Demand Forecasting System for Small Business

Optimization
Mary Cherian, Aniket Jane, Deepak, G Sinchana, MD. Modasscer
Department of Information Science and Engineering
Dayananda Sagar Collage of Engineering

Abstract- Demand forecasting is a fundamental element in efficient inventory

control and supply chain management through the ability of firms to anticipate
future demand and devise methods to service that demand accordingly. The
present paper explores two major approaches to demand forecasting: the SARIMA
model and the DeepAR model. SARIMA is a more traditional statistical approach
and exploits autoregressive, integrated, and moving average components along
with seasonal adjustments to pick up trends and seasonality in the sales data.
DeepAR is the most advanced deep learning-based probabilistic approach in
modelling complex time series dynamics and relationships within interconnected
time series. This paper details the process of demand forecasting starting from data
collection, preparation, and cleaning stages to model building, testing, and
validation. SARIMA was able to clearly prove its fit for short- to medium-term
forecasting, enabling it to be very efficient in keeping track of inventory stocks and
preventing stockouts. However, DeepAR does outperform SARIMA if using large
dynamic datasets where non-linear patterns come into play, providing
probabilistic forecasts that quantify uncertainty. Our findings have illuminated the
strengths and weaknesses of both methods. Given the stable, well-structured
datasets, SARIMA is a robust choice. However, DeepAR's advanced capabilities
are well-suited for businesses dealing with complex and dynamic conditions. Firms
can optimize operations to adapt to market fluctuations and even make data-
driven decisions to enhance efficiency and customer satisfaction by using these
tools.

Keywords— Demand Forecasting, Machine Learning, Hybrid Model, DeepAR, SARIMA.

