=Paper=
{{Paper
|id=Vol-2853/short6
|storemode=property
|title=Neural Network Forecasting Using Big Data
|pdfUrl=https://ceur-ws.org/Vol-2853/short6.pdf
|volume=Vol-2853
|authors=Sergey Ivanov,Nataliia Maksyshko,Mykola Ivanov
|dblpUrl=https://dblp.org/rec/conf/intelitsis/IvanovMI21
}}
==Neural Network Forecasting Using Big Data==
https://ceur-ws.org/Vol-2853/short6.pdf
Neural Network Forecasting Using Big Data
Sergey Ivanova, Nataliia Maksyshkoa and Mykola Ivanova
a
Zaporizhzhia National University, Zhukovsky str., 66, Zaporizhzhia, 69063, Ukraine
Abstract
To increase the speed of data transfer while providing access to multidimensional data there
is the use of Big Data as a tool. On the base of the MapReduce model you can use modern
tools for working with big data. Therefore, in the paper researched Big Data as a single
centralized source of information for the entire subject area. In addition, in this paper the
structure of a neural network forecasting system is proposed, which include many databases,
where transactions are processed in real time. For neural network forecasting of
multidimensional data, a network in Matlab is considered and built. A matrix of input data
and a matrix of target data, which determine the input statistical information, are used to
teach the neural network.
The application of the Levenberg-Marquardt algorithm for training a neural network is
considered. Also, the results of the training process of neural network in Matlab are
presented. The obtained forecasting results are presented, which allows to conclude about the
advantages of a neural network in multivariate forecasting.
Keywords 1
Big Data, MapReduce model, neural network, forecasting, matrix
1. Introduction
As Laurent Flores explained [1], the success of the use of information technology is determined by
the success of digital economic, thus how they are measured and used. However, attention is paid to
forecasting in the digital economy, taking into account intelligent systems.
Intelligent systems and the use of multidimensional communication determined the emergence of a
new concept by the German economist Klaus Schwab, his economic forum in Davos [2]. According
to this concept, it is argued that we are in the era of the fourth industrial revolution (Industry 4.0),
when the virtual world is combined with the physical world using information technology. The fourth
industrial revolution is characterized by a change in economic relations and the widespread use of
intelligent technologies (general technologies, big data, artificial neural networks, and others).
It should be noted that with the use of digital technologies, D2C models have come to be used. The
D2C (Direct to Consumer) model represents a direct selling system, where companies themselves
manufacture, promote, sell and deliver their product without the involvement of intermediaries. So in
the above paper (https://www.emarketer.com/content/nikes-d2c-sales-will-comprise-third-of-its-
business) it is established that, in contrast to traditional promotion through retail chains, companies
using D2C model develop their own distribution channels. Thanks to their good positioning, these
companies not only have a competitive advantage in the market, but also have their own structure on
the Internet. These companies have changed the producer-consumer relationship and are reducing the
distance between them.
Today, any customer can contact the manufacturer directly, ask their question and make a
purchase, avoiding extra charges and saving time. Renowned manufacturers have recognized the need
to develop their own D2C strategies based on analytics. The authors of the paper acknowledge that
IntelITSIS’2021: 2nd International Workshop on Intelligent Information Technologies and Systems of Information Security, March 24–26,
2021, Khmelnytskyi, Ukraine
EMAIL: flydaiver@gmail.com (S. Ivanov); maxishko@ukr.net (N. Maksyshko); nn_iva@ukr.net (M. Ivanov)
ORCID: 0000-0003-1086-0701 (S. Ivanov); 0000-0002-0473-7195 (N. Maksyshko); 0000-0002-1908-0763 (M. Ivanov)
© 2021 Copyright for this paper by its authors.
Use permitted under Creative Commons License Attribution 4.0 International (CC BY 4.0).
CEUR Workshop Proceedings (CEUR-WS.org)
the use of D2C opens up additional opportunities for companies. According to the authors, Nike is a
prime example, with D2C sales accounting for a third of total revenues by the end of 2020 based on
its Consumer Direct strategy.
Given the widespread use of digital marketing, author Rimma Katz in paper [3] explored social
commerce, which fostered development along the D2C model. Today, social commerce is used to
increase the reach of consumers, those who may know about direct contact with the manufacturer.
