Combating the Bullwhip Effect in Rival Online Food Delivery

Platforms Using Deep Learning

Tisha Ghosh

Abstract
The wastage of perishable items has caused significant health and economic crises, leading to
uncertainty in the business environment and fluctuating customer demand. In modern times, this
issue has intensified with the rise of online food delivery services, where frequent and unpredictable
orders contribute to inefficiencies in supply chain management. These fluctuations contribute to the
arXiv:2503.22753v1 [cs.LG] 27 Mar 2025

bullwhip effect, a major supply chain issue that results in stockouts, excess inventory, and inefficien-
cies. Accurate demand forecasting and predictive analysis help stabilize inventory, optimize supplier
orders, and reduce resource wastage. This paper proposes a Third-Party Logistics (3PL) supply
chain model involving key stakeholders— restaurants, online food apps, and customers—alongside a
deep learning-based demand forecasting model using a two-phase Long Short-Term Memory (LSTM)
network. Phase 1 (Intra-day demand forecasting) captures short-term variations within a day, while
Phase 2 (Daily demand forecasting) predicts overall daily demand. A two-year historical dataset
(January 2023 – January 2025) with daily sales data from Swiggy and Zomato is used to capture
nonlinear demand trends through discrete event simulation. The grid search method is employed
to select optimal LSTM hyperparameters, ensuring precise forecasting. The proposed method is
evaluated and compared using Root Mean Square Error (RMSE), Mean Absolute Error (MAE), and
R² score, along with optimized training time, among which R² serves as the primary measure of
forecasting accuracy. Phase 1 achieves an R² score of 0.69 for Zomato and 0.71 for Swiggy with a
training time of 12 minutes, while phase 2 shows substantial improvement, with R² values of 0.88
(Zomato) and 0.90 (Swiggy), achieving a training time of 8 minutes, demonstrating strong predic-
tive performance in demand forecasting. To mitigate demand fluctuations, restaurant inventory is
dynamically managed using the newsvendor model, adjusted based on forecasted demand. The pro-
posed framework significantly reduces the bullwhip effect, enhancing forecasting accuracy and supply
chain efficiency. For Phase 1, the overall supply chain instability decreases from 2.61 to 0.96, while
for Phase 2, it reduces from 2.19 to 0.80, demonstrating the model’s effectiveness in minimizing food
waste and maintaining optimal inventory levels of restaurants.

Keywords: Time series forecasting, Demand forecasting, Supply chain, Deep learning, LSTM,
Discrete event simulation, Bullwhip effect.

1 Introduction
Supply Chain Management (SCM) has been a critical area of research for a long time, with a particular
focus on the management of perishable goods (Blackburn and Scudder, 2009). The food supply chain, in
particular, faces numerous challenges, including stockouts, overstocking, ingredient waste, bottlenecks,
and the bullwhip effect (Li et al., 2014; Folkerts and Koehorst, 1997). In today’s fast-paced world, the
younger generation has embraced the convenience of online food delivery apps, making them an essential
part of daily life. However, this growing reliance on food delivery services has led to unpredictable
demand patterns, resulting in frequent stockouts and inventory fluctuations. To compensate, restaurants
and suppliers often overstock to avoid shortages, which results in excessive wastage of perishable items.
This issue is a manifestation of the bullwhip effect, where minor changes in consumer demand lead
to increasingly significant fluctuations in orders and stock levels as the supply chain moves upstream
(Lee et al., 1997a). Usually, it is caused by factors like inaccurate demand forecasting, order batching,
pricing changes, and information flow delays. To ensure smoother supply chain operations, managing the
bullwhip effect requires improved supply chain coordination, demand forecasting techniques and accurate
information sharing. Demand is one of the most crucial data points that must be communicated among
supply chain participants (Abolghasemi et al., 2020), and effective demand forecasting play a crucial
role in addressing these challenges (Haberleitner et al., 2010; Kumar et al., 2020), ensuring customer
satisfaction while preventing supply chain disruptions.

1
 Traditional forecasting models, such as ARIMA (AutoRegressive Integrated Moving Average), Holt-
Winters moving average, and exponential smoothing, have been widely used for demand prediction
(Moroff et al., 2021; Abbasimehr et al., 2020). However, these methods are primarily linear and often
fail to accurately capture real-world demand patterns, which tend to be non-linear. To address these
limitations, non-linear statistical models like ARCH (Autoregressive Conditional Heteroskedasticity) and
GARCH (Generalized Autoregressive Conditional Heteroskedasticity) have been introduced (Khashei and
Bijari, 2011). While these models offer improvements over purely linear approaches, they still struggle
to fully capture complex non-linear dependencies in demand fluctuations.
As a result, companies are seeking more advanced methods to closely align with customers’ needs. In
response to these challenges, businesses have increasingly turned to machine learning (ML) techniques,
which have demonstrated superior predictive capabilities across various industries (Min, 2010). ML-
based forecasting models have been successfully applied to demand prediction in sectors such as furniture
(Abbasimehr et al., 2020), energy (Szul et al., 2020), cash flow at ATMs (Catal et al., 2015), tourism
(Shahrabi et al., 2013), and natural gas supply chains (Potočnik et al., 2019). Some of the most widely
used ML techniques for time series forecasting include Artificial Neural Networks (ANNs), Support
Vector Machines (SVMs), K-Nearest Neighbors (KNNs), and Adaptive Neuro-Fuzzy Inference Systems
(ANFIS)(Abbasimehr et al., 2020).
Among these, Recurrent Neural Networks (RNNs) have gained significant attention due to their
ability to recognize sequential dependencies in time-series data and predict future trends based on past
observations (Amirkolaii et al., 2017). Recurrent Neural Networks (RNNs) are widely applied across
various areas of the supply chain. For instance, (Aktepe et al., 2021) demonstrated the superiority of
RNNs in forecasting spare parts demand. Similarly, Rahman et al. (2018) showcased their effectiveness
in medium- and long-term electricity consumption predictions. Additionally, Potočnik et al. (2019)
compared RNNs with linear regression and extreme machine learning algorithms for forecasting natural
gas demand, incorporating factors such as historical temperatures and time-based variables, including
holidays and other seasonal events. Their findings highlighted the superiority of RNNs. However, research
also indicates that RNNs are not always the optimal choice for supply chain applications (Amirkolaii et al.,
2017). Their limitations stem from short memory retention and the vanishing gradient problem, which
complicates training and reduces their effectiveness in certain scenarios. Unlike traditional ANNs, RNNs
incorporate feedback loops, allowing information to persist across time steps, making them particularly
effective for time-dependent forecasting tasks. However, RNNs suffer from short-term memory limitations
(Hochreiter and Schmidhuber, 1997) and the vanishing gradient problem, which reduces their effectiveness
for long-term forecasting applications. To overcome these drawbacks, Long Short-Term Memory (LSTM)
networks were developed. LSTMs are specifically designed to retain relevant information over extended
periods by distinguishing between short-term and long-term dependencies. This characteristic makes
them highly effective for time series forecasting, as they can learn patterns in demand fluctuations over
both short and long durations (Wu et al., 2018; Siami-Namini et al., 2019). After successfully capturing
demand patterns, companies stabilize their inventory.
In this study, a two-phase multilayer LSTM model is developed and optimized using the grid search
technique, which systematically explores different hyperparameter configurations to enhance forecasting
accuracy. The proposed model is applied to third-party logistics (3PL) providers in the online food
delivery industry who forecast demand, such as Uber Eats, DoorDash, and Zomato. They act as com-
petitive demand aggressors (CDA) by connecting business-to-consumer (B2C) with business-to-business
(B2B) restaurant suppliers in highly active Indian cities where online food delivery is most prevalent,
utilizing synthetic sales data for 2 years on two competitive platforms, such as Zomato and Swiggy, to
validate their performance. This research contributes to the field of supply chain in several ways. First,
a synthetic data set is simulated with relevant columns, replicating real-world online food sales data
from highly active Indian cities where Zomato and Swiggy operate. Second, it develops and fine-tunes a
two-phase multilayer LSTM model using the GridSearch method to achieve optimal forecasting precision-
Phase 1 (Intra-day demand forecasting) captures short-term fluctuations within a single day and Phase
2 (Daily demand forecasting) predicts overall daily demand trends for improved inventory planning. It
develops and fine-tunes a multi-layer LSTM model using the Grid Search method to achieve optimal
forecasting precision. Third, inventory replenishment is optimized based on dynamically forecasted de-
mand fluctuations, preventing both overstocking and shortages. Fourth, the reduction in the bullwhip
effect is measured by analyzing demand fluctuations and the stabilization of inventory replenishment for
restaurants using online food delivery platforms, both at the individual platform level and overall within
the supply chain.
The remainder of this study presents a detailed analysis of the proposed approach. A review of

