature selection procedure, but are always included if there is a seasonal component in the time series. Thirdly, when techniques can be used both with and without variables, past sales values need to be encoded as independent variables. We therefore need to determine how many past values will be included into the model. This hyper parameter is selected on the same validation set as the other hyper parameters, and has possible values ranging from one month to seven months. Furthermore, we define methods with external factors as techniques that use both past sales data and external parameters as independent variables. In this case, the number of past months to use as input to the model, is therefore again a hyper parameter.
Finally, we note that the list of forecasting techniques is not exhaustive. Two types of methods are notably underrepresented: ensemble methodologies and deep learning methods. We opted to include only one technique of each category in order to keep the scope of the paper manageable, i.e. Random Forests and LongShort Term Memory Neural Networks respectively. However, the obvious next step of this research is to take a closer look at these types of methodologies.
3.5 Evaluation
Evaluation for forecasting benchmarks is often entirely based on accuracy metrics. There has been a lot of discussion in the past about which metric gives the best overview of performance when comparing techniques, as many commonly used measures can exhibit strange behavior (Hyndman and Koehler, 2006; Kolassa, 2016; Tashman, 2000). In this paper, we therefore propose a combination of frequently used accuracy metrics and the expected profit function that was defined above, to select the bestperforming models. In the first category, we take into account the Mean Absolute Percentage Error (MAPE) and the Root Mean Squared Error (RMSE), as defined in Equations 6 and 7. Furthermore, we include the seasonal version of the Mean Absolute Scaled Error, which was first defined in (Hyndman and Koehler, 2006), based on the seasonality of the time series data. The formula for this metric can be found in Equation 8 with as the seasonality of the time series. This metric compares a technique’s performance to the insample error of a seasonal naïve model, which makes it perfect for truly benchmarking techniques. Next to the expected profit function, we also consider the computation time of each forecast as an approximate of the model complexity. We therefore include a total of five quantitative performance metrics in our analysis.
(6) 
(7) 
(8) 
4 Results
The result section of this paper will firstly take a look at the experimental results, which are based on forecasting the 35 datasets with 17 different forecasting techniques. Secondly, we will discuss the implications of these results, while we also comment on the limitations of this study.
4.1 Experimental results
Model  MAPE  RMSE  MASE  Profit  Time 
Without external factors  
ARMAGARCH (GARCH)  12.16 (0.00)  12.14 (0.00)  12.16 (0.00)  12.58 (0.00)  19.68 (0.00) 
Conditional Inference Regression Tree (CtreeUni)  12.93 (0.00)  12.92 (0.00)  12.88 (0.00)  12.79 (0.00)  8.16 (0.00) 