So in the presented report by Jasmine Enberg (Updating social networks for the 4th quarter of
2020 - https://www.emarketer.com/content/social-media-pdate -q4-2020), a summary of the main
events, their analysis for marketing management, solving problems strategic development of
companies. Report author Jasmine Enberg determines that the global forecast for monthly social
media users in 2020 has increased due to the effects of the pandemic. However, no platform will be
able to maintain the growth it picked up at the beginning of the year. Therefore, in 2021, the growth
rate will begin to normalize. Recent product launches, including Facebook live shopping and
Instagram shopping tags, show that e-commerce continues to be a priority for the two platforms.
Snapchat and Twitter focused more on effective marketing, namely the release of new sets of
promotional offers with direct consumer response and others.
Therefore, the modern companies are based on direct selling (D2C) models. That allows you to get
images of buyers and segment them. In addition to these tasks, it is necessary to solve forecasting the
market, which changes every year.
It should be noted that worldwide targeted statistics for the entire sales system. This information is
stored in cloud storage (Big Data). The information used includes data from the time of attracting a
new consumer to the required resource information about the number, including repeated ones.
In this paper, we propose a neural network analytics method to forecast the main parameters for
companies.
The innovation in the application of neural networks is the formation and application of neural
network learning matrices. The peculiarity of these matrices is that their values have a random
relationship with each other. For example, each family buys a conditional loaf of bread a day. This
sales process has a pattern and can be attributed to a deterministic process. The next day, the family
buys 0.5 white and 0.5 gray, and so on in different proportions. This sales process is random and
belongs to a stochastic process.
2. Formal problem statement
Making forecasts is part of any firm's overall analytics. An important role in the forecasting
method is paid by the multidimensionality of information and methods of their processing. The use of
Big Data with OLAP technologies requires new approaches to processing and applying large amounts
of data. This is due to the wide range of communication systems used in the novated companies.
Therefore, neural network forecasting based on Big Data is an urgent task, which is discussed in
this paper.
3. Literature review
Modern companies use many available forecasting methods, they not only improve the quality of
their products, but also get information about the needs of customers. Neural network forecasting
models are a great way to predict customer preferences and apply new ways to stand out from the
competition. Using practical forecasting models today is the best way to get the most effective and
complete data to improve decisions. In this case, forecasting methods in digital economy should
include not only customer surveys, their age, interests and price, but also the characteristics of the
product, brand, logistics and others.
In the work of B. Yonathan [4], the problem of fitting mathematical models to numerical data was
considered. Such a fit is often performed by the least squares method, regardless of previous
knowledge of parameter values or the statistical nature of measurement errors.
According to the results of a study of the opinions of experts, A. Ashton [5] considered a scenario
in which the results of the opinions of experts differ significantly from the polls of intentions. For this
case, the author considered the problem of predicting market behavior.
Further development of the principle of intentions was found in the work of V. Morwitz [6]. The
author proposed the principles of using intentions when solving the forecasting problem. In this paper
is included research people behavior and how they would solve problems in different situations. For
this, the method of polling the intentions of people was used. Intent surveys are widely used in
marketing when sales data is unknown, for example, to forecast new products.
A continuation of these works found themselves in the work of J. S. Armstrong [7], where the role
of a person as a dominant factor was considered. This task was solved as a role-playing game for
making predictions of the behavior of people who interact with others. A key tenet of this approach is
to provide realistic simulation of interactions.
This forecasting method is currently rarely used. G. Rowe and G.Wright [8] considered the
application of the Delphi method as a procedure. The authors found that the accuracy of expert
predictions can be improved through the use of Delphi structured methods. One of the principles of
the method is that experts' forecasts should not depend on each other. Expert groups sometimes
violate this principle; as a result, the data should not be used in forecasting.
D. Wittink and T. Bergestuen [9] considered the problem about “Intentions” in their work. This
paper examines the intention as an indicator of a consumer to purchase a product under the influence
of various factors. The consumer can declare his intentions to make a purchase of various goods. This
method is based on the following principles, namely, using a new design to create an acceptable
situation.
The formation of a digital marketing strategy was considered in the work of P. Mandal and N.
Joshi [10]. The authors of the paper emphasize that digital technologies make marketing more
effective, since it allows to identify individual consumer interests and how to make better manage
campaigns and improve the product. In this paper, the authors propose a flowchart for developing
marketing strategies.
M. Ivanov [11] discusses the analysis of the state of the digital economy and digital marketing in
his paper. The author in the paper shows that the dynamics of processes in the economy is quite high
and requires a quick analysis of multidimensional data. The author proposes a conceptual model and a
method for assessing consumer demand in the target market, aimed at the prospective management of
trading floors using Big Data.