2
related research on demand forecasting and supply chain management is provided in Section 2. Section
3 outlines the problem description, including the simulation approach, mathematical formulation of the
LSTM model, and strategies for reducing the bullwhip effect. This is followed by an explanation of
the methodology used to develop the two phase LSTM model, optimize its hyperparameters using Grid
Search, and preprocess the dataset in Section 4. In Section 5, the numerical setup and process are
discussed, including all numerical values used. Section 6 presents the results and discussion, analyzing
the impact of the proposed model on reducing the bullwhip effect and improving inventory stability.
Finally, in Section 7, the study concludes with a summary of key findings and managerial implications,
outlining actionable steps for companies to implement the proposed model.

2 Literature review
In this section, the literature in four different but related streams of our proposed model is reviewed,
namely, (i) bullwhip effect in SCM, (ii) data simulation for bullwhip effect analysis (iii) online food
delivery applications and challenges, and (iv) demand forecasting techniques.

2.1 Bullwhip effect in SCM

Over the past few decades, the bullwhip effect on product orders has become a widely studied topic
among researchers and practitioners. Early studies attempted to demonstrate the existence of the bull-
whip effect on product orders and identify its underlying causes (Forrester, 1997, 2012). The bullwhip
effect is a phenomenon in which minor variations in customer demand lead to progressively larger fluc-
tuations in orders as they propagate upstream in the supply chain (Lee et al., 1997a). Currently, most
research focuses on quantifying the bullwhip effect and exploring strategies to mitigate its impact. Lee
et al. (2000) provided a formal definition of the bullwhip effect on product orders and systematically
analyzed four main causes: demand signal processing, order batching, price fluctuations, and rationing.
Additionally, they proposed several countermeasures, including reducing frequent updates of demand
forecasts, breaking order batching, stabilizing prices, and eliminating gaming behavior in shortage situa-
tions (Lee et al., 1997a,b). Several studies have emphasized the role of demand forecasting in influencing
the bullwhip effect (Sodhi et al., 2012). Researchers have also identified the bullwhip effect across various
industries, including retail (Yao and Dresner, 2008; Ramanathan, 2014), automotive (Bai and Sarkis,
2020), healthcare and pharmaceuticals (Costantino et al., 2015), electronics and semiconductor man-
ufacturing (Lee et al., 1997b; Chen and Lee, 2012) and food supply chains (Yang et al., 2021). Over
time, various strategies have been developed to mitigate the bullwhip effect. Early studies identified key
causes and solutions, such as information sharing (Lee et al., 1997a), emphasizing accurate demand data
to reduce uncertainty. Order batching reduction emerged as another approach, as Lee et al. (1997a)
advocated just-in-time replenishment, later supported by Peng and Xiao (2014) for Haier Group’s case.
Price stabilization was recommended by Lee et al. (1997a) to minimize demand spikes. In the 2000s,
demand forecasting improvement was explored by Chen et al. (2000), showing how better forecasting
reduces variability, while lead time reduction was highlighted by Lee et al. (1997a) for its role in faster
demand response. In the 2010s, vendor-managed inventory (VMI) gained attraction when Ravichandran
(2008) proved its effectiveness in consumer durables and also in semiconductors (Diaz et al., 2022). Agile
and lean practices were also studied, with Peng and Xiao (2014) demonstrating their success at Haier.
Additionally, modern forecasting models improved with Li and Dörfler (2024) by applying robust con-
trol theory to minimize order fluctuations. These studies illustrate a progressive shift from fundamental
supply chain coordination mechanisms to sophisticated, data-driven, and technology-enhanced solutions
for bullwhip effect reduction.

2.2 Data simulation for bullwhip effect analysis

Simulation techniques play a crucial role in studying and mitigating the bullwhip effect by replicat-
ing real-world supply chain dynamics. Various studies have explored simulation techniques to miti-
gate the bullwhip effect. System dynamics simulations were used by Sterman (1989) and Dejonckheere
et al. (2003) to model demand amplification and optimize replenishment policies. Agent-based modeling
(ABM) was applied by Nair and Vidal (2011) to analyze decentralized decision-making impacts on de-
mand variability. Monte Carlo simulations and bootstrapping have been used for sales data generation,
with Sterman (1989) applying them to retail forecasting. GANs were introduced by Tkachuk et al. (2025)
to simulate realistic sales patterns. Time-series simulations often use ARIMA models (Zhang, 2003) or

3
stochastic distributions like Poisson and Gaussian (Syntetos and Boylan, 2005). While Chatfield et al.
(2004) used DES to measure demand amplification under different inventory policies. Wangphanich et al.
(2010) integrated stochastic demand patterns into DES to assess forecasting impacts. In our study, we
adopt DES with stochastic distributions to model demand and lead time variability, allowing for a more
realistic assessment.

2.3 Online food delivery applications and Challenges

Online food delivery (OFD) applications have revolutionized the food industry, improving convenience
but also introducing significant supply chain challenges. Chopra and Meindl (2001) highlighted the im-
portance of real-time tracking to enhance supply chain efficiency, while Hwang and Kim (2020) explored
drone-based delivery as a potential future solution. Demand fluctuations create inefficiencies, with sudden
order surges leading to stockouts and delays, as noted by Suali et al. (2024). However, through machine
learning forecasting accuracy has been improved (Aamer et al., 2020). Logistics and last-mile delivery
constraints, such as traffic congestion and workforce shortages, impact delivery speed, but predictive
analytics and dynamic routing provide optimization strategies (Chu et al., 2023). Inventory manage-
ment is another concern, with perishable food wastage resulting from inaccurate predictions (Osman
et al., 2023). Environmental concerns, including excessive packaging waste and carbon emissions, have
led to sustainability initiatives such as eco-friendly packaging and electric vehicles (Maheshwari et al.,
2021), with regulatory policies further promoting green delivery practices (Pal, 2023). Technological
advancements like blockchain and IoT are also transforming supply chains, with blockchain improving
transparency and traceability (Zhang and Wei, 2022) and IoT-enabled temperature monitoring ensuring
food safety in transit (Mohsin and Yellampalli, 2017). These studies highlight the complexities of OFD
supply chains and the role of innovation in addressing them. This problem is particularly pronounced on
competitive food delivery platforms, where third-party logistics (3PL) providers serve as competitive de-
mand aggregators (CDAs), connecting B2C customers with B2B restaurant suppliers. Unlike traditional
supply chains, food delivery networks face highly volatile demand patterns, short fulfillment cycles, and
perishable inventory, necessitating more advanced forecasting and inventory optimization techniques.

2.4 Demand forecasting techniques

Effective demand forecasting and inventory smoothing are critical for maintaining supply chain efficiency,
reducing costs, and improving customer satisfaction. Various techniques, ranging from traditional statis-
tical models to advanced machine learning approaches, have been explored to optimize inventory man-
agement. To improve forecasting accuracy and mitigate demand variability, researchers have explored
more sophisticated forecasting methods such as ARIMAX models, which integrate external factors into
forecasting (Assad et al., 2023; Babai et al., 2022). More recently, machine learning-based forecast-
ing techniques have been explored for their potential to reduce the bullwhip effect. Feizabadi (2022)
showed that machine learning models outperform traditional statistical approaches while Paruthipattu
(2021) compared AI-based techniques such as Random Forest, XGBoost, and LSTM, concluding that
they provide superior accuracy in demand prediction. Sarıcıoğlu et al. (2024) further investigated LSTM
networks, demonstrating their effectiveness in mitigating supply chain fluctuations. Other studies, such
as those by Rezki and Mansouri (2024), explored hybrid models integrating deep learning with sta-
tistical forecasting, while Kiuchi (2024) examined reinforcement learning algorithms, highlighting their
adaptability to dynamic demand patterns (Borucka, 2023; Zhu et al., 2021).
The above literature review demonstrates effective supply chain management, and the evolution of
techniques has mitigated the issues starting from the very beginning of traditional supply chain models.
However, managing the current perishable food supply chains, which are highly volatile and competitive
in the modern era, is a big challenge. Therefore, as motivations to develop the current model, the
following key research gaps are highlighted:
• Many studies using simulation-based approaches primarily focus on long-term demand trends while
overlooking short-term variations, seasonal changes, and event-driven spikes (e.g., holidays, pro-
motions, or sudden market shifts).
• Advanced forecasting techniques, particularly deep learning-based models such as LSTM and rein-
forcement learning, often require high computational power and careful hyperparameter tuning.
• Limited integration of AI-driven forecasting models with adaptive inventory control systems tai-
lored for perishable goods.

4
This paper addresses these challenges by simulating sales data with short-term fluctuations, seasonal
variations, and event-driven demand spikes. Unlike studies focusing only on long-term trends, our
LSTM-based model captures both short- and long-term patterns through optimized hyperparameter
tuning, enhancing forecasting accuracy and mitigating the bullwhip effect. Table 1 summarizes key
recent papers in each of these streams, highlighting their contributions and relevance to our study

Deep Learning Bullwhip Ef- Online Food

Author(s) Simulation
Forecasting fects Delivery
Viloria et al. (2019) ✓
Alabdulkarim (2020) ✓ ✓
(Aamer et al., 2020) ✓ ✓
Chiadamrong and Sarnrak (2021) ✓ ✓
Ramírez (2021) ✓ ✓
Yang et al. (2022) ✓ ✓
Tjen et al. (2023) ✓ ✓
Udenio et al. (2023) ✓ ✓
Rezaeefard et al. (2024) ✓ ✓
Kleinemolen (2024) ✓ ✓
Chen et al. (2024) ✓ ✓
This study ✓ ✓ ✓ ✓

Table 1: Comparison of our study with the existing literature

Our study addresses these gaps by proposing a novel AI-enhanced framework that integrates pre-
dictive demand analytics with adaptive inventory management strategies to enhance the stability and
efficiency of online food delivery supply chains while mitigating bullwhip effects.

3 Problem description
This supply chain model follows an online food delivery system, integrating key stakeholders: restaurants
(as suppliers and service providers), third-party logistics (3PLs), customers, and the online platform as
the central coordinator.
The process begins with restaurants managing inventory and sourcing ingredients. When a customer
places an order through the online app, the platform directs it to the restaurant. The restaurant prepares
the meal and hands it over to a 3PL delivery executive, who ensures timely delivery. Once the customer
receives the order, they provide feedback via the app, contributing to service improvements for all
stakeholders, as shown in Figure 1.

5
Figure 1: Supply chain model for online food delivery platforms representing the interactions between
customers, Third-party logistics providers (3PLs), and restaurant suppliers

This section outlines the supply chain model developed to address the bullwhip effect in competitive
online food delivery platforms. The proposed framework integrates three key components: discrete
event simulation (DES), long short-term memory (LSTM) for demand forecasting, and bullwhip effect
calculation.

3.1 Discrete event simulation (DES)

Discrete event simulation (DES) is a modeling technique used to analyze complex systems where events
occur at distinct points in time. It is widely applied in supply chain and logistics management to sim-
ulate real-world operations and evaluate system performance under varying conditions. In this study,
DES is employed to replicate and analyze the operations of two competitive online food delivery plat-
forms, accounting for demand fluctuations, dynamic pricing, lead times, and external disruptions such as
weather, holidays, and special events. DES is particularly useful in this context as it captures dynamic
interactions between multiple stakeholders, examines how fluctuations in demand impact order fulfill-
ment and inventory levels, and evaluates the bullwhip effect. Additionally, it provides valuable insights
for optimizing pricing strategies, delivery efficiency, and inventory management, ultimately helping to
develop strategies that mitigate demand uncertainties and improve overall supply chain resilience.

3.1.1 Simulation framework

The simulation models the demand for Swiggy and Zomato, two leading competating online food delivery
apps, based on patterns from two established Kaggle datasets 1 . It operates on a time-based framework
from January 2023 to January 2025, where events are triggered at five specific timesteps daily: morning,
noon, evening, night, and midnight. The simulation runs continuously, generating demand for a 24-hour
cycle using SimPy.

3.1.2 Key assumptions

This simulation model is built upon several theoretical assumptions that govern the demand patterns,
pricing mechanisms, and supply chain dynamics.
a) Demand functions
1 https://www.kaggle.com/datasets/bhanupratapbiswas/food-delivery-time-prediction-case-study/data,

https://www.kaggle.com/datasets/cbhavik/swiggyzomato-order-information

6
 • Long-term trends and seasonality: The demand for food delivery services follows both a long-
term trend (growth or decline) and seasonal variations across five distinct periods.
• Cyclical demand patterns: The fluctuation of demand throughout the day is modeled using a
sinusoidal function 2 : π 
C(t) = Bphase sin t+c +h (1)
P
where Bphase represents the phase amplitude, P controls periodicity, c is a phase shift constant,
and h is a baseline adjustment for both swiggy and zomato.
• Seasonality factors (S(t)): Various external factors contribute to fluctuations in demand,
including
– Time of day: Demand varies significantly across different hours, peaking during meal
times.3
– Weather conditions: Extreme or mild weather conditions influence customer ordering
behavior.4
– Holidays and special events: Demand experiences spikes during major holidays and cul-
tural festivals.5
• Stochastic demand across platforms: The model accounts for two distinct platforms, each
exhibiting its own stochastic demand curve with random noise, following a normal distribution
to reflect real-world variability.
b) Market and customer segmentation: Different customer groups exhibit distinct ordering patterns,
influenced by factors such as demographics, purchasing power, and regional preferences.
c) Dynamic pricing mechanism: Pricing is dynamically adjusted based on market trends, platform-
specific pricing strategies, and demand-supply equilibrium.
d) Order fulfillment and logistics: The order fulfillment process is modeled using an M/G/∞ queuing
system, where order arrival rate follows an exponential distribution (M), capturing the randomness
of incoming orders. Service time (G) accounts for both food preparation and delivery duration,
where preparation time is assumed to follow a lognormal distribution and delivery time follows a
uniform distribution with infinite servers.
e) Inventory uncertainty: The availability of food ingredients fluctuates based on demand patterns,
supplier restocking schedules, and external disruptions.

3.1.3 Mathematical Formulation

The demand for food delivery platforms is influenced by price, lead time, and external factors. The
proposed mixed demand model, as stated in (Chopra and Meindl, 2007), captures these dynamics as
given below.
a) Demand Function
The demand functions of the two food delivery platforms are given below

DZ (ti ) = αZ (t) − (βZ (ti ) · PZ (ti )) − (γ(t) · LZ (ti ))
(2)

+ τ (t)(PS (ti ) − PZ (ti )) − δ(t)(LZ (ti ) − LS (ti )) + ϵ(t) × F (t)


DS (ti ) = αS (t) − (βS (ti ) · PS (ti )) − (γ(t) · LS (ti ))
(3)

+ τ (t)(PZ (ti ) − PS (ti )) − δ(t)(LS (ti ) − LZ (ti )) + ϵ(t) × F (t)

where Z refers Zomato and S refers Swiggy and

– t = The number of days the dataset is simulating
2 https://mtimpa.weebly.com/uploads/1/9/3/5/19359645/chap06.2.pdf
3 https://tfresource.org/topics/Model_Validation_and_Reasonableness_Checking_Time_Of_Day.html
4 https://climatechange.chicago.gov/climate-impacts/climate-impacts-agriculture-and-food-supply
5 https://www.sonwillogistics.com/unwrapping-the-impact-of-the-holiday-season-on-the-food-industry/

7
 – ti = Each day taking 5 times (i= morning, noon, evening, night, midnight)
– DZ (ti ), DS (ti ) = Demand for each platform
– αZ (t), αS (t) = Baseline demands
– βZ (ti ), βS (ti ) = Price sensitivity coefficients
– PZ (ti ), PS (ti ) = Platform prices for different food
– τ (t), γ(t), δ(t) = Price competition intensity factor, lead time sensitivity coefficient, lead time
competition intensity factor
– LZ (ti ), LS (ti ) = Platform lead times
– ϵ(t) = Random Noise follows Normal(µ(t), σ 2 (t))
– T(t)= Time of the day
– E(t)= Event or holiday effect
– βcustomer = Customer segmentation
– F(t) = Cumulative effect of seasonal and external factors

F (t) = T (t) × W (t) × E(t) × βcustomer

Taking the expectation on both sides

E[DZ (ti )] = αZ (t) − βZ (ti )PZ (ti ) − γ(t)E[LZ (ti )]
+ τ (t)(PS (ti ) − PZ (ti )) − δ(t)(E[LZ (ti )] − E[LS (ti )]) (4)

+ E[ϵ(t)] × F (t)

E[DS (ti )] = αS (t) − βS (ti )PS (ti ) − γ(t)E[LS (ti )]
+ τ (t)(PZ (ti ) − PS (ti )) − δ(t)(E[LS (ti )] − E[LZ (ti )]) (5)

+ E[ϵ(t)] × F (t)
and the variances are

Var(DZ (ti )) = (γ 2 (t)+δ 2 (t))Var(LZ (ti ))+δ 2 (t)Var(LS (ti ))−2δ(t)Cov(LZ (ti ), LS (ti ))+σ 2 (t) (6)

Var(DS (ti )) = (γ 2 (t)+δ 2 (t))Var(LS (ti ))+δ 2 (t)Var(LZ (ti ))−2δ(t)Cov(LZ (ti ), LS (ti ))+σ 2 (t) (7)

Here LZ (ti ) and LS (ti ) are uncorrelated as they are lead time of two different platforms.
b) Lead time function
The lead time for an order on a platform (e.g., zomato, swiggy) is defined as:

Lk (ti ) = Ts,k + Tr,k (8)

where:
- Ts,k = Time when the order delivery is fulfilled to the customer.
- Tr,k = Time when the restaurant accepts the customer’s order.
- k= swiggy, zomato
We are using the M/G/∞ model as it effectively represents the dynamics of food delivery systems.
These systems experience random order arrivals and variable service times while ensuring that
each order is assigned to a delivery executive without delay. This model captures the following
components
– Markovian or poisson process (M): The orders arrive randomly over time, and the inter-arrival
times Tr,i follow an exponential distribution with rate λi

P (Tr,i > t) = e−λi t , t≥0 (9)

8
 – General Service Time Distribution (G): The total service time for an order consists of two
components: the food preparation time and the delivery time. This can be expressed as

Ts,i = Tp,i + Td,i (10)

Where (Tp,i ) represents the food preparation time and (Td,i ) represents the delivery time. The
food preparation time(Tp,i ) is assumed to follow a lognormal distribution. This choice is based
on the positive skewness typically observed in cooking times, which arise due to variability in
preparation speed, human involvement, and different cooking methods (Thecharmsofs et al.,
2001). The lognormal distribution is defined as

Tp,i ∼ Lognormal(µi , σi ) (11)

Where µi , σi are the parameters of the lognormal distribution, representing the mean and
standard deviation of the logarithm of the preparation times, respectively. The probability
density function (PDF) for the lognormal distribution is:

(ln t − µi )2
 
1
fTp,i (t) = √ exp − , t>0 (12)
tσi 2π 2σi2

The cumulative distribution function (CDF) for this distribution is given by

 
1 1 ln t − µi
FTp (t) = + erf √ (13)
2 2 σi 2

Next, the delivery time (Td ) is assumed to be influenced by the distance between the restau-
rant and the customer. Given that delivery times are bounded within a certain range, we
model Td using a uniform distribution. This reflects the fact that the delivery time can vary
depending on the distance but is limited by geographical constraints and operational condi-
tions(Christopher, 2022). Thus, we assume

Td,i ∼ Uniform(ai , bi ) (14)

For a uniform distribution, the probability density function (PDF) is

1
fTd,i (t) = , a i ≤ t ≤ bi (15)
bi − ai
where ai and bi are the lower and upper bounds of the delivery time, respectively.
The cumulative distribution function (CDF) for the uniform distribution is

t − ai
FTd,i (t) = , a i ≤ t ≤ bi (16)
bi − ai
– ∞ (Infinite Servers): customers wait if all servers are busy in traditional queues. However, in
food delivery, each order is immediately assigned to a delivery executive as soon as it is ready.
Since there is no limit on the number of delivery executives available, we assume an infinite
number of servers. This prevents congestion and ensures smooth operation.
c) Price function
A real-time dynamic pricing strategy is formulated as

Pi = (Pb × Gi ) + Fi + Bi + (Ri × Di ) + Si (17)

Where - i= swiggy, zomato

- Pb = Base prices for for food as per Indian chart
- Gi = GST multiplier
- Fi = Platform fees for zomato and swiggy
- Bi = Base delivery fees for zomato and swiggy
- Ri = Per km delivery rate

9
 - Di = Distance from restaurant to customer house
-
Small_order_fee, if Pb ≤ threshold

Si =
0, otherwise
This pricing formula is obtained directly from the platform and reflects the real-time dynamic
pricing model used by food delivery aggregators.
d) Restaurant inventory ordering
Perishable food ingredients have a limited shelf life, making inventory management a critical chal-
lenge. To minimize the risk of overstocking while ensuring sufficient supply, the newsvendor model
provides an effective approach (Hadley and Whitin, 1963). In our case, demand originates from
two independent sources, such as swiggy and zomato. This study employs a two-phase inventory
management approach using the newsvendor model, integrating both historical demand-based in-
ventory calculation and LSTM forecast-driven optimization. The methodology is divided into two
cases
– Case 1: Inventory Calculation Using 5-Time-Per-Day Demand Predictions
– Case 2: Inventory Calculation Using Daily Average Demand Predictions
In case 1 at each time step ti (where i= represents morning, noon, evening, night, and midnight),
the total demand is given by

D(ti ) = DZ (ti ) + DS (ti ) (18)

Since demand follows a distribution, the expected total demand (mean) is

µ5-time, total = E[DZ (ti )] + E[DS (ti )] (19)

The total standard deviation is

q
σ5-time,total = 2
σZ(t i)
2
+ σS(t i)
+ 2ρσZ(ti ) σS(ti ) (20)

where:
– σZ(ti ) and σS(ti ) are the standard deviations of Zomato and Swiggy demands for each time of
a day.
– ρ is the correlation coefficient between DZ (ti ) and DS (ti ).
The inventory level is determined as Hadley and Whitin (1963)

Q∗5-time = µ5-time,total + z · σ5-time,total (21)

where z is the Z score of the standard normal distribution.
In case 2, inventory is updated once per day, using the daily average demand instead of frequent
intra-day updates.
5
1X
µdaily = µ5-time,total (22)
5 t=1

The daily standard deviation is

5
1X
σdaily = σ5-time,total (23)
5 t=1

Thus, the daily inventory level follows

Q∗daily = µdaily + z · σdaily (24)

After performing the two-phase LSTM approach, the inventory is updated as follows:

10
 For intra-day demand predictions using Phase 1 LSTM, the inventory is updated at each time step
ti :

Q∗5-time,lstm = µ5-time,total lstm + z · σ5-time, total lstm (25)

where the demand at each time ti is updated using LSTM predictions:

D(ti ) → D(ti , lstm) (26)

For the daily-level demand prediction using Phase 2 LSTM, the inventory is updated as:

Q∗daily,lstm = µdaily lstm + z · σdaily lstm (27)

where the daily demand prediction is computed as

5
1X
D(t, lstm) = D(ti , lstm) (28)
5 i=1

This two-phase approach ensures that inventory is dynamically adjusted based on intra-day fluc-
tuations while maintaining a stable daily-level forecast.

3.2 Long short-term Mmemory (LSTM) model architecture

LSTM networks are a specialized type of RNN designed to capture both long- and short-term dependen-
cies. Since its introduction by Hochreiter and Schmidhuber in 1997(Hochreiter and Schmidhuber, 1997),
they have been refined and widely adopted. While traditional RNNs can theoretically model long-term
dependencies, their effectiveness is limited due to issues like vanishing gradients.

(a) LSTM RNN architecture. A single memory block is (b) LSTM RNN architectures.
shown for clarity.

Figure 2: Illustration of LSTM architectures.

LSTMs address this limitation by incorporating memory cells and a gating mechanism that selectively
retains or discards information. These gates help regulate the flow of historical information, ensuring
relevant past data is preserved while redundant details are discarded. Figures 2a & 2b depict the
description of the LSTM memory block to capture sequential dependencies in demand forecasting follows

11
that of Sak et al. (2014). The specific steps of the LSTM algorithm are as follows:
it = σ(Wi ht−1 + Wx Xt + bi ) (Input Gate)
ft = σ(Wf ht−1 + Wx Xt + bf ) (Forget Gate)
ot = σ(Wo ht−1 + Wx Xt + bo ) (Output Gate)
c̃t = tanh(Wc ht−1 + Wx Xt + bc ) (Candidate Cell State)
ct = ft · ct−1 + it · c̃t (Final Cell State Update)
ht = ot · tanh(ct ) (Hidden State Update)
where,
• it , ft , ot are the input, forget, and output gate activations.
• ht represents the hidden state at time t.
• ct is the cell state at time t, while c̃t is the candidate cell state.
• Wi , Wf , Wo , Wc , Wx are weight matrices governing transformations.
• bi , bf , bo , bc are bias terms.
The hyperbolic tangent activation function is given by

ex − e−x
tanh(x) = (29)
ex + e−x
Here, we use a two-phase LSTM approach.
• Phase 1 (Intraday forecasting): The model takes a sequence of five time steps within a single day
and predicts the next day’s five time steps. This phase helps in capturing short-term variations in
demand that occur within a single day.
• Phase 2 (Daily forecasting): The model takes sequences spanning six days and predicts the demand
for the next day. This phase is essential for capturing long-term temporal patterns and trends across
multiple days.
In both phases, the LSTM processes sequential data, leveraging its memory cell to store and update
information effectively. The intraday phase ensures that short-term fluctuations are well modeled, while
the daily phase allows the model to account for broader trends and cyclical patterns in demand.

3.3 Measure of bullwhip effect

The bullwhip effect refers to the amplification of demand variability as it moves upstream in the supply
chain, leading to inefficiencies such as excessive inventory holding, stockouts, and unpredictable order
fluctuations. This effect is quantified using the bullwhip coefficient B, defined as (Rezaeefard et al., 2024)
σI2
B=
σd2
where σI2 represents the variance of inventory present at the restaurant and σd2 represents the variance
of customer demand (either for an individual platform or total across platforms).
If B > 1, demand amplification is occurring, leading to inefficiencies such as excessive inventory,
stockouts, and unstable order patterns. If B < 1, demand is well-regulated, meaning order variability is
lower than demand variability, resulting in a stable supply chain.
In the context of a restaurant using Swiggy and Zomato as selling platforms, minimizing the bullwhip
effect is crucial to maintaining a stable and efficient supply chain. The bullwhip coefficients for each
platform are given by
2 2
σI,Zomato σI,Swiggy
BZomato = 2 , BSwiggy = 2
σd,Zomato σd,Swiggy
The overall bullwhip coefficient across both platforms is
2
σI,Total
BOverall = 2
σd,Total
The objective is to minimize B at both individual and overall levels (such that B < 1), so that demand
amplification is reduced and a stable, efficient order fulfillment system is maintained.

12
 Figure 3: Flowchart of the proposed method

4 Proposed methodology
Designing a deep learning model for time-series forecasting in an online food delivery application requires
a structured approach to handle demand fluctuations efficiently. Selecting the optimal forecasting model
involves leveraging modern deep learning algorithms to capture complex demand patterns.
The proposed methodology for food order demand forecasting is illustrated in Figure 3), outlining the
key steps involved in predicting order volumes, optimizing inventory, and enhancing delivery efficiency.

4.1 Data preprocessing

The data used in this study represents the demand history of an online food delivery service, specifically
for Zomato and Swiggy. The time series dataset spans from January 2023 to January 2025, capturing
fluctuations in order volumes over this period. Before delving into the analysis of the time series, it is
essential to properly refine and structure the raw data. In this study, feature selection, category encoding,
standardization, and additional feature selection were performed as key preprocessing steps to ensure
the dataset is well-prepared for accurate demand forecasting.

4.1.1 Categorical encoding

Machine learning models require numerical inputs, but categorical variables contain valuable information.
Encoding these variables prevents data loss while ensuring they are properly interpreted by the model. In
our model, we applied one-hot encoding to categorical features such as weather conditions (sunny, rainy,

13
cloudy, etc.) and food categories (fast food, desserts, beverages, etc.), as these are nominal variables
with no inherent order.
For one-hot encoding, each categorical variable is transformed into a vector of binary values

Xencoded = {x1 , x2 , . . . , xn }, xi ∈ {0, 1} (30)

For event importance (high, medium, low), we used ordinal encoding since it follows a meaningful order.
We assigned numerical values as follows

2, if high

Xevent = 1, if medium
0, if low



4.1.2 Feature selection

Selecting relevant features improves model performance by reducing noise, lowering computational com-
plexity, and preventing overfitting. In our model, the dependent variable (Y ) represents demand, while
the independent variables (X) were chosen based on their impact on order forecasting for both Swiggy
and Zomato

X = {Pi (t), Li (t), ∆i (t), E(t)} (31)

where Pi (t) represents the price of the food item, Li (t) denotes the lead time, ∆i (t) denotes the
delivery distance (i=Swiggy, Zomato), and E(t) captures external factors, including the public holidays,
special events, Weather conditions. These features were selected separately for Swiggy and Zomato to
account for platform-specific variations while ensuring an optimal demand forecasting model.

4.1.3 Standardization
The goal of standardization is to scale the values of the source dataset to have a mean of 0 and a standard
deviation of 1, which improves the stability and efficiency of the learning process. To achieve this, we
used the StandardScaler method from the scikit-learn Python library.
Let a time series of length N be represented as Vi , i = 1, 2, . . . , N . The equation for standardization
is as follows
Vi − µ
Vi′ = (32)
σ
where Vi′ represents the standardized value, Vi denotes the observed values in the dataset, µ is the mean
of V , and σ is the standard deviation of V .

4.2 Modeling
4.2.1 Structuring time series data for supervised learning
To adapt time series data for supervised learning, it is first transformed into structured instances with
predefined input and output attributes. These instances are then divided into separate training and
test sets to facilitate model learning and evaluation. This step ensures the construction of appropriate
datasets for training, testing, and validation.

4.2.2 Sequential data preparation for LSTM

LSTM networks are designed to handle sequential data, making them suitable for time series forecasting.
To effectively utilize LSTM, we need to transform the dataset into sequences of fixed-length time steps.
To effectively train an LSTM model for time series forecasting, we structure the dataset in two phases,
each designed to capture different levels of temporal dependencies.
• Phase 1 (Intraday demand sequence for LSTM)
In this phase, we represent time series demand data in structured daily groups, where each day
consists of five discrete time slots. Let the demand for the day be represented as

Vd = {Vd,1 , Vd,2 , Vd,3 , Vd,4 , Vd,5 }

14
 where Vd,t denotes the demand at time slot t on day d. To construct overlapping sequences
for training and prediction, we segment the data into input-output pairs using a sliding window
approach over n days.

Xd = [Vd,1 , Vd,2 , Vd,3 , Vd,4 , Vd,5 , Vd+1,1 , . . . Vd+n−1,5 ]

where Xd captures a total of 5n demand values. The corresponding target sequence Yd represents
the demand values for the next day (d+n) across all five time slots

Yd = [Vd+n,1 , Vd+n,2 , Vd+n,3 , Vd+n,4 , Vd+n,5 ]

This approach ensures that each input Xd includes sufficient historical context, while the target Yd
provides the full demand pattern for the subsequent day.
• Phase 2 (Aggregated daily demand for LSTM training)
In this phase, we aggregate the demand values from the five discrete time slots into a single daily
demand value. This transformation simplifies the time series representation, making it suitable for
models that predict daily demand trends rather than intra-day variations.
To achieve this, we compute the average demand for each day

5
1X
Vdavg = Vd,t
5 t=1

where Vdavg represents the average demand for day d.

After transforming the time series into daily averages, we prepare the dataset for LSTM training
using a sliding window approach over six days. The input sequence for each training instance
consists of the average demand over six consecutive days

Xd = [Vdavg , Vd+1
avg avg
, Vd+2 avg
, Vd+3 avg
, Vd+4 avg
, Vd+5 ]

The corresponding target value represents the predicted daily demand for the seventh day

avg
Yd = Vd+6

This approach ensures that the LSTM model captures longer-term temporal dependencies by learn-
ing from daily aggregated patterns. It is particularly useful for forecasting overall demand trends
rather than fluctuations within a day.

4.2.3 Splitting data into training, testing, and validation sets

When dealing with extensive datasets, it is beneficial to divide the data into three distinct subsets: the
training set (train the model by learning demand patterns from past data), the validation set (fine-
tune the model by selecting the optimal hyperparameters and preventing overfitting), and the test set
(evaluation on unseen data).

4.2.4 Hyperparameter optimization

Selecting the most suitable model configuration is a crucial and time-intensive process. Optimizing
hyperparameters enhances model performance, improves generalization, and reduces reliance on manual
trial and error. Since hyperparameters are independent of the dataset, they act as control settings that
influence how the model learns and adapts.
A common approach for tuning hyperparameters is grid search, which systematically explores differ-
ent parameter combinations to identify the best-performing configuration. However, as the number of
parameters increases, grid search can become computationally expensive. In this study, we optimize the
LSTM model using the following hyperparameters

• Epochs (E): The number of times the entire dataset is passed through the model during training,
which helps in adjusting model weights and improving learning.

15
 • LSTM units (U ): The number of neurons in the LSTM layer, which determines the model’s ability
to capture temporal dependencies in data.
• Batch size (B): The number of samples processed before updating model weights, influencing
training speed and stability.

• Dropout rate (D): A regularization technique that randomly deactivates a fraction of neurons
during training to prevent overfitting.
• Learning rate (α): The step size at which the model updates weights during training, affecting
convergence speed and accuracy.

• Optimizer: The optimization algorithm used to update model weights and minimize errors. We
use the Adam optimizer to efficiently adjust weights and reduce the mean squared error (MSE).
• Layer configuration: The depth of the LSTM network, affecting its ability to capture complex
patterns. Our model ranges from a monolayer to multiple stacked layers.

• Training time: The total time required to train the model, which depends on the dataset size,
network complexity, and computational resources.

By tuning these hyperparameters, we aim to improve the LSTM model’s ability to predict demand
accurately while maintaining computational efficiency.

4.2.5 Performance assessment

The predicted model is evaluated using a test dataset separately for both phases. The performance
metrics are computed for both Swiggy and Zomato demands :

a) Root mean squared error (RMSE) measures the average magnitude of prediction errors, providing
an indicator of model accuracy:
v
u n
u1 X
RM SE = t (Yi − Ŷi )2 (33)
n i=1

A lower RMSE value indicates better model performance.

b) Mean absolute error (MAE) measures the average absolute difference between actual and predicted
values, providing a straightforward interpretation of prediction errors
n
1X
M AE = Yi − Ŷi (34)
n i=1

A lower MAE indicates better model performance, as it represents smaller deviations between
predictions and actual values.
c) R-Squared (R2 ) score evaluates the proportion of variance in the dependent variable explained by
the model
(Yi − Ŷi )2
P
R2 = 1 − P (35)
(Yi − Ȳ )2
where Ȳ is the mean of the actual values. An R2 value closer to 1 indicates a better fit of the
model to the data.

5 Experimental framework
This section analyses the numerical setup and experimental process of the supply chain model.

16
5.1 Numerical Setup
The dataset used in this study simulates online food delivery demand forecasting for Zomato and Swiggy,
covering the period from January 2023 to January 2025. Demand is recorded at five time intervals per
day—morning, noon, evening, night, and midnight—resulting in a total of 3,675 observations across 19
variables. It includes temporal attributes such as the week, date, day, and time of order placement, allow-
ing for an analysis of time-dependent demand variations. Order characteristics, including food category,
price, and customer demand, provide insights into consumer purchasing patterns on both platforms.
Delivery-related metrics, such as lead time and delivery distance, reflect logistical constraints and service
efficiency. Additionally, the dataset incorporates external factors that impact demand, including supplier
inventory levels, public holidays, special events categorized by importance (high, medium, or low), and
weather conditions.
For the cyclical demand pattern in Equation (1), we assume βphase = 0.475, with P = 5 time periods
per day and a peak shift of c = 1.5 to align with dinner hours. The baseline demand is set at h = 0.525
to reflect realistic daily fluctuations.
The base demand values are defined as αs (t) = 12, 000 for Swiggy and αz (t) = 10, 000 for Zomato,
based on observed market trends. The pricing structure follows average Indian food prices, and price
sensitivity parameters βs (ti ) and βz (ti ) are selected based on empirical studies Andreyeva et al. (2010).
For order arrivals, we assume an M/G/∞ queuing model, with the arrival rate λarrival varying across
demand periods

3 − 10 peak hours

λarrival = 2 − 6 standard times
1 − 3 late-night hours



The service time Tp,i is assumed to follow a log normal distribution with a mean of 30 minutes and a
standard deviation of 10 minutes, reflecting realistic variation in food preparation times. These values
correspond to lognormal parameters µ = 3.35 and σ = 0.32. Delivery time depends on the restaurant-
to-customer distance ∆(ti ), assumed to range between 1 km and 15 km, following a uniform distribution
(in minute) to reflect real world variations.6

U(15, 20) if distance < 5

Td,i = U(30, 35) if 5 ≤ distance < 10
U(40, 45) if distance ≥ 10



Lead time sensitivity τ (t) is taken from prior research Negi and Wood (2019), while competitive factors
γ(t) = 0.5 and δ(t) = 0.5 reflect moderate competition between Swiggy and Zomato.
ϵt follows a normal distribution with a mean of 0 and standard deviation of 20. We use z = 1.96,
the standard safety factor for optimal order quantity calculations. The correlation coefficient ρ = 0.93
represents the relationship between Swiggy and Zomato demand.
Demand fluctuations are influenced by external factors, with extreme weather increasing demand by
40–60% (Liu et al. (2022);Yao et al. (2023)), holidays and special events causing a 20–50% surge 7 , and
loyal customers generating 10–20% more orders 8 compared to others. These variations highlight the
importance of accurate forecasting in optimizing inventory and supply chain management. The detailed
derivation of all model parameters and assumptions is provided in Appendix

5.2 Experiment process

The experiment follows a structured approach to prepare and train the model effectively. The key steps
include the following:

5.2.1 Feature selection and target variable identification

The dependent variable for demand forecasting includes Zomato Demand DZ (ti ) and Swiggy Demand
DS (ti ), representing food orders on each platform. Independent variables include price PZ (ti ), PS (ti ),
6 https://economictimes.indiatimes.com/industry/services/retail/what-indians-ordered-from-food-to-pet-\

care-here-are-the-trending-deliveries-in-2024/articleshow/117236304.cms?
7 https://www.qsrmagazine.com/growth/consumer-trends/whats-next-restaurant-delivery/
8 https://www.adjust.com/blog/hunger-for-food-delivery-apps-grows-in-2022-against-all-odds/

17
lead time LZ (ti ), LS (ti ), and distance ∆Z (ti ), ∆S (ti ). External factors like event importance E(t) and
weather conditions W (t) also influence demand fluctuations.

5.2.2 Exploratory data analysis (EDA)

To analyze demand patterns, we compute the daily average demand to smooth short-term fluctuations:
5
1X
Davg,t = Dt,i (36)
5 i=1
Long-term trends are captured using the cumulative mean:
t
1X
CMt = Davg,j (37)
t j=1

Demand stability is assessed via the 7-day rolling variance, considering deviations from the rolling
mean:
t
1 X
RVt = (Davg,j − D̄avg,t )2 (38)
7 j=t−6

Additionally, we analyze the demand distribution using a histogram to identify skewness and outliers,
and a Q-Q plot to compare the demand distribution against a normal distribution.

5.2.3 Train-validation-test split

For the train-validation-test split, both Swiggy and Zomato demand data are simulated from January 1,
2023, to January 1, 2025. To ensure a proper distribution, we allocate 70% of the data (January 2023
– April 2024) for training, 15% (May 2024 – September 2024) for validation, and the remaining 15%
(October 2024 – January 2025) for testing.

5.2.4 Hyperparameter optimization via GridSearch

Table 2 presents the selected values for each hyperparameter, which will be used for configuring and
training the model in both the phases. These values have been determined based on prior experimentation
and literature review to ensure optimal performance.

Hyperparameter Value Range

Epochs (E) 50 – 200
LSTM Units (U ) 32 – 128
Batch Size (B) 16 – 64
Dropout Rate (D) 0.1 – 0.5
Learning Rate (α) 0.001 – 0.01
Layers 1-3

Table 2: Hyperparameter values for GridSearch

5.2.5 Bullwhip effect measurements

The calculation is performed in three stages: First, the bullwhip effect is computed using the training
dataset, establishing a baseline measurement of demand fluctuations in historical data. Next, the same
calculation is performed on the test dataset for Zomato and Swiggy demand, along with restaurant
inventory for both phases. Lastly, calculate bullwhip effects on the predicted dataset after forecasting
demand, with the inventory adjusted for both individual platforms and overall for both phases.

6 Results and discussion

This section presents the results of the demand forecasting model and its impact on supply chain stability.
The analysis follows a structured approach:

18
6.1 EDA results
Figure 4 presents the cumulative mean and rolling variance of demand for Zomato and Swiggy over time.
These visualizations help in understanding long-term trends and demand fluctuations.

(a) Cumulative mean of demand over time

(b) Rolling variance of demand over time

Figure 4: Cumulative mean and rolling variance of demand for Zomato and Swiggy

The cumulative mean of demand, shown in Figure 4a, highlights the overall trend for both platforms.
Initially, Swiggy exhibits a higher average demand than Zomato, but both platforms stabilize over
time, with Zomato settling at a lower demand level. The rolling variance, presented in Figure 4b,
reveals fluctuations in daily demand. Swiggy exhibits consistently higher volatility compared to Zomato,
indicating greater variability in its order patterns. These findings provide key insights into demand
stability, helping to refine forecasting models. Additionally, to assess the distribution of demand, we plot
the Histogram & QQ Plot given in Figures 5 and 6.

19
 Figure 5: Histogram of demand

Both Zomato and Swiggy demand exhibit a bimodal distribution in Figure, indicating two distinct
demand clusters. The positive skewness suggests occasional periods of significantly higher demand,
requiring strategic resource planning. The QQ plots provide visual evidence that the demand for both

Figure 6: QQ plot of demand

Zomato and Swiggy deviates from a normal distribution, particularly in the tails, suggesting the presence
of more frequent extreme values than a normal distribution would predict.

6.2 Grid search & hyperparameter tuning

For both phase 1 and phase 2, we perform hyperparameter tuning using Grid Search on the validation
dataset, evaluating various hyperparameter combinations mentioned in Table 2 to identify the optimal
configuration that minimizes the validation loss. Figure 7 and Figure 8 suggest model 2 is the best.
It shows the lowest and most stable validation loss over the epochs, indicating it generalizes the best to
unseen data for both phases. Using the best hyperparameter combination, we predict future demand for
both Swiggy and Zomato across both phase 1 and phase 2.

20
 Figure 7: Validation loss curve for phase 1 (intraday demand prediction)

Figure 8: Validation loss curve for phase 2 (daily aggregated demand prediction)

6.3 Performance analysis

Figure 9, 10 present the actual and predicted demand plots for the test dataset, demonstrating that the
LSTM model effectively captures complex nonlinear patterns in each case.

Figure 9: Actual vs. predicted demand for phase 1 (intraday demand prediction)

21
 Figure 10: Actual vs. predicted demand for phase 2 (daily aggregated demand prediction)

After predicting future demand for both Swiggy and Zomato, we analyze the impact on restaurant
inventory. The inventory levels are adjusted based on the forecasted demand to ensure optimal stock
availability while minimizing wastage.

Figure 11: Restaurant inventory levels after predictive demand for phase 1

22
 Figure 12: Restaurant inventory levels after predictive demand for phase 2

Figure 11 and Figure 12 illustrate the restaurant inventory trends after incorporating predictive
demand insights for both phases, respectively. These plots show how LSTM smooths out demand un-
certainty, leading to a more stable inventory that aligns with the predicted demand. Table 3 presents
the evaluation metrics for both phases, which measure the model’s accuracy in predicting demand for
Swiggy and Zomato.

Phase Platform RMSE MAE R2 Training Time(minute)

Zomato 5531.67 4130.31 0.69 12
Phase 1
Swiggy 6768.82 4992.76 0.71 12
Zomato 536.25 414.51 0.88 8
Phase 2
Swiggy 579.40 468.78 0.90 8

Table 3: Evaluation metrics for phase 1 and phase 2

6.4 Measurement of bullwhip effect

We calculate the bullwhip effect for a restaurant through Zomato and Swiggy across different data seg-
ments, including training data, test data, and predicted demand, using LSTM with inventory smoothing.
As shown in Table 4, a high reduction in the bullwhip effect suggests improved demand stability. This
analysis is conducted separately for both phases to evaluate how short-term and long-term demand
variations impact inventory management.

Training Testing Predicted

Phase
Zomato Swiggy Overall Zomato Swiggy Overall Zomato Swiggy Overall

Phase 1 2.88 2.35 2.61 2.06 1.96 2.01 1.02 0.91 0.96

Phase 2 2.32 2.07 2.19 1.97 1.76 1.86 0.72 0.89 0.80

Table 4: Bullwhip effect analysis across phase 1 and phase 2

23
7 Conclusion & managerial insights
The primary objective of this study was to develop an LSTM-based demand forecasting model to improve
both intra-day and aggregated daily demand predictions for restaurants, utilizing real-world sales data
from Swiggy and Zomato. By addressing demand fluctuations, the model aimed to enhance inventory
planning and reduce the bullwhip effect, a common challenge in restaurant supply chains. The results
validate the effectiveness of neural networks in mitigating these inefficiencies, aligning with findings
from (Rezaeefard et al., 2024), which demonstrated the superiority of LSTM over traditional forecasting
models. To achieve optimal performance, the model was fine-tuned using the grid search method, ensuring
it captured the complex patterns and seasonality present in the time-series data. A lower bullwhip effect
indicates a more stable and predictable supply chain, minimizing inefficiencies such as overstocking or
shortages. The results show that in Phase 1, the bullwhip effect decreased from 2.61 to 0.96, with R2
values of 0.69 for Zomato and 0.71 for Swiggy. While this demonstrates a significant reduction in demand
variability, Phase 2 performed even better, with the bullwhip effect further declining from 2.19 to 0.80.
The higher R2 values of 0.88 for Zomato and 0.90 for Swiggy in Phase 2 indicate improved prediction
accuracy, meaning the model was better at explaining demand variations. In conclusion, the results
demonstrate that LSTM-based demand forecasting, combined with inventory smoothing, is an effective
approach to minimizing the bullwhip effect in restaurant supply chains. The proposed model achieved
strong predictive accuracy, and its integration into supply chain management can significantly enhance
efficiency and improve decision-making for restaurants relying on customer orders placed through Swiggy
and Zomato.
The findings of this research have significant practical implications for restaurant operators, supply
chain managers, and online food delivery platforms. By implementing the proposed LSTM-based demand
forecasting model with inventory smoothing, these stakeholders can enable them to make more informed
decisions. For restaurant operators, accurate intraday forecasting ensures optimal ingredient availability
during peak hours while minimizing waste during low-demand periods. At the aggregate daily level, it
enables better procurement planning, reducing last-minute orders and supply chain disruptions. For sup-
ply chain managers, improved intraday demand predictions allow for better synchronization of supplier
deliveries with restaurant needs, reducing storage costs and ensuring a steady supply of perishable goods.
On a daily scale, stable demand forecasts lead to more efficient production planning and cost-effective
logistics, minimizing uncertainty in order quantities. For online food delivery platforms, precise intra-
day demand forecasting helps allocate delivery resources efficiently, reducing wait times and improving
customer satisfaction. At the aggregate level, demand insights optimize dynamic pricing, targeted pro-
motions, and restaurant recommendations, driving higher order volumes and platform engagement. By
leveraging LSTM-based forecasting for both intraday and daily demand, these stakeholders can enhance
efficiency, reduce costs, and improve overall supply chain stability, ensuring long-term sustainability in
the competitive food delivery industry.
While this study demonstrates the effectiveness of LSTM-based forecasting in reducing the bullwhip
effect and improving demand prediction for restaurants, certain limitations remain. One key gap is that
due to the unavailability of industry used data, the study relies on simulated data rather than real-world
transactional data from restaurants and food delivery platforms. Although the simulation effectively
captures realistic demand fluctuations, some assumptions are made intuitively due to the lack of relevant
literature. Additionally, while the model successfully predicts demand and mitigates the bullwhip effect,
it does not optimize inventory replenishment in real time, which is crucial for minimizing costs and
ensuring supply chain efficiency.
To address these gaps, future research should focus on implementing the model using real-world
datasets from restaurants and food delivery platforms to validate its practical effectiveness. Further-
more, integrating reinforcement learning or optimization algorithms could enhance real-time inventory
management by dynamically adjusting order quantities based on demand predictions.

Appendix
A1. Cyclical Demand Patterns
Food delivery demand follows a cyclical pattern, peaking during dinner hours (7 PM – 9 PM) and
reaching its lowest point in the early morning (3 AM – 5 AM). On New Year’s Eve, Swiggy and Zomato

24
processed 9,500 and 4,254 orders per minute, respectively.9 ,10 On a normal day, peak demand is four
times lower, with Swiggy handling 2,375 orders/min and Zomato 1,064 orders/min. Off-peak demand is
approximately 5% of peak demand (Swiggy: 119 orders/min, Zomato: 53 orders/min).
Over a 120-minute peak window, Swiggy processes 285,000 orders, while Zomato processes 127,680
orders. During off-peak periods, order volumes drop to 14,280 and 6,360 orders, respectively. The
amplitude of demand fluctuations is calculated as:
285, 000 − 14, 280 127, 680 − 6, 360
BSwiggy = = 0.475, BZomato = = 0.475 (39)
2 × 285, 000 2 × 127, 680
The corresponding baseline demand levels are:
285, 000 + 14, 280 127, 680 + 6, 360
hSwiggy = = 0.525, hZomato = = 0.525 (40)
2 × 285, 000 2 × 127, 680
Dividing the day into five periods (P = 5) and aligning the peak at dinner time (c = 1.5), the
sinusoidal demand model is given by
π 
γt = 0.475 sin t + 1.5 + 0.525 (41)
5
This formulation ensures that peak demand occurs at dinner time and declines during the early
morning hours.

A2. External demand multipliers

The numerical values for demand multipliers are estimated based on empirical research, market trends,
and food delivery platform reports. These multipliers capture variations due to time of day, weather
conditions, special events, and customer behavior.

Time of day multiplier: 

1.2 − 1.5 for peak hours

T (t) = 1.0 − 1.2 for off-peak hours (42)
0.8 − 1.0 for late-night hours



Weather influence multiplier:


1.4 − 1.6 for extreme weather conditions

W (t) = 1.1 − 1.3 for mild weather (43)
for clear weather

1.0


Holiday and special event multiplier:

(
1.2 − 1.5 for major holidays and special occasions
E(t) = (44)
1.0 for regular days

Customer segmentation multiplier:


0.7 − 0.9 for mismatched preferences

βcustomer = 1.1 − 1.2 for loyal customers (45)
for general customers

1.0


A3. Service time food preparation modeling

The service time Tp,i follows a lognormal distribution with a mean of 30 minutes and a standard deviation
of 10 minutes. The mean and standard deviation of a lognormal distribution are given by the formulas
σ2
m = eµ+ 2 = 30
9 Swiggy receives 9,500 orders per minute on New Year’s Eve - https://www.business-standard.com
10 Zomato records 4,254 orders per minute on New Year’s Evehttps://m.economictimes.com

25
 q
s= (eσ2 − 1)e2µ+σ2 = 10
To determine the lognormal parameters µ and σ, we use the standard transformations. The location
parameter µ is given by:

m2
 
µ = ln √
s2 + m2
Substituting m = 30 and s = 10, we obtain:

302
 
µ = ln √ = ln(28.47) ≈ 3.35
102 + 302
Similarly, the scale parameter is computed as

s2
 
2
σ = ln 1 + 2
m

which evaluates to:

σ 2 = ln(1.111) ≈ 0.105
Taking the square root:
√
σ= 0.105 ≈ 0.32
Thus, the service time is modeled as:

Tp,i ∼ Lognormal(3.35, 0.32)

A4. Baseline Demand and Market Correlation

We simulate daily demand over two years at five time intervals per day. The baseline demand values for
Swiggy and Zomato are determined using their relative market shares. Reports indicate that Swiggy’s
order volume was 1.2 times higher than Zomato’s during this period 11 while Zomato processed over 647
million orders in the financial year.12
Based on these observations, the baseline demand values are defined as

αs (t) = 12, 000, αz (t) = 10, 000 (46)

The correlation coefficient ρ between Swiggy and Zomato demand is calculated as:
P
(Xi − X̄)(Yi − Ȳ )
ρ = pP pP (47)
(Xi − X̄)2 (Yi − Ȳ )2
For this dataset, the computed correlation value is ρ = 0.93, indicating a strong positive relationship
between demand trends for both platforms.

References
Aamer, A., Eka Yani, L. and Alan Priyatna, I. (2020). Data analytics in the supply chain manage-
ment: Review of machine learning applications in demand forecasting, Operations and Supply Chain
Management: An International Journal. 14(1), 1–13.
Abbasimehr, H., Shabani, M. and Yousefi, M. (2020). An optimized model using lstm network for demand
forecasting, Computers & industrial engineering. 143, 106435.

Abolghasemi, M., Beh, E., Tarr, G. and Gerlach, R. (2020). Demand forecasting in supply chain:
The impact of demand volatility in the presence of promotion, Computers & Industrial Engineering.
142, 106380.
11 Swiggy-Zomato Market Share Analysis
12 Zomato Order Statistics (2019-2023)

26
Aktepe, A., Yanık, E. and Ersöz, S. (2021). Demand forecasting application with regression and artifi-
cial intelligence methods in a construction machinery company, Journal of Intelligent Manufacturing.
32(6), 1587–1604.
Alabdulkarim, A. A. (2020). Minimizing the bullwhip effect in a supply chain: a simulation approach
using the beer game, Simulation. 96(9), 737–752.
Amirkolaii, K. N., Baboli, A., Shahzad, M. and Tonadre, R. (2017). Demand forecasting for irregular
demands in business aircraft spare parts supply chains by using artificial intelligence (ai), IFAC-
PapersOnLine. 50(1), 15221–15226.
Andreyeva, T., Long, M. W. and Brownell, K. D. (2010). The impact of food prices on consumption: a
systematic review of research on the price elasticity of demand for food, American journal of public
health. 100(2), 216–222.
Assad, D. B. N., Cara, J. and Ortega-Mier, M. (2023). Comparing short-term univariate and multi-
variate time-series forecasting models in infectious disease outbreak, Bulletin of Mathematical Biology.
85(1), 9.
Babai, M. Z., