A New Product’s Demand Forecasting Using Artificial Neural
Natapat Areerakulkan
, Chanicha Moryadee
, Lamphai Trakoonsanti
Martusorn Khaengkhan
and Natpatsaya Setthachotsombut
College of Logistics and Supply Chain, Suan Sunandha Rajabhat University, Dusit, Bangkok, Thailand
Keywords: Demand Forecasting, Mobile Phone, Artificial Neural Network.
Abstract: This paper presents the means to improve new product (mobile phone) demand forecasting that led to total
cost reduction and more efficient inventory management. The selected forecast methods, namely Holt-Winters
(HW), Autoregressive Integrated Moving Average (ARIMA), Exponential Smoothing (ETS), and Artificial
Neural Network (ANN), are implemented, where the most accurate method, ANN is selected to forecast
demand of the new product (sixth generation mobile phone) for the following year. In addition, the
comparison between the original and ANN method shows that ANN is 51.28% more accurate. After that, we
develop the proposed solution plan that links improved demand forecasting to calculate the suitable inventory
quantities and production rates for both finished goods and work in process. The proposed solution scenario
when compared with problem scenario can reduce loss sales and inventory carrying costs by $1,400,626.80
or equivalent to 27.71%.
Demand forecasting is a prediction of product
demand or service demand for a period in the future.
It relies on historical demand data using mathematical
techniques to obtain appropriate forecasting methods
and accurate forecasting values. These precise
demand forecasts result in effective planning,
whether it is planning the use of resources such as
machinery, personnel, as well as purchasing raw
materials necessary to produce finished goods.
Therefore, if there is a lack of accurate forecasting, it
may affect the productivity of the whole production
line. In addition to that, in terms of inventory
management for retail stores, inaccurate demand
forecasting can have far-reaching negative
consequences. For instance, ordering more products
than customers need will result in the problem of
deterioration of products, especially perishable
products such as fruits or fresh food. In addition,
when we store these overordered products for too
long, we may end up disposing of them as waste. In
addition, it may lead to the problem of overstocking
in warehouse management causing loss without cause
of necessary storage space. On the other hand,
inaccurate demand forecasts can also result in
shortages, which is why accurate forecasting is
The research will focus on finding suitable
forecasting methods using time-series data analysis
for the case study of the company's upcoming new
mobile phone products. At present, there is still a
problem of insufficient products to meet the needs of
customers. This originates product shortages causing
customers to wait for products for a long time and
causing customers to change their minds to buy
products from competing companies, resulting in loss
of customers and revenue. By looking into past data,
it shows product shortages and customer waiting
times of 8-10 weeks, see Table 2. Therefore, case
study companies want to analyse historical data to
solve such problems in releasing new products to the
Areerakulkan, N., Moryadee, C., Trakoonsanti, L., Khaengkhan, M. and Setthachotsombut, N.
A New Product’s Demand Forecasting Using Artificial Neural Network.
DOI: 10.5220/0012734800003690
Paper published under CC license (CC BY-NC-ND 4.0)
In Proceedings of the 26th International Conference on Enterprise Information Systems (ICEIS 2024) - Volume 1, pages 417-424
ISBN: 978-989-758-692-7; ISSN: 2184-4992
Proceedings Copyright © 2024 by SCITEPRESS – Science and Technology Publications, Lda.
market for the following year. The remainder of this
paper is structured as follows. Firstly, section 2
presents the objectives of this research and section 3
presents related literature. Then, section 4 presents
research methodology where in-depth problem
analysis and proposed methodology are explained.
Continued to section 5, results of the experiment are
presented to verify the superiority of the proposed
methodology (forecasting using ANN). Lastly,
section 6, summary, discussion, and future research
perspectives are concluded and presented.
The objectives of this research are twofold; first to
conduct in-depth analysis on the current loss sales
problem of a mobile phone manufacturer and second
to find the means to solve such problems and
demonstrate the improvement results.
Normally, we can divide forecasting methods into
three main categories, namely traditional statistics,
machine learning based, and hybrid methods, (Ingle
et. al., 2021) where we summarize those related
literatures as follows.