4. The purpose of the paper
The paper is devoted to neural network forecasting based on the use of neural networks. The
information base of neural network forecasting is Big Data. The process of constructing training
matrices, training a neural network and making the predictions are presented in the result of paper.
5. Neural network forecasting modeling
The modern economy is characterized by the rapid dynamics of economic processes. Under these
conditions, forecasting models of neural networks acquire new significance in the process of making
managerial decisions. The process of the importance of making management decisions in digital
marketing systems is shown in Figure 1.
In this Fig. 1 in the field of traditional marketing analytics (MTA), which is based on the use of
classical approaches and forecasting methods. MTA is aimed at solving economic forecasting
problems, where the time for processing and obtaining forecast results takes a long time. This time for
solving forecasting problems can range from several days to a month or a year, which is
incommensurate with the dynamic processes in the digital economy. Therefore, the forecast results
become outdated, lose their relevance and can lead to negative results.
Therefore, decisions related to forecasting processes for a short period of time are an urgent task.
The scope of this problem is depicted in Figure 1 as Marketing Actual Analytics (MRA). Marketing
Actual Analytics (MRA) and is aimed at solving the problem of predicting fast processes, namely
from the moment information appears up to several hours.
Today the number of information sources of data in the world is growing rapidly.
Figure 1: The importance of management decision making in digital marketing
Therefore, storage technologies and their processing of information are becoming more and more
in demand. By storing information, one can single out the use of Big Data for which the basic
principles of work can be formulated:
1. Horizontal scalability, which takes into account that the data can be arbitrarily large from any
system. They have the ability to handle big data.
2. Tolerance to failures, which use the principles of horizontal scalability and apply methods of
clustering systems; locality of data, which allows in large distributed systems to separate data from a
large number of data centers. All modern tools for working with big data, one way or another, follow
these three principles. The first principle is based on the MapReduce model. The MapReduce model
provides for distributed data processing proposed by Google and is shown in Fig. 2.
Figure 2: Data processing according to the MapReduce model
MapReduce provides that data is organized as relational or multidimensional data (OLAP). The
data processing method takes place in three stages. The first stage is aimed at executing the Map ()
function. At this stage, the data is preprocessed using the Map () function, which is defined by the
user. The work of this stage is to preprocess and filter the data.
The second stage of the model is performed by the Shuffle () function. This stage goes unnoticed
by the user. At this stage, the Map () function performs the data immersion procedure similarly to the
formation of data marts (Data Mart), that is, one Map () data output corresponds to each mart. In the
future, these showcases will serve as an input for the Reduce () function.
The third stage of the model is aimed at executing the Reduce () function. Each data mart, which is
formed in the second stage, transfers information to the input of the Reduce () function. The Reduce ()
function is user defined and calculates the result for individual storefronts. The set of all values
returned by Reduce () is the result in this method. Therefore, Big Data technology is consider as a tool
that allows you to increase the speed of data transfer while providing a large capacity of information
carriers.
In addition, this technology can improve the availability of cloud applications and data services.
Thus, digital marketing is shape around the mainstream e-commerce models. The interconnection of
the main models (B2B, B2A, D2C, C2A and C2C) of e-commerce systems based on the systems for
collecting, storing and analyzing information in real time, which based on subsequent storage in
historical data layers.
For the implementation of systems that perform Marketing Relevant Analytical tasks using data,
OLAP data systems are used, which are structured according to the principle of multidimensional
information presentation [11]. Reducing the cost of creating multidimensional warehouses can be
achieve by using Data Mart. A data mart can only contain thematically aggregated data. Big Data is
today a single, centralized source of information for the entire subject area. The structure of the
marketing analytical system can be represent as follows (Fig. 3).
In a marketing analytical system, there are many databases. where transaction processing is done
in real time. Therefore, online data source systems (ODS) provide information for processing in
OLTP. OLTP systems provide storage and processing of information in real time. The processed data
in OLTP is transferred to the Data Mart systems with the subsequent construction of multidimensional
OLAP data cubes.
This multidimensional data is aimed at presenting information on thematic sections both on
marketing information and other information from different areas of the economy.
Figure 3: The structure of the marketing analytical system
The marketer has the ability to access multidimensional data in the repository, as well as complete
economic information for conducting an MRA. The advantages of this approach are:
• simplicity of creating and filling OLAP, since filling comes from reliable sources of data marts;
• reducing the load on working with multidimensional data, namely, one multidimensional query
processes multiple OLAP layers.