HoltWinters exponential smoothing (HW)  10.52 (1.00)  10.52 (1.00)  10.54 (1.00)  10.74 (0.00)  5.61 (0.00) 
K Nearest Neighbors Regression (KNNUni)  13.10 (0.00)  13.10 (0.00)  13.12 (0.00)  12.95 (0.00)  17.35 (0.00) 
Long Short Term Memory RNN (LSTMUni)  17.23 (0.00)  17.26 (0.00)  17.27 (0.00)  16.78 (0.00)  25.05 (0.00) 
Multiple Linear Regression (LRUni)  13.49 (0.00)  13.49 (0.00)  13.46 (0.00)  13.93 (0.00)  3.50 (0.61) 
Multivariate Adaptive Regression Splines (MARSUni)  13.48 (0.00)  13.48 (0.00)  13.46 (0.00)  13.70 (0.00)  6.68 (0.00) 
Random Forests (RFUni)  14.07 (0.00)  14.07 (0.00)  14.05 (0.00)  13.91 (0.00)  12.16 (0.00) 
Random walk (RW)  18.43 (0.00)  18.41 (0.00)  18.42 (0.00)  18.11 (0.00)  2.52 (/) 
Recursive Partitioning Regression Tree (RpartUni)  13.58 (0.00)  13.59 (0.00)  13.57 (0.00)  13.45 (0.00)  6.15 (0.00) 
Seasonal ARIMA (SARIMA)  10.45 (/)  10.47 (/)  10.48 (/)  10.70 (/)  18.40 (0.00) 
Seasonal decomposition by Loess model (DM)  11.80 (0.07)  11.80 (0.08)  11.81 (0.07)  12.12 (0.03)  9.67 (0.00) 
Seasonal random walk (SRW)  12.40 (0.00)  12.43 (0.00)  12.45 (0.00)  13.14 (0.00)  2.62 (1.00) 
Simple Multilayer Perceptron (MLPUni)  12.52 (0.00)  12.61 (0.00)  12.61 (0.00)  12.59 (0.00)  21.95 (0.00) 
Support Vector Regression (SVRUni)  12.29 (0.00)  12.29 (0.00)  12.30 (0.00)  12.45 (0.00)  17.55 (0.00) 
With external factors  
Conditional Inference Regression Tree (CtreeMulti)  12.93 (0.00)  12.93 (0.00)  12.88 (0.00)  12.81 (0.00)  9.76 (0.00) 
K Nearest Neighbors Regression (KNNMulti)  15.07 (0.00)  15.07 (0.00)  15.10 (0.00)  14.90 (0.00)  18.53 (0.00) 
Long Short Term Memory RNN (LSTMMulti)  17.05 (0.00)  16.93 (0.00)  16.95 (0.00)  16.78 (0.00)  25.02 (0.00) 
Multiple Linear Regression (LRMulti)  12.89 (0.00)  12.88 (0.00)  12.86 (0.00)  12.77 (0.00)  6.05 (0.00) 
Multivariate Adaptive Regression Splines (MARSMulti)  13.25 (0.00)  13.25 (0.00)  13.24 (0.00)  13.22 (0.00)  11.87 (0.00) 
Random Forests (RFMulti)  14.14 (0.00)  14.12 (0.00)  14.10 (0.00)  13.89 (0.00)  15.61 (0.00) 
Recursive Partitioning Regression Tree (RpartMulti)  14.28 (0.00)  14.31 (0.00)  14.29 (0.00)  13.98 (0.00)  8.70 (0.00) 
Seasonal ARIMAX (SARIMAX)  10.81 (1.00)  10.80 (1.00)  10.82 (1.00)  11.16 (0.00)  20.78 (0.00) 
Simple Multilayer Perceptron (MLPMulti)  14.13 (0.00)  14.11 (0.00)  14.12 (0.00)  13.45 (0.00)  24.20 (0.00) 
Support Vector Regression (SVRMulti)  14.04 (0.00)  14.01 (0.00)  14.04 (0.00)  14.06 (0.00)  19.13 (0.00) 
Vector Autoregression (VAR)  13.97 (0.00)  13.99 (0.00)  14.00 (0.00)  14.05 (0.00)  14.31 (0.00) 
Friedman test  
ChiSquared  1299.8  1285.8  1286.7  1059.3  18708 
Pvalue  <2.2e16  <2.2e16  <2.2e16  <2.2e16  <2.2e16 
The results of the experiments are based on a total of 21840 forecasts, as we performed 24 onemonthahead forecasts on 35 time series with 26 different models. We only take into account the results for models that have completed both the parameter tuning and feature selection procedures that were explained in the methodology section of this paper, see Section 3. Other model setups were disregarded for the final analyses.
In order to compare all of these models to one another, we apply two ranking tests for the 26 forecasting techniques according to five evaluation measures: MAPE, RMSE, MASE, Profit and computation time. Concretely, we rank all of the methods for each of the 840 unique forecasts and then display the average over these forecasts. This methodology ensures a fairer comparison between the techniques than, e.g., simply taking an average MAPE of the 840 forecasts. Furthermore, we can verify if the differences in rank are significantly separate from one another. The Friedman test (Friedman, 1940) is a nonparametric statistical test that verifies whether the difference between two treatments is significant or not. In this benchmark, the 26 forecasting techniques constitute the ’treatments’, defined as in Equation 9, while the 35 time series datasets are the ’blocks’, in Equation 9, which form groups of similar units. The Friedman test will rank the treatments according to a given evaluation criterion and will compare this ranking for each block. Therefore, three different average rankings are made for these experiments, according to the three evaluation measures. The pvalue of the Friedman test then indicates if there exists a significant difference between any of the treatments.
(9) 
If this test is significant, a posthoc analysis must follow, as we are interested to know which techniques differ from one another. As all of the Friedman tests are indeed significant, we turn to a second step, which consists of a pairwise Nemenyi test (Nemenyi, 1962) for the three rankings. Concretely, the test determines if the average ranks of the models are at least at a critical distance of:
(10) 
with as critical values, which consist of the Studentized range statistic divided by .
The results of these tests can be found in Table 5. The first column contains the method’s name, together with its abbreviation in all following figures. This table then displays the average rank of all of the forecasting methods according to MAPE, RMSE, MASE, expected profit and computation time. The numbers between brackets are the pvalues from the pairwise Nemenyi test that compares the given method to the best technique according to the evaluation metric at hand, which is indicated in bold for each evaluation measure. All pvalues with a significance level are underlined. Figure 2 displays the relative rankings according to each evaluation metric from best at the top to worst on the bottom. The grey boxes contain the models that are not significantly different from the best model at a significance level. The lines connect the relative rankings of the same technique according to the different measures.
We can now turn back to our research questions, which are firstly focused on selecting the best type of technique for a tactical sales forecast. In our introduction, we posed three sub research questions that each require the techniques to be divided into two categories. Firstly, the techniques can be divided into techniques with and without external features. Furthermore, we contrast seasonal and nonseasonal methods, and distinguish between machine learning and statistical techniques. In order to ascertain whether one category significantly outperforms the other, we perform a Wilcoxon signed rank test on the average ranks of each category for each forecast. The null hypothesis of this test states that there is no difference between the two groups. For each of the categorizations, this test rejected the null hypothesis and proved to be significant at the
significance level. Firstly, univariate techniques without variables outperform the ones that add external factors with a pvalue smaller than 2.2e16. Seasonal models have a significantly lower average rank, with a pvalue that is also smaller than 2.2e16. Finally, the ML techniques consistently lead to worse results than the statistical models, as the Wilcoxon test was highly significant with a pvalue of 0.00235. In short, time series models that explicitly model seasonality and do not incorporate external factors still seem to be the best candidate for our set of tactical sales forecasts, which all display a trend and seasonality. In order to select the top performing techniques, we take Table 5 and Figure 3 into consideration again, which show a clear top four models in terms of both accuracy and profit. The bestperforming forecasting techniques are SARIMA, SARIMAX, HoltWinters and the seasonal decomposition model (DM). The only exception to the seasonal, univariate and nonML rule is SARIMAX, which also incorporates external drivers into the model.In Figures 3, we take a closer look at these bestperforming models in terms of both accuracy and the expected profit. These figures contain the distributions of the pairwise differences of SARIMA, SARIMAX, HoltWinters and the seasonal decomposition model (DM). Grey boxplots indicate a significant difference between the two models that are mentioned on the Yaxis. DM seems to consistently perform worse than the other three models, while the remaining three time series models all perform equally in terms of both MAPE and expected profit.
The last performance metric that can still make a difference in the selection of the bestperforming technique, is computation time. This measure is indicative of the complexity of the model, but can also have an effect on the final costs of the model as computation efforts also lead to additional expenses. The average computation time of the top four models is summarized in Table 6 below. Clearly, the training of HoltWinters and DM take the least amount of time by far. However, the average computation times of both SARIMA and SARIMAX are still below 10 seconds per forecast. Furthermore, it is logical that these last two techniques require more time to train, given the feature and hyper parameter optimization according to profit for both of them. In conclusion, a top three of equally performing time series models remains: HoltWinters exponential smoothing, Seasonal ARIMA and Seasonal ARIMAX, but HoltWinters will significantly save on computation time if there is a large number of time series to forecast.
Model  Average computation time (seconds) 

HoltWinters  0.02 
Seasonal decomposition model  0.05 
SARIMA  3.89 
SARIMAX  9.35 
Finally, we also take a closer look at the interpretability of these top three techniques. As time series models, they are all transparent methodologies that attribute weights to the autoregressive, trend and seasonal components of the time series. Additionally, SARIMAX displays the weights of the added external factors, indicating their impact on the sales, which greatly adds to the explanatory power of the model. This therefore entails a large advantage for the SARIMAX technique in terms of business value. On the other hand, the feature selection procedure leads to a higher computation time and effort, so these two aspects need to be weighed against one another. In the end, the univariate time series models perform equally to SARIMAX, but additional information on the external influences on the sales might be preferable in a business context. Note that this refers to two completely different objectives, i.e. predicting versus explaining. Before the final selection of the best technique, businesses need to clearly outline the objective of a forecasting model. In terms of variable selection in this paper, Figure 4 shows the average percentage of selected variables for each of the variable types, illustrated for each of the two data sources. From these charts, we can conclude that weather and macroeconomic variables are selected the most for all datasets. On average, 2 weather variables and 2.5 macroeconomic variables were selected for the CocaCola Company datasets, while 1.78 weather variables and 3.89 macroeconomic variables were chosen for the public datasets.
The second research question focused on the integration of the expected profit function into the model selection process. We can clearly see from Table 5 and especially Figure 2 that the ranking of the techniques according to MAPE, RMSE and MASE are virtually the same, while the ranking according to the expected profit function looks a bit different. Although the top methods perform well according to all of these evaluation measures, the changes in the ranking already indicate that it is valuable to compare models according to profit as well, as it might lead to a different ranking of the possible techniques. For example, the pvalues in Table 5 of the DM technique are not significantly different from the top three time series models in terms of the accuracy measures, but they are significantly different from them when we look at the expected profit function. In Figure 5, we will look at some pairwise differences of other models according to MAPE and Profit as well. In this figure, we can clearly see that techniques can significantly differ in terms of Profit and not in terms of MAPE, or vice versa. Specifically, we compare the univariate cases of Multiple Linear Regression (LRUni) and Support Vector Regression (SVRUni), and the variant with external factors of Simple Multilayer Perceptron (MLPMulti). The pairwise differences between SVR and LR, and SVR and MLP show that there is a clear difference between the two evaluation measures. It is also important to note that these changes do not exist in pairwise differences when we only compare the accuracy metrics.
4.2 Discussion
The results of this study have three interesting implications for model selection in sales forecasting from a business perspective. Firstly, we proposed a profitdriven approach that provides a completely automated framework for model building and selection. The expected profit function that we implement, is completely adaptable to any sales forecasting situation by combining business expertise with traditional accuracybased evaluation. Furthermore, this profit function can be used as an evaluation criterion that gives a different view on which technique is truly the best one in a benchmarking exercise. While the results in this paper are consistently in line with the accuracy measures, the overall ranking according to profit is still significantly different than the accuracybased ones. This indicates that a ranking according to profit might yield a different result in model selection. In this paper, however, the top three models’ performance was consistenly very close to one another, while the same models outperformed others by a significant margin. In other cases, when model performance between techniques is closer to one another, the expected profit function can provide an additional perspective into final model selection. Furthermore, this paper adds to the scarce literature on the use of profitdriven analytics in forecasting and regression analysis.
Secondly, we notice that univariate time series models that explicitly capture seasonality, perform the best in this benchmarking study, although the Seasonal ARIMAX method is an exception to the univariate characteristic. However, this technique only performs equally to the aforementioned univariate methods and we can therefore raise the question if the addition of external variables is truly useful in this context. While other studies have shown the value of adding external drivers into the models for sales forecasting on a strategic level (Sagaert et al., 2017), this research shows that we can forecast as well without any independent variables in the model. When we take into account the additional cost of data collection and model maintenance, we conclude that forecasting the sales on a product category level is easier achieved by univariate models without compromising on accuracy or profit. Although we recognize the added explanatory value of integrating features, we question if it is worth the effort when achieving the best forecast is the goal.
Finally, we compared two categories of forecasting techniques to one another: statistical methods and machine learning techniques. In the case of tactical sales forecasting, we clearly see that simpler models outperform the others significantly for these 35 datasets. This leads us to conclude that the more traditional models are actually still performing the best when tackling this type of time series problem. These findings are in line with (Makridakis and Hibon, 2000), but contradict (Crone et al., 2011). To conclude, seasonal time series models tend to outperform other techniques for a tactical sales forecast. From a business perspective, this conclusion is especially positive, as these models are easy to interpret and have a faster computation time.
5 Conclusion
In this paper, we introduced a new, completely automated and profitoriented strategy to sales forecasting, which integrates an expected profit function into several steps of model selection. This function can be implemented in any sales forecasting context by letting business experts and previous data set the profit margins for every product. Furthermore, our research has proven that simpler time series models tend to outperform more complex techniques for 35 sales datasets. All of the applied ML techniques achieve significantly worse results than the traditional models, both in accuracy and profit. This implies that less complex techniques are still the best type of method to handle tactical sales forecasting. Finally, we found that univariate time series models that are able to explicitly model the seasonality of a time series, perform best. This indicates that the addition of external variables is unnecessary, especially when we consider the additional costs that are linked to maintaining models with external drivers.
In terms of possible limitations of this study, we recognize some shortcomings in this paper. Firstly, it is impossible to come up with an exhaustive list of forecasting techniques. However, we have attempted to implement common methods from all three categories of techniques that are frequently used for forecasting. Furthermore, this research consists of 35 monthly time series, which is significantly less than the larger benchmarks and competitions in the field (Athanasopoulos et al., 2011; Crone et al., 2011; Makridakis and Hibon, 2000). However, this paper particularly focuses on one field, i.e. sales forecasting, and is one of the larger benchmarking studies in this specific area. Furthermore, we have added to the generalizability and reproducibility of the study by including several publicly available datasets as well. Finally, we only implemented individual forecasting methods without considering ensemble methods. This type of methodology has become extremely popular in forecasting (Lessmann et al., 2015) and it has been proven that this approach can significantly impact the accuracy of forecasts. Potential future research therefore includes an expansion of this study in three aspects. Firstly, we can include more sales time series in order to further underpin our statements. Secondly, we can implement more techniques and include ensemble methods. Finally, this study can be further expanded to other fields than sales forecasting as well. However, given the range of techniques and the number of datasets that were already used in this paper, we can state that simple, seasonal time series models are still the best choice for a highlevel tactical sales forecast.
Acknowledgements
We would like to acknowledge The CocaCola Company for funding this research and providing us with the necessary business expertise and data to conduct our experiments.
Bibliography
References
 A comparison of multivariate and univariate time series approaches to modelling and forecasting emergency department demand in western australia. Journal of biomedical informatics 57, pp. 62–73. Cited by: §2.1.
 A novel approach to model selection in tourism demand modeling. Tourism Management 48, pp. 64–72. Cited by: §2.1.
 Making progress in forecasting. International Journal of Forecasting 22 (3), pp. 433–441. Cited by: §1, §1.
 Findings from evidencebased forecasting: methods for reducing forecast error. International Journal of Forecasting 22 (3), pp. 583–598. Cited by: §1.

A hybrid seasonal autoregressive integrated moving average and quantile regression for daily food sales forecasting
. International Journal of Production Economics 170, pp. 321–335. Cited by: §2.1.  The tourism forecasting competition. International Journal of Forecasting 27 (3), pp. 822–844. Cited by: §2.1, §2.1, §3.3, §5.
 Tuning data mining methods for costsensitive regression: a study in loan chargeoff forecasting. Journal of Management Information Systems 25 (3), pp. 315–336. Cited by: §2.2.
 Assessing and hedging the cost of unseasonal weather: case of the apparel sector. European Journal of Operational Research 244 (1), pp. 261–276. Cited by: §3.1.
 Are more data always better for factor analysis?. Journal of Econometrics 132 (1), pp. 169–194. Cited by: §3.3.
 Forecasting the value effect of seasoned equity offering announcements. European Journal of Operational Research 214 (2), pp. 418–427. Cited by: §2.1, §2.2.
 A combination selection algorithm on forecasting. European Journal of Operational Research 234 (1), pp. 127–139. Cited by: §1, §2.1.

A nested heuristic for parameter tuning in support vector machines
. Computers & Operations Research 43, pp. 328–334. Cited by: §3.4.  Advances in forecasting with neural networks? empirical evidence from the nn3 competition on time series prediction. International Journal of Forecasting 27 (3), pp. 635–660. Cited by: §1, §1, §1, §2.1, §2.1, §4.2, §5.
 Utility based data mining for time series analysis: costsensitive learning for neural network predictors. In Proceedings of the 1st international workshop on Utilitybased data mining, pp. 59–68. Cited by: §2.2.
 Consumer behaviour and sales forecast accuracy: what’s going on and how should revenue managers respond?. Journal of Revenue and Pricing Management 9 (4), pp. 374–376. Cited by: §3.1.
 Distribution of the estimators for autoregressive time series with a unit root. Journal of the American statistical association 74 (366a), pp. 427–431. Cited by: §3.4.
 A review of datasets and load forecasting techniques for smart natural gas and water grids: analysis and experiments. Neurocomputing 170, pp. 448–465. Cited by: §2.1.
 The forecasting journals and their contribution to forecasting research: citation analysis and expert opinion. International Journal of forecasting 22 (3), pp. 415–432. Cited by: §1.
 The forecasting performance of various models for seasonality and nonlinearity for quarterly industrial production. International Journal of Forecasting 21 (1), pp. 87–102. Cited by: §2.1.
 A comparison of alternative tests of significance for the problem of m rankings. The Annals of Mathematical Statistics 11 (1), pp. 86–92. Cited by: §4.1.
 Tourism in the canary islands: forecasting using several seasonal time series models. Journal of Forecasting 27 (7), pp. 621–636. Cited by: §2.1.
 Forecasting international city tourism demand for paris: accuracy of uniand multivariate models employing monthly data. Tourism Management 46, pp. 123–135. Cited by: §2.1.
 The value of competitive information in forecasting fmcg retail product sales and the variable selection problem. European Journal of Operational Research 237 (2), pp. 738–748. Cited by: §3.1.
 Another look at measures of forecast accuracy. International journal of forecasting 22 (4), pp. 679–688. Cited by: §3.5.
 How to measure the impact of a forecast error on an enterprise?. The Journal of Business Forecasting 22 (1), pp. 21. Cited by: §1, §3.2.
 Evaluating predictive count data distributions in retail sales forecasting. International Journal of Forecasting 32 (3), pp. 788–803. Cited by: §3.5.
 Benchmarking stateoftheart classification algorithms for credit scoring: an update of research. European Journal of Operational Research 247 (1), pp. 124–136. Cited by: §2.1, §5.
 Car resale price forecasting: the impact of regression method, private information, and heterogeneity on forecast accuracy. International Journal of Forecasting 33 (4), pp. 864–877. Cited by: §1.

Demand forecasting with high dimensional data: the case of sku retail sales forecasting with intraand intercategory promotional information
. European Journal of Operational Research 249 (1), pp. 245–257. Cited by: §2.1, §3.1.  The m3competition: results, conclusions and implications. International journal of forecasting 16 (4), pp. 451–476. Cited by: §1, §2.1, §2.1, §4.2, §5.
 Distributionfree multiple comparisons. In Biometrics, Vol. 18, pp. 263. Cited by: §4.1.
 Forecasting with cue information: a comparison of multiple regression with alternative forecasting approaches. European Journal of Operational Research 180 (1), pp. 354–368. Cited by: §2.1.
 Seasonality and the order of integration for consumption. Oxford Bulletin of Economics and Statistics 50 (4), pp. 361–377. Cited by: §3.4.
 Social network analytics for churn prediction in telco: model building, evaluation and network architecture. Expert Systems with Applications 85, pp. 204–220. Cited by: §2.2.
 Feature selection based on mutual information criteria of maxdependency, maxrelevance, and minredundancy. IEEE Transactions on pattern analysis and machine intelligence 27 (8), pp. 1226–1238. Cited by: §3.3.
 ‘Horses for courses’ in demand forecasting. European Journal of Operational Research 237 (1), pp. 152–163. Cited by: §1, §1, §2.1.
 Performance of state space and arima models for consumer retail sales forecasting. Robotics and computerintegrated manufacturing 34, pp. 151–163. Cited by: §1, §2.1.
 Tactical sales forecasting using a very large set of macroeconomic indicators. European Journal of Operational Research. Cited by: §3.1, §4.2.
 Comparing univariate and multivariate models to forecast portfolio valueatrisk. Journal of financial econometrics 11 (2), pp. 400–441. Cited by: §2.1.
 Importance of input data normalization for the application of neural networks to complex industrial problems. IEEE Transactions on Nuclear Science 44 (3), pp. 1464–1468. Cited by: §3.4.

Profit maximizing logistic regression modeling for customer churn prediction
. In Data Science and Advanced Analytics (DSAA), 2015. 36678 2015. IEEE International Conference on, pp. 1–10. Cited by: §2.2.  Supply chain forecasting: theory, practice, their gap and the future. European Journal of Operational Research 252 (1), pp. 1–26. Cited by: §3.1.
 Outofsample tests of forecasting accuracy: an analysis and review. International journal of forecasting 16 (4), pp. 437–450. Cited by: §3.5.
 A comparison of univariate methods for forecasting electricity demand up to a day ahead. International Journal of Forecasting 22 (1), pp. 1–16. Cited by: §2.1.
 ProfARIMA: a profitdriven order identification algorithm for arima models in sales forecasting. Applied Soft Computing. Cited by: §1, §3.2, §3.2.
 Profit driven business analytics: a practitioner’s guide to transforming big data into added value. John Wiley & Sons. Cited by: §2.2.
 New insights into churn prediction in the telecommunication sector: a profit driven data mining approach. European Journal of Operational Research 218 (1), pp. 211–229. Cited by: §2.2.
 A novel profit maximizing metric for measuring classification performance of customer churn prediction models. IEEE Transactions on Knowledge and Data Engineering 25 (5), pp. 961–973. Cited by: §2.2.
 Supply chain forecasting: best practices & benchmarking study. Technical Paper. Lancaster Centre for Forecasting, pp. 1–42. Cited by: §2.1.
 Nonfixed and asymmetrical margin approach to stock market prediction using support vector regression. In Neural Information Processing, 2002. ICONIP’02. Proceedings of the 9th International Conference on, Vol. 3, pp. 1398–1402. Cited by: §2.2.
 Neural network forecasting for seasonal and trend time series. European journal of operational research 160 (2), pp. 501–514. Cited by: §3.4.
 An extended tuning method for costsensitive regression and forecasting. Decision Support Systems 51 (3), pp. 372–383. Cited by: §2.2.
Comments
There are no comments yet.