3.1 Traditional Statistical Forecasting
Most traditional methods use historical sales data to
make forecasts of future demand. It uses time series
analysis methods, namely Autoregressive Integrated
Moving Average (ARIMA), Exponential Smoothing
(ETS), and Holt-Winters. We briefly summarize
these research works as the following paragraphs.
Ghosh (2020) forecast food demand using
ARIMA model, where the ideal model is ARIMA
(1,0,1). In their work, they use Akaike, Schwarz
Bayesian, Maxi-mum likelihood, and Standard Error
to evaluate accuracy of forecasting.
Huber et. al. (2017) forecast demand in a
hierarchical pattern at different organizational levels
where they use multivariate ARIMA model to
forecast daily demand for the bakery supply chain.
They find out that ARIMA is effective and can reduce
the problem of inaccurate forecasting.
Silva et. al. (2019) make demand forecast for the
food industry. They show that Exponential smoothing
method is an accurate and easy-to-use method for
improving production planning effectively.
Kimes et. al. (1998) predict the demand for
various menu dishes in a restaurant. They show that
the Holt-Winters model can effectively forecast these
demands that have both seasonal variation and trend
Sinthukhammoon et. al. (2023) forecast the
demand of Okra for planting community enterprise
located in Kamphaeng Saen District, Thailand. The
forecast methods implemented in their study are
Exponential Smoothing with trend and Seasonal
index. The result shows that Seasonal index method
provide lesser error therefore selected to forecast the
next year Okra demand data.
3.2 Machine Learning Based
Forecasting Model
Machine learning (ML) models use algorithms that
learn from data over time in an automated form.
When compared with traditional forecasting, it is
more accurate, flexible, and easily adjusted according
to various situations, however, traditional methods
are much easier to understand and use. Well known
ML models are Regression, Decision Tree, and Deep
Learning. We summarize the related literatures as
Reynolds et. al. (2013) make forecasts for future
sales in the restaurant industry. In their research, they
implement multiple regression (MR) model
constructed based on 41 years past sales data, where
the presented models were precise and acceptable.
Ma et. al. (2016) perform demand forecasting in
case of high dimensional data for retail product
SKUs. The results show that the use of Multi stages
Lasso regression (LR) plays a significant role in
selecting variables and estimating models. However,
the main problem with LR is that the explanatory
variable space will increase rapidly if we include
promotional matching data in the forecast model.
Priyadarshi et. al. (2019) implement forecasting
models such as ARIMA, long short-term memory
(LSTM) networks, support vector regression (SVR),
and gradient boosting regression (GBR) for
forecasting selected vegetables demand. The results
show that the machine learning algorithms, namely
LSTM and SVR provide more accurate forecasting
when compared to other models.
Ramya and Vedavathi (2020) implement XG
Boost algorithm to predict Rossmann sales data over
eight thousand drug stores. The result shows that XG
Boost gives an excellent sales forecast over ARIMA,
in addition it can assist shops to increase income by
ICEIS 2024 - 26th International Conference on Enterprise Information Systems
analysis of extra information such as advertisement
recommendations, holiday, and competitors.
In addition, there are works that implemented a
hybrid forecasting method where the use of a system
consisting of more than one method together. For
example, Aburto and Weber (2007) used hybrid
model between ARIMA and Neural Network models
to forecast daily product sales. By using ARIMA
outputs as input into Neural Network model, the
results can give more accurate predictions.
3.3 ANN Based Model
Nuanmeesri et. al. (2022) present the combination of
Multilayer Perceptron Neural Network with feature
selection method to predict students drop out during
COVID-19 pandemic of Suan Sunandha Rajabhat
University. The results show that the proposed
method gives the prediction accuracy of
Luckyn and Alabere (2024) determine the sale of
diapers within the retail sector using ANN, where
their historical data contain seasonality patterns,
promotion activities, economic indicators, and
demographic characteristics. The results show that
the ANN can predict diaper sales with high accuracy
and improve consumer satisfaction by decreasing
stockouts and overstock situation.
Rumbe et. al. (2024) introduce two distinct
approaches namely Holt-Winter method and ANN to
forecast tent sales under seasonal influences. The
results show the superiority of ANN over Holt-Winter
method. In addition, the paper explores influential
factors affecting commercial tent sales and
identifying key supply chain players.