The data coming from the OID is transferred to the OLTP and the data marts are moved. OLAP
stores data as multidimensional layers of measures and dimensions [11].
For neural network forecasting, a multidimensional query is formed to multidimensional data,
which allows obtaining the following information input stream (Inflow) - formed by data from OLTP
and DMN subsystems:
𝐼𝐼 = �𝑝𝑝𝑗𝑗 = �𝑔𝑔𝑗𝑗 , 𝑖𝑖𝑖𝑖𝑗𝑗 , 𝑚𝑚𝑚𝑚𝑗𝑗 ��, 𝑗𝑗 = �����
1, 𝑁𝑁 , (1)
where 𝑔𝑔𝑗𝑗 is the product included in the analyzed sets of 𝑁𝑁 — the object of research;
𝑖𝑖𝑖𝑖𝑗𝑗 — indicator of income 𝑗𝑗 of the product;
𝑚𝑚𝑚𝑚𝑗𝑗 — product indicators 𝑗𝑗.
Datasets from set 𝐼𝐼 stored in OLAP and on demand allow you to obtain and conduct analysis with
the subsequent storage of data, which are called transactions. Description of a transaction to set 𝐼𝐼 as
follows:
𝑇𝑇 = �𝑖𝑖𝑖𝑖𝑗𝑗 �𝑖𝑖𝑖𝑖𝑁𝑁 ∈ 𝐼𝐼�. (2)
Such transactions for retail outlets on the Internet correspond to the nomenclature of goods that the
consumer buys and the data stored in OLAP as multidimensional data cubes (OLAP).
Then solving the problem of neural network forecasting, data arrays of its training are formed. The
forecasting technique using a neural network is formalized through the problem of pattern
recognition. Data on the predicted economic indicators of a product for a certain period of time form
an image, the class of which is determined by the values of the predicted indicators.
In the proposed methodology, the dimension of the multidimensional array will determine both the
forecasting interval and the number of predicted indicators. Each next line of the array is formed as a
result of a shift by one interval equal to the prediction interval.
The neural network is trained on the generated training array of product indicators and adjusts its
weights accordingly. As a result, the neural network is trained to solve the forecasting problem for a
certain forecasting horizon. It should be noted that two forecasting approaches are used: one-step and
multi-step. One-step forecasting is used for short-term forecasts and multistep forecasting is used for
long-term forecasting.
In general, the forecasting model can be represented as follows. Let the time interval [𝑡𝑡0 , 𝑡𝑡𝑘𝑘 ] be
given, the indicators 𝑔𝑔𝑗𝑗 , 𝑖𝑖𝑖𝑖𝑗𝑗 , 𝑝𝑝𝑝𝑝𝑗𝑗 of the product are defined, where 𝑡𝑡0 is the initial time value, 𝑡𝑡𝑘𝑘 is the
current time value. To find the predicted values on the prediction interval ∆, a method is proposed that
includes the following stages:
1. Analytical analysis of indicators and the formation of a learning matrix from selected values
from historical slices of multidimensional databases (OLAP technology).
A learning matrix (ML - matrix learning) can be written as input data for representation in a neural
network:
𝑖𝑖𝑖𝑖01 = 𝑓𝑓01 (𝑡𝑡0 ) 𝑖𝑖𝑖𝑖02 = 𝑓𝑓02 (𝑡𝑡0 + ∆) ⋯ 𝑖𝑖𝑖𝑖0𝑚𝑚 = 𝑓𝑓0𝑚𝑚 (𝑡𝑡0 + (𝑚𝑚 − 1)∆)
𝑀𝑀𝐿𝐿𝐼𝐼𝐼𝐼𝐼𝐼𝐼𝐼𝐼𝐼 = � 𝑖𝑖𝑖𝑖11 =⋯𝑓𝑓11 (𝑡𝑡0 ) 𝑖𝑖𝑖𝑖12 = 𝑓𝑓⋯ 12 (𝑡𝑡0 + ∆) ⋯ 𝑖𝑖𝑖𝑖1𝑚𝑚 = 𝑓𝑓1𝑚𝑚 (𝑡𝑡0 + (𝑚𝑚 − 1)∆) � , 𝑚𝑚 = �����
1, 𝑘𝑘 , (3)
⋯ ⋯
𝑖𝑖𝑖𝑖𝑁𝑁1 = 𝑓𝑓𝑁𝑁1 (𝑡𝑡0 ) 𝑖𝑖𝑖𝑖𝑁𝑁2 = 𝑓𝑓𝑁𝑁2 (𝑡𝑡0 + ∆) ⋯ 𝑖𝑖𝑖𝑖𝑁𝑁𝑁𝑁 = 𝑓𝑓𝑁𝑁𝑁𝑁 (𝑡𝑡0 + (𝑚𝑚 − 1)∆)
where ∆ — is the horizon (time interval) of forecasting.