1. INTRODUCTION improving and estimating the range of

predictions. Most of these methods are SARIMA
Demand forecasting is critical to modern models, which are mainly used in the modeling
business practice in order to maintain a of time-dependent processes thanks to the
competitive edge: it works together with efficient incorporation of structure: trends, seasonality
management of inventories and scheduling of and other regimes present in the time series data.
production and it controls supply chain activities. SARIMA works very well for industries such as
Being familiar with the meaning of demand retail, manufacturing and logistics those have a
forecasting is very crucial since it involves the demand curve with a certain level of oscillation,
intervals for which the demands for the particular because of the prevailing seasonality, for short-
goods or services are expected. Good forecasting medium range of forecasting periods. However,
prevents any costs associated with overstocking these types of models often have a challenge in
and understocking, improves customer service capturing the irregular swings and structural
levels, and improves the efficiency of the supply changes when there is intricate structure
chain. Forecasting methods in time series contained in the data base used even within the
analysis have been developed with the aim of dynamic markets.
 inclination and situation related assessment
Advanced types of machine learning techniques criteria, Mean Squared Error (MSE) and Mean
such as DeepAR emerged as a strong source for Adjusted Error (MAE), etc. The outcomes
demand forecasting. DeepAR is a deep learning- showed that ELM and MLP were in all likelihood
based stochastic forecasting method which is better than traditional approaches meaning that
able to learn various time series patterns. This they could be depended upon as far as maps of
method is good at dealing with wide nonlinear their structure were concerned.
systems, high noise levels and stopping the
market instability. This is the primary reason that [3] José Eduardo de Carvalho Lima et al. made a
it operates particularly well in business lines like systematic review that provides some areas for
e-commerce and digital services where the improvement regarding demand forecasting and
demand pattern is highly unpredictable. planning of production systems, but it is not very
useful practically since it does provide a
2. LITERATURE SURVEY comparative assessment of different models. On
the background of the study, some models are
[1] Tereza Šustrová et. al. from the Brno very complex and therefore, applicability can be
University of Technology, explored the a challenge. In this regard, the scope of research
challenges posed by the conventional inventory is imbalanced as some crucial areas remain
models. With this, they recommended the inadequately addressed. In any case, the paper is
application of artificial neural networks (ANNs) helpful in identifying weaknesses of the existing
due to the complexities posed by supply chains. research, but it could have improved its
The authors experimented with different contributions through more comprehensive
configurations of ANNs, while the training treatment of some particular techniques used.
function TRAINGDX and transfer function
TANSIG were employed. This research applied [4] Aditya Chawla et al. analysed applications of
data analysis and pre-processing with MS Office ANNs in demand forecasting. Scenarios
Excel and the model creation, testing was done presented the capability of ANNs in demand
with the Neural Network Toolbox of MATLAB forecasting. Results thereby attained reflected
software. In this study, the optimal model that satisfying for using ANNs in forecasting, mainly
had eight hidden neurons was very effective (R² in the integration of complex data patterns. The
= 0.99) and had very low Mean Squared Error huge computational costs involved with models
(MSE = 0.029) and found that Managing Neural developed using ANNs will be one major
Networks can help even to determine how much drawback to this kind of model. The
order to place in order to enhance management of computational complexity does cause a particular
order cycle dealing with inventory levels problem when very large data sets are involved,
problems. since tools like MATLAB, unfortunately, tend to
limit the ability to scale up models that would
[2] Maryam Zohdi et al. of Sharif University of accommodate large volumes of data satisfactorily.
Technology and Hamad bin of Khalifa
University have put forward the idea of [5] Muhammad Adnan Khan et al. at Lahore
advancing demand prediction by machine Garrison University designed and implemented a
learning algorithms. The authors discussed the business intelligence driven demand forecasting
limitations of oft-used statistical approaches such system which integrates machine learning into its
as ARIMA when applied to complex non-linear workings. Once they applied the deepAR
data. Following this, they, perhaps excitedly, algorithm on AWS Sagemaker, they could
started looking up some of the more advanced analyse sales and inventory data with up to 92.38%
approaches like Extreme Learning Machine accuracy. Most importantly, this model can be of
(ELM), Decision Trees, Gradient Boosting, substantial assistance to companies those want to
Multi-Layer Perceptron (MLP), K-nearest reduce wasteful production and improve the
Neighbors (KNN), and so on. The models used utilization of available resources, especially
acquired data from a marketing campaign those in need of new products but have little data