Binesh et. al. (2023) propose advanced recurrent
neural network (LSTM) against five traditional
forecasting models to forecast hotel room price under
COVID-19 pandemic situation. The results show that
the LSTM outperform traditional methods such that
the simplest LSTM model is more accurate than that
of the traditional methods.
Raza (2017); Fischera and Kraussb (2017); Xiong
et. al. (2015), use deep learning to predict financial
markets in terms of stock market performance, stock
price, and stock volatility, respectively. The results
show that for nonlinear and large volumes data, deep
learning methods, namely long short-term memory
(LSTM), artificial neural network (ANN), and
generative adversarial networks (GAN), have proven
to be more accurate forecasts compared to traditional
statistical methods or other machine learning
methods. Somehow, one important disadvantage of
deep learning is that it adds computational complexity
and require understanding and computer
programming capabilities.
According to the literature review, the forecasting
method suitable for our research will be the statistical
forecasting method mentioned in section 2.1 and the
Neural Network method (section 2.2), due to main
reasons explained as follows.
1. Firstly, in our research the entrepreneur is
interested in only forecasting one variable,
namely the new product demand. Therefore,
for simplicity it is not necessary to use
complex multi variables forecasting methods
such as decision tree-based method or
regression analysis.
2. Secondly, demand data is stable and clearly
3. Lastly, in our research the entrepreneur needs
forecasting methods that are more convenient
to use and easy to understand over those
complex methods.
This research will begin by thoroughly exploring the
problem to study the root cause of the problem,
collect the necessary data for analysis, then conduct
analysis to find solutions to problems. After that, we
conduct experiments to determine the comparison
results of before and after solving the problem.
Finally, we will propose appropriate measures to
solve the problem, explaining in detail for each step
as follows:
4.1 In-Depth Problem Analysis
For in-depth problem analysis, we collected historical
data to understand what happened during the release
of last year’s products (sixth generation) to market.
Table 1 shows the data of such events.
Table 1 shows underestimation of demand
forecasts every week except in week one. This causes
the customer to not receive the product, resulting in
the cancellation of the order or not placing an order.
Table 2 shows the impact of this problem on false
production planning.
From Table 2, we can identify significant
problems, namely, insufficient finished goods stock,
from week fifty-one to week eight, to meet either
actual demand or forecast. This might originate from
lacking connection among forecasts, production, and
inventory planning. Moreover, finished goods (FG.)
A New Product’s Demand Forecasting Using Artificial Neural Network
Table 1: Historical data of mobile phone released to the
market last year (sixth generation).
Demand (Units) Forecast (Units) Deviation
capacity and work in process (WIP.) capacity are
insufficient to meet the level of market demand, either
forecasted or actual demand.
The main reason of the insufficient FG. stock
problem stems from in-accurate forecasting led to
mis-planning of both FG. and WIP inventory
volumes. Accordingly, we examine the comparison
data between (see Table 1.) forecast and actual sales
of sixth generation mobile phones, we can calculate
the average percentage of absolute error (MAPE) is
as high as 16.46%. As a result of high MAPE, we
must improve demand forecast accuracy by finding a
better forecast method than the original method for
product demand. Noted that, the mobile phone
demand has a combination of both trend and seasonal
4.2 Forecasting Methodology
To achieve the objective mentioned above we
implement various forecasting methods as:
Holt-Winter’s model (HW)
Exponential Smoothing (ETS)
Artificial Neural Network (ANN)
Table 2: Production planning data of mobile phone (sixth generation) released to market last year.
Events Week Pr. Quantity
Level (FG.)
Pr. Quantity
Start WIP.
40 0 0 0
41 0 0 0
42 0 0 0
43 0 0 0
Start FG.
44 42,000 42,000 42,000
45 42,000 84,000 84,000 34,000 120,000
46 42,000 126,000 126,000 34,000 112,000
47 42,000 168,000 168,000 34,000 104,000
48 42,000 210,000 210,000 34,000 96,000
49 42,000 252,000 252,000 34,000 88,000
Release FG.
50 42,000 56,550 56,550 34,000 80,000
51 42,000 0 -78,890 34,000 72,000
52 42,000 0 -149,006 34,000 64,000
1 42,000 0 -170,889 34,000 56,000
2 42,000 0 -157,503 34,000 48,000
3 42,000 0 -136,076 34,000 40,000
4 42,000 0 -110,484 34,000 32,000
5 42,000 0 -78,034 34,000 24,000
6 32,000 0 -52,595 34,000 26,000
7 30,000 0 -26,754 34,000 30,000
8 28,000 0 -1,862 34,000 36,000
9 28,000 23,368 23,368 34,000 42,000
ICEIS 2024 - 26th International Conference on Enterprise Information Systems
Figure 1: Historical weekly sales data for past mobile phone generation.
We visualize the historical demand data, as shown
in Figure 1 which illustrates historical data from the
past 6 years mobile phone generation. Obviously,
sales will be the most in the first week of product
release, after which there will be a significant
decrease in sales from weeks 1-6. From week six
onwards, sales fell in steady volumes to the lowest
sales in week twelve.
Then, we divide the training data set to be demand
data from year 1(first gen) to year 5 (fifth gen). For
the testing data set we use the demand data for year 6
(sixth gen). We measure the accuracy of the
forecasting method by using MAD, RMSE, and
MAPE evaluation as following equations.
Noted that, to determine appropriate parameters
for each forecasting model, we implemented Time
series package in R software.
The next step after obtaining the suitable
forecasting. According to the case study company's
problem survey, we found that there is still a lack of
placement to link among inventory quantities,
demand forecasts, and production rates, whether
finished goods (FG.) or work in process (WIP.).
Therefore, it is necessary to establish this link to
manage inventory efficiently. We summarize the
linking steps as follows.
Step: 1 Calculate the number of weeks to stock
FG. and WIP. for peak demand using equation (4).
Step 2: Make production planning of FG. and
WIP. in accordance with demand and inventory
levels, as shown in Table 5.
where n = number of weeks required to stock FG.
and WIP.
= The forecast amount of demand that
exceeds capacity.
w = weeks in which the forecast demand
exceeds capacity.
k = number of weeks where demand exceeds
Lastly, we conduct costs comparison between
proposed scenario versus that of problem scenario
based on two costs components as approximated loss
sales and inventory carrying costs.
As previously mentioned, we preliminary select
forecasting methods that are suitable for this problem.
These are (1) Holt-Winters' Seasonal Method (HW.),
(2) Autoregressive Integrated Moving Average
(ARIMA), (3) Exponential Smoothing (ETS.), and
(4) Artificial Neural Network (ANN.).
A New Product’s Demand Forecasting Using Artificial Neural Network
For various forecasting methods, it is necessary to
identify appropriate parameters to make the most
accurate forecasts. We summarize these parameters
for each forecast method as follows.
HW forecasting showed that the optimal
parameters with the least forecast error were
multiplicative, with smoothing coefficients
of α = 0.0463, β=0.0443, and γ=0.8344.
For ARIMA models, the appropriate
parameters that cause the least tolerances
are: ARIMA (0,0,1) (1,1,1) [12] with drift.
The Exponential Smoothing (ETS) forecast
method found that optimal parameters were
α = 0.0242, β=0.0242, and γ=0.9758.
The last method, the Neural Network (ANN)
method, found that the ideal model was
NNAR (2,1,2) [12] that was seasonal, and
lagged 1, 2, and 12 (y
.) of each
season were inputs, with 2 Neurons in the
Hidden Layer.
Table 3: Forecast errors for training dataset (gen1-gen5).
1,210.91 949.97
2,354.12 1427.46
2.99 5.41
As shown in Table 3, ANN. method provides the
smallest forecast errors for MAD. and RMSE. cases,
while ETS. provides the smallest error in MAPE. As
mentioned in (Vandeput, 2021), selecting the suitable
demand forecast method based on using MAPE., the
forecast value is often lower than it should be. While
using RMSE., the forecast value is about the average
value. On the other hand, MAD. often gives higher
forecast value than it should be. Therefore, in this
research, we prefer average forecast value, so RMSE.
seems appropriate. Therefore, we select the right
forecasting method based on RMSE., where the best
forecast method in this case is ANN.
For testing data set, by using the ANN method, we
forecast demand of sixth generation mobile phone
and compare it with the original forecast method, see
Table 4 for the ANN. forecast results. From that table,
it shows that forecast values obtained by ANN
method are more accurate than those of the original
forecast. The original method has forecast error based
on calculated MAD, RMSE, and MAPE as 7972.17,
12410.48, and 16.46, respectively. While proposing
ANN has MAD = 3030.044, RMSE = 6046.08, and
MAPE = 5.41, respectively. In other words, ANN is
51.28% more accurate than the original forecast
method based on RMSE.
Table 4: Forecast results comparison for gen sixth model.
Week Demand Old Forecast New Forecast
Then, we calculate number of weeks to stock FG.
and WIP. using equation (4). WIP. should start at
week thirty-seven while FG. should start at week
forty-two, respectively. Additionally, the capacity
should increase from 42,000 to 50,000 units. The
proposed production planning for solution scenario of
sixth generation mobile phone is conducted and
shown in Table 5.
In this paper, we found the major problem which is
inaccurate demand forecast that causes mis planning
of both FG. and WIP inventory volumes. Therefore,
we present the better forecast method based on ANN,
which gives 51.28% more accurate than that of the
original method. Then, we develop the proposed
solution plan that links together inventory quantities,
demand forecasts, and production rates, for both
finished goods and work in process.
We then compare the proposed solution scenario
with the problem scenario based on two cost
components as approximated loss sales and inventory
carrying costs. By knowing the sales margin, in this
case $155.00, we calculate the total sales loss value
of the problem scenario as $3,085,070.00, where the
proposed solution scenario has zero loss sales cost (no
For the inventory carrying cost based on weekly
carrying cost for WIP inventory/unit =$0.53, and
weekly carrying cost for FG inventory/unit = $1.20,
we calculate inventory carrying cost for problem
scenario as $1,969,441.60 and that of solution
ICEIS 2024 - 26th International Conference on Enterprise Information Systems
Table 5: Production planning for solution scenario of sixth
generation mobile phone.
Events Week Pr.
37 0 0 37,500
38 0 0 37,500
39 0 0 37,500
40 0 0 37,500
Start FG.
41 0 0 37,500
42 50,000 50,000 37,500
43 50,000 100,000 37,500
44 50,000 150,000 37,500
45 50,000 200,000 37,500 137,500
46 50,000 250,000 37,500 125,000
47 50,000 300,000 37,500 112,500
48 50,000 350,000 37,500 100,000
49 50,000 400,000 37,500 87,500
50 50,000 215,000 37,500 75,000
51 50,000 90,000 37,500 62,500
52 50,000 30,000 37,500 50,000
1 50,000 18,000 37,500 37,500
2 27,061 18,061 0 10,439
3 10,439 9,500 18,944 18,944
4 14,974 9,474 0 3,970
5 3,970 4,944 8,237 8,237
6 5,409 4,853 0 2,828
7 2,828 4,181 3,423 3,423
8 2,355 4,086 0 1,068
9 1,068 2,604 2,462 2,462
scenario as $3,653,885.23. Therefore, summing loss
sales and carrying costs together we obtain the total
cost for problem and solution scenario as
$5,054,512.03 and $3,653,885.23, respectively. On
the other hand, upon implementing solution scenario,
we can reduce costs by $1,400,626.80 or by 27.71%.
Somehow, in this research, we propose the
solution scenario on forecasting only product demand
without considering other important variables such as
promotion and competitors. Therefore, for future
research, it would be more practical if we included
these variables into building the forecast model.
Aburto, L., Weber, R. (2007). A sequential hybrid
forecasting system for demand prediction. In: Perner, P.
(ed.) MLDM 2007. LNCS (LNAI), vol. 4571, pp. 518–
532. Springer, Heidel-berg.
Binesh, F., Belarmino, A. M., Van der Rest, J. P., Singh, A.
K., & Raab, C. (2024). Forecasting hotel room prices
when entering turbulent times: a game-theoretic
artificial neural network model. International Journal of
Contemporary Hospitality Management, 36(4), 1044-
Fischera, T., Kraussb, C. (2017). Deep learning with long
short-term memory networks for financial market
predictions. FAU Discussion Papers in Economics.
Ghosh, S. (2020). Forecasting of demand using ARIMA
model. American Journal of Applied Mathematics and
Computing, 1(2), 11-18.
Huber, J., Gossmann, A., & Stuckenschmidt, H. (2017).
Cluster-based hierarchical demand forecasting for
perishable goods. Expert systems with applications, 76,
Ingle, C., Bakliwal, D., Jain, J., Singh, P., Kale P., and
Chhajed, V. (2021). Demand Forecasting: Literature
Review on Various Methodologies, 12th International
Conference on Computing Communication and
Networking Technologies (ICCCNT), Kharagpur,
India, pp. 1-7, doi: 10.1109/ICCCNT51525.2021.958
Kimes, S. E., Chase, R. B., Choi, S., Lee, P. Y., & Ngonzi,
E. N. (1998). Restaurant revenue management:
Applying yield management to the restaurant industry.
Cornell Hotel and Restaurant Administration Quarterly,
39(3), 32-39.
Luckyn, B. J., Alabere, I., & Ogra, O. (2024). Predictive
Modelling for Diaper Sales in Retail: An Artificial
Neural Network Approach. Iconic Research and
Engineering Journals, 7(8).
Ma, S., Fildes, R., & Huang, T. (2016). Demand forecasting
with high dimensional data: The case of SKU retail
sales forecasting with intra-and inter-category
promotional information. Euro-pean Journal of
Operational Research, 249(1), 245-257.
Nuanmeesri, S., Poomhiran, L., Chopvitayakun, S., &
Kadmateekarun, P. (2022). Improving dropout
forecasting during the COVID-19
pandemic through
feature selection and multilayer perceptron neural
network. International Journal of Information and
Education Technology, 12(9), 851-857.
Priyadarshi, R., Panigrahi, A., Routroy, S., & Garg, G. K.
(2019). Demand forecasting at retail stage for selected
vegetables: a performance analysis. Journal of
Modelling in Management, 14(4), 1042-1063.
Ramya, B. S. S., & Vedavathi, K. (2020). An advanced
sales forecasting using machine learning algorithm.
International Journal of Innovative Science and
Research Technology, 5(5), 342-345.
Raza, K. (2017). Prediction of stock market performance by
using machine learning techniques. In: International
A New Product’s Demand Forecasting Using Artificial Neural Network
Conference on Innovations in Electrical Engineering
and Computational Technologies (ICIEECT), pp. 1.
Reynolds, D., Rahman, I., & Balinbin, W. (2013).
Econometric modelling of the US restaurant industry.
International Journal of Hospitality Management, 34,
Rumbe, G., Hamasha, M., & Al Mashaqbeh, S. (2024). A
comparison of Holts-Winter and Artificial Neural
Network approach in forecasting: A case study for tent
manufacturing industry. Results in Engineering, 21,
Silva, J. C., Figueiredo, M. C., & Braga, A. C. (2019).
Econometric Demand forecasting: A case study in the
food industry. In Computational Science and Its
Applications–ICCSA 2019: 19th International
Conference, Saint Petersburg, Russia, July 1–4, 2019,
Proceedings, Part III 19 (pp. 50-63). Springer
International Publishing.
Sinthukhammoon, K., Hiranphaet, A., Sooksai, T.,
Chumkad, K., & Sonong, S. (2023, March). Demand
forecasting of Okra: case study of Okra planting
community enterprise of Kamphaeng Sean district:
International Academic Multidisciplinary Research
Conference, (pp. 198-202).
Vandeput, N. (2021). Data Science for Supply Chain
Forecasting, Berlin, Boston: De Gruyter.
Xiong, R., Nichols, E. P., & Shen, Y. (2015). Deep learning
stock volatility with google domestic trends. arXiv
preprint arXiv:1512.04916.
ICEIS 2024 - 26th International Conference on Enterprise Information Systems