The input data in the neural network in Matlab is presented in the form of a 5x10 matrix, which
represents static information on 10 products for five days of the week and has the following form
(Fig. 4).
Figure 4: “ВataInp” matrix view
Target data that determine the output statistical information for the neural output can be
represented in the form of learning value matrix (𝑇𝑇𝑇𝑇𝑇𝑇𝑂𝑂𝑂𝑂𝑂𝑂𝑂𝑂𝑂𝑂𝑂𝑂 ):
𝑖𝑖𝑖𝑖01 = 𝑓𝑓01 (𝑡𝑡0 + ∆) 𝑖𝑖𝑖𝑖02 = 𝑓𝑓02 (𝑡𝑡0 + (𝑘𝑘 + 1)∆) ⋯ 𝑖𝑖𝑖𝑖0𝑚𝑚 = 𝑓𝑓0𝑚𝑚 (𝑡𝑡0 + 𝑚𝑚∆) (4)
𝑇𝑇𝑇𝑇𝑇𝑇𝑂𝑂𝑂𝑂𝑂𝑂𝑂𝑂𝑂𝑂𝑂𝑂 = � 𝑖𝑖𝑖𝑖11 = 𝑓𝑓11 (𝑡𝑡0 + ∆) 𝑖𝑖𝑖𝑖12 = 𝑓𝑓12 (𝑡𝑡0 + (𝑘𝑘 + 1)∆) ⋯ 𝑖𝑖𝑖𝑖1𝑚𝑚 = 𝑓𝑓1𝑚𝑚 (𝑡𝑡0 + 𝑚𝑚∆) �.
⋯ ⋯ ⋯ ⋯
𝑖𝑖𝑖𝑖𝑁𝑁1 = 𝑓𝑓𝑁𝑁1 (𝑡𝑡0 + ∆) 𝑖𝑖𝑖𝑖𝑁𝑁2 = 𝑓𝑓𝑁𝑁2 (𝑡𝑡0 + (𝑘𝑘 + 1)∆) ⋯ 𝑖𝑖𝑖𝑖𝑁𝑁𝑁𝑁 = 𝑓𝑓𝑁𝑁𝑁𝑁 (𝑡𝑡0 + 𝑚𝑚∆)
The target data, which determines the output statistical information in Matlab for neural output, is
presented in Fig. 5.
Figure 5: Target Output Matrix “DataOut”
2. NS training. The process of training a neural network is to match to each 𝑀𝑀𝑀𝑀𝐼𝐼𝐼𝐼𝐼𝐼𝐼𝐼𝐼𝐼 element the
value of the 𝑇𝑇𝑇𝑇𝑇𝑇𝑂𝑂𝑂𝑂𝑂𝑂𝑂𝑂𝑂𝑂𝑂𝑂 matrix corresponding to the mapping in the value of the elements of the
weight matrix 𝑤𝑤𝑗𝑗 :
𝑤𝑤𝑗𝑗 : 𝑀𝑀𝐿𝐿𝐼𝐼𝐼𝐼𝐼𝐼𝐼𝐼𝐼𝐼 → 𝑇𝑇𝑇𝑇𝑇𝑇𝑂𝑂𝑂𝑂𝑂𝑂𝑂𝑂𝑂𝑂𝑂𝑂 . (5)
In the process of training the neural network, the task of minimizing the objective function is
solved. With this approach, an algorithm is used for training, which is the most efficient not only in
terms of errors, but also in time. The neural network in Matlab is trained using the Levenberg-
Marquardt error backpropagation algorithm. The Levenberg-Marquardt algorithm uses a scalable
conjugate gradient backpropagation.
Therefore, the training of the neural network is represented in time, and the network is tuned in
accordance with its error. The magnitude parameter is used to measure the generalization of the neural
network and stop learning when the generalization stops improving. The test score itself does not
affect training and provides an independent assessment of the performance of the neural network
during and after training.
The choice of the algorithm, as well as the learning process of the neural network, is shown in Fig.
6-7.
Figure 6: The process of learning the neural network
Figure 7: The choice of the learning algorithm
The number of neural network training epochs can be written as follows:
𝑡𝑡 −𝑡𝑡 (6)
𝑒𝑒𝑒𝑒𝑒𝑒𝑒𝑒ℎ𝑠𝑠 = � 𝑘𝑘 0 �.
∆
In the process of forecasting by a neural network, it is necessary to take into account the
forecasting horizon. In the Matlab system, the sim (net,[;;]) function is implemented, which allows
you to supply a variety of input values and get a solution at the output of a neural network. The
forecast results for the sale of 10 goods are shown in Fig. 8.
Therefore, the created neural network does indeed make multiple predictive decisions. It allows
you to solve the sales function of marketing and consider the dynamics of the sale of many products
in real time. Neural network forecasting looks at the number of future periods that the forecast will
cover. That is, you may need a forecast 7 days ahead, with data for every day. In this case, the period
is a day, and the horizon is 7 days.
Finally, the prediction interval is the frequency with which a new prediction is made. Often the
prediction interval coincides with the prediction period. The choice of the forecasting period and
horizon is usually dictated by the conditions for making decisions.
Figure 8: The result of forecasting the sale of 10 products
Choosing these two parameters is one of the hardest parts of neural network forecasting. For
forecasting to be meaningful, the forecasting horizon must be no less than the time required to
implement the decision made on the basis of the forecast.
Thus, forecasting is highly dependent on the nature of the decision being made.
6. Conclusion
In this paper, the authors consider the solution to the problem of forecasting a neural network and
the use of Big Data as a tool to increase the data transfer rate while providing access to
multidimensional data (OLAP). The paper proposes the structure of a neural network for solving the
forecasting problem, which uses the training matrices of this network. Matrix data can be built from
data that provides information in OLAP.
The paper proposes the use of a neural network for predicting multidimensional data, which is
built in the Matlab system. To solve the forecasting problem, the authors proposed the construction of
the input and target data matrices for training a neural network. The paper also presents the
construction procedures and the results of the neural network training process in Matlab.
The obtained forecasting results allow us to conclude about the advantages of a neural network for
multivariate data prediction. Multidimensional data and their level of detail are important for solving
the forecasting problem.
The wider use of digital systems makes it possible to apply the proposed approach to forecasting.
It should be noted that the use of neural networks in forecasting could have the following
disadvantages. So, in the process of forming training matrices, if there are no data in the initial
information or data presented in an indistinct form (a linguistic variable, the appearance of a new
product or there is no demand for a product), then an error may appear.
Therefore, this drawback in solving the forecasting problem can be eliminated if systems with
fuzzy sets are used, which will allow to formalize fuzzy variables.
7. References
[1] L. Florès, How to Measure Digital Marketing, Springer, London, 2014.
[2] Kl. Schwab (Ed.), The Fourth Industrial Revolution, World Economic Forum, Cologny/Geneva,
2017.
[3] R. Kats, Mucinex Eyes Social Commerce to Help Bolster D2C Business, emarketer Web, 2020.
URL: https://www.emarketer.com/content/mucinex-eyes-social-commerce-help-bolster-d2c-
business/.
[4] B. Yonathan (Ed.), Nonlinear Parameter Estimation, New York, 1974.
[5] A.H. Ashton, Does consensus imply accuracy in accounting studies of decision making?, The
Accounting Review, vol. 60, no. 2 (1985) 173–185. URL: www.jstor.org/stable/246784.
[6] V. Morwitz, Methods for forecasting from intentions data, Principles of Forecasting (2001) 33-
56. doi:10.1007/978-0-306-47630-3_3
[7] J.S. Armstrong (ed.), Role Playing: A Method to Forecast Decisions, Springer, Boston, 2001, pp.
15-30. URL: https://repository.upenn.edu/marketing_papers/152.
[8] D.R. Wittink, T. Bergestuen, Forecasting with conjoint analysis, In: Armstrong J.S. (eds),
Principles of Forecasting, International Series in Operations Research & Management Science,
volume 30, Springer, Boston, MA, 2001, pp. 147-167. doi:10.1007/978-0-306-47630-3_8
[9] G. Rowe, G. Wright, Expert opinions in forecasting role of the Delphi technique, Springer,
Boston, 2001, pp. 125-144. doi:10.1007/978-0-306-47630-3_7
[10] P. Mandal, N. Joshi, Understanding Digital Marketing Strategy, International Journal of
Scientific Research and Management 5(6) (2017) 5428-5431. doi:10.18535/ijsrm/v5i6.11
[11] M. Ivanov, Cloud-based Digital Marketing, CEUR Workshop Proceedings, 2422, 2019.