referred to as Black Friday, with some customer to consult.
 combining of human judgement and models with
[6] Real Carbonneau et al., of Concordia demand forecasting within changing
University, studied some machine learning environments.
techniques, particularly NN, RNN, and SVM, in
addressing bullwhip effect in supply chain [9] Fatima Zohra Benhamida et al. published a
demand forecasting. These advanced techniques review paper on demand forecasting methods for
were then compared with traditional methods in time-series data. The paper was proposing a
the form of moving averages and regression novel demand forecasting tool for inventory
models with both simulated data and real data. control. In such a study, the strength in
Though RNNs and SVMs have shown an integrating statistical and machine learning
improvement in the accuracy of their forecasting, methods is presented. However, such a study is
the differences weren't statistically significant. lacking in practical comparative insights into its
SVMs provided a solid mathematical base to application across diversified industries. Given
excel in dealing with complex time series with the background of the study, the complexity in
minimum overfitting. Therefore, the use of hybrid methods gives rise to problems in
machine learning is expected to increase the adoption. Additionally, the scope is somewhat
quality of forecasts in order to streamline the imbalanced because it focuses less on the area of
supply chain. intermittent demand forecasting and new product
introduction forecasting. The paper identifies
[7] Fattah et al. investigated demand forecasting substantial gaps and is very helpful for
using the ARIMA model in a Moroccan food developing robust tools for forecasting.
company, employing the Box-Jenkins
methodology with historical data from 2010– [10] Kanik et al. investigated the application of
2015. They developed and evaluated several regression-based machine learning techniques
ARIMA models based on criteria like AIC, SBC, for forecasting sales within supply chain
and maximum likelihood, selecting ARIMA (1, 0, operations. The study focused on leveraging
1) as the most suitable. The chosen model predictive analytics to analyse sales data,
demonstrated high accuracy in predicting identifying demand trends and patterns. Their
demand, facilitating improved production findings highlighted how regression models
planning, inventory management, and cost could support supply chain optimization and
reduction. This study highlighted ARIMA’s enhance inventory management decisions.
effectiveness in forecasting demand for
perishable goods, ensuring better decision- [11] Sukolkit et al. proposed an innovative
making and enhanced supply chain efficiency. approach to inventory management that
integrates demand forecasting techniques
[8] Naoufel Cheikhrouhou et al. performed a specifically tailored for the steel industry. The
review on collaborative forecasting that study emphasized the importance of open
combines mathematical models with event-based innovation in addressing inventory challenges,
fuzzy judgments. The study highlights the utilizing advanced forecasting methods to predict
problems of using these methods because of demand trends accurately. Their research
different perceptions by the forecasters and the provided insights into optimizing inventory
difficulty of integrating contextual knowledge. management processes in a dynamic and
Although it emphasizes collaboration within complex industrial environment.
firms, it pays less attention to inter-firm
processes, thus creating a gap in the overall [12] Khan et al. (2020) introduced a demand
analysis. The research explains the limitations of forecasting model enhanced by business
conventional forecasting models and proposes intelligence and machine learning techniques.
systematized solutions to improve predictions. The study highlighted the integration of machine
However, it does not offer a critical comparison learning algorithms to improve the accuracy and
of various techniques available and, therefore is efficiency of demand predictions. By leveraging
less useful practically. Nevertheless, the study business intelligence tools, the model provided
outlines important insights related to the actionable insights, supporting strategic
decision-making processes across various
business operations. The SARIMA model is a statistical method of
time series forecasting aimed at data modelling
3. Proposed System that follows the seasonal patterns. Its
methodology is essentially an extension of the
A. Data Source ARIMA model, taking the seasonal into account,
and is particularly well-suited for datasets in
which seasonality makes a significant difference
to demand. The three components on which
SARIMA is based are Autoregressive, Integrated,
3. Methodology and Moving Average, with their seasonal
counterparts. Notation SARIMA (p, d, q) (P, D,
Q, m) where p, d, and q depict the non-seasonal
AR, differencing and MA terms. Their seasonal
equivalents, P, D, Q, and m with m as the number
of observations in each season are denoted,
Figure 1. Overview of System Data Source
The SARIMA method focuses on an exhaustive
analysis of the data to work out the underlying
The proposed system relies on two key data trends, seasonality, and noise. This involves
sources: sales data and inventory data, both of exploratory data analysis, where visualizations
which play a vital role in generating accurate and various statistical tests, such as the
predictions for unit sales. Each dataset is autocorrelation and partial autocorrelation
categorized using unique Stock Keeping Unit functions, help identify possible seasonal lags
identities (SKU IDs) and Store identities, and check for stationarity. Stationarity is a crucial
ensuring that the forecasts are highly specific to assumption in SARIMA modelling, since it
individual products and their respective store ensures that the statistical properties of the time
locations. Historical sales data captures trends in series, mean and variance, are invariant with
customer demand, while inventory data ensures respect to time. In the case of non-stationary in
that predictions are aligned with stock data, differencing is done. Seasonal differencing
availability, helping to avoid overstocking or is specially used to remove periodic fluctuations,
stock shortages. while regular differencing removes overall trends.
This data is uploaded into a web interface where
it undergoes preprocessing, such as cleaning and This paper will focus on the usage of a robust
formatting, to prepare it for machine learning methodology for forecasting demand through
analysis. Advanced algorithms are then applied SARIMA and DeepAR models and a stronger
to identify patterns and correlations between methodology for accurate and more interpretable
various factors influencing demand. By results. The dataset has historical sales data
combining real-time and historical data labelled supplemented with economic indicators and
with SKU IDs and Store IDs, the system seasonal trends to capture demand fluctuations
produces reliable, product-specific, and location- comprehensively. Data was sourced from
based forecasts. The system can also incorporate internal company repositories and external
external factors, such as seasonal trends or publicly available databases. To ensure sufficient
market shifts, to enhance prediction accuracy coverage of patterns and anomalies, a daily
further. This comprehensive approach enables granularity spanning five years is chosen,
businesses to make informed decisions, offering both short-term and long-term
streamline inventory management, and improve perspectives critical for effective forecasting.
overall operational efficiency, ultimately
boosting customer satisfaction. Before model application, rigorous data pre-
B. Machine Learning Models processing steps are undertaken. Missing values,
a common challenge in time series datasets, are
SARIMA AND DEEPAR: handled using linear interpolation, preserving the
temporal continuity essential for accurate particularly useful in providing an intuitive
forecasting. Outliers could skew model interpretation of the accuracy of forecasts
predictions. They are detected using interquartile relative to observations. These metrics
ranges and replaced with median values, hence collectively ensure a balanced assessment of
upholding the integrity of the dataset. The model efficacy across different dimensions of
variance of the data is stabilized through log error sensitivity and interpretability.
scaling, and differencing is used to ensure The final choice of model is guided by
stationarity; this is critical for models as comparisons on the validation dataset. Both
complicated as SARIMA. These pre-processing SARIMA and DeepAR are compared to the
steps ensure that the dataset is optimized for evaluation metrics defined. SARIMA's
reliable analysis and modelling. interpretability together with its statistical basis
is compared with the ability of DeepAR to model
For model choice, both SARIMA and DeepAR non-linear dependencies as well as the capability
are chosen for their complementary strengths in of integrating covariates. In situations where both
handling seasonal and non-linear time series data. have similar performance, operational factors,
SARIMA is most suited for datasets exhibiting such as computational efficiency and scalability,
clear seasonal patterns, as it can include play a decisive role. The model with the
autoregressive, moving average, and seasonal minimum average Mean absolute percentage
components. This makes it strong in modelling error over validation datasets is preferred for
periodic demand variations. However, DeepAR, deployment with consideration of practical
a probabilistic forecasting model that uses implementation aspects, aligning with business
recurrent neural networks (RNN), proves objectives. With this methodology, then, the
superior in grasping complex dependencies SARIMA and DeepAR are combined in order to
present in the data. The ability of deep learning provide robust, accurate, and interpretive
models to deal with covariates, and hence demand forecasts. Thus, by incorporating
outputting probabilities for the underlying rigorous data pre-processing, systematic
predictions, is extremely helpful in dynamic parameter tuning, and comprehensive evaluation,
patterning of missing data. this approach will ensure that the model chosen
is not only technically sound but also
The train is split into 70% for training, 20% for operationally viable. It will allow businesses to
validation, and 10% for testing so as to allow make informed decisions and optimize inventory
robust model evaluation and generalization. levels and, hence, improve general supply chain
SARIMA's parameters, comprising efficiency.
autoregressive (p), differencing (d), and moving
average (q) components, together with seasonal
counterparts (P, D, Q, m) are selected with grid
search and the model is assessed with Akaike
Information Criterion(AIC). In DeepAR, hyper
parameters like the number of RNN layers,
hidden units, and learning rate are tuned with
Bayesian search to ensure the model is both
efficient and accurate.

Model performance is evaluated using several

metrics, namely Mean Absolute Error, Mean
Squared Error, and Mean Absolute Percentage
Error. Mean Absolute Error gives a simple

average measure of deviation, while the Mean

Squared Error punishes larger errors, making it
more sensitive to significant deviations in
forecasts. Mean absolute percentage error is
 the combined forecast over
the prediction period,
Figure: 2 System Architecture
4. SIMULATION illustrating sales trends and
RESULTS fluctuations for better
decision-making. By integrating visualization
This simulation demonstrates demand tools, the interface provides actionable insights,
forecasting using two models, SARIMA and enabling businesses to respond proactively to
DeepAR, to predict product sales for a specific demand fluctuations. The ability to fine-tune
SKU (216419) at a particular store (Store ID: parameters for both models ensures flexibility
8091). The interface allows users to upload data, and adaptability to diverse forecasting scenarios,
set a prediction horizon (default of 12 weeks), enhancing decision-making and operational
and input store and SKU details. The SARIMA efficiency. This simulation, therefore shows how
model, a traditional statistical approach, provides effective advanced deep learning approaches,
predictions with an accuracy of 50.76%, such as DeepAR, are in demand forecasting and
reflecting its limitation in capturing complex the importance of amalgamation of traditional
temporal patterns and irregularities. In contrast, and modern approach towards achieving
the DeepAR model, which is based on deep improved accuracy metrics to allow comparison.
learning, achieves a much higher accuracy of
81.0%, indicating its ability to better model time-
series data and identify trends. To further
enhance reliability, a combined model is created
by merging predictions from both SARIMA and
DeepAR, resulting in an overall accuracy of
77.14%.This is a hybrid approach that balances
the strengths of both models, hence offering a
robust forecasting solution. All three models'
predicted values are presented with accuracy the

metrics to allow comparisons. A graph visualizes

 Figure: 3 Web Interface [1] M. Zohdi, M. Rafiee, V.
Kayvanfar, and A.
REFERENCES Salamiraad, "Demand

forecasting based machine

Figure: 4 Prediction Interface
learning algorithms on
customer information: an applied approach," no. 2, 2021.
International Journal of Information Technology,
vol. 14, no. 2, Feb. 2022. [7] J. E. de Carvalho Lima, P. R. A. Firmino,
and L. A. O. Rocha, "Demand forecasting,
production planning, and control: a systematic
[2] N. Cheikhrouhou, F. Marmier, O. Ayadi, and
literature review," Advanced Structured
P. Wieser, "A collaborative demand forecasting
Materials, 2020.
process with event-based fuzzy judgements,"
Computers & Industrial Engineering, vol. 61, no.
[8] A. Chawla, A. Singh, A. Lamba, N.
3, July 2011.
Gangwani, and U. Soni, "Demand Forecasting
using Artificial Neural Networks: A Case Study
[3] T. Šustrová, "A Suitable Artificial of American Retail Corporation," International
Intelligence Model for Inventory Level Conference on Artificial Neural Networks,
Optimization," Trendy Ekonomiky a 2017.
Managementu – Trends Economics and
Management, vol. 25, no. 1, 2016. [9] J. Fattah, L. Ezzine, Z. Aman, H. E.
Moussami, and A. Lachhab, "Forecasting of
demand using ARIMA model," International
[4] R. A. Carbonneau, K. Laframboise, and R. Journal of Engineering Business Management,
Vahidov, "Application of machine learning vol. 10,Oct. 2018.
techniques for supply chain demand
forecasting," European Journal of Operational [10] K. Kanik, A. C. Shukla, and S. Punjabi,
Research, vol. 184, no. 3, Feb. 2008. "Sales Forecasting Using Regression- Based
Machine Learning Algorithms in Supply Chain
[5] G. Ranasinghe, K. Wickramasinghe, and P. Environment," Thesis, Rajiv Gandhi
Perera, "Demand Forecasting of a Multinational Proudyogiki Vishwavidyalaya, Bhopal, June
Retail Company using Deep Learning," 2023.
Proceedings of the 8th International Conference
on Advances in ICT for Emerging Regions [11] N. Sukolkit, S. Arunyanart, and A.
(ICTer), 2019. Apichottanakul, "An open innovative inventory
management-based demand forecasting
approach for the steel industry," Journal of Open
[6] F. Z. Benhamida, O. Kaddouri, T. Ouhrouche, Innovation: Technology, Market, and
M. Benaichouche, D. Casado-Mansilla, and D. Complexity, vol. 10, Oct. 2024.
Lpez-de-Ipiña, "Demand Forecasting Tool for
Inventory Control: .
[12] M. A. Khan et al., "Effective Demand
Stock&Buy Case Study,"Journal of Forecasting Model Using Business Intelligence
Communications Software and Systems, vol. 17, Empowered With Machine Learning," IEEE
Access. 2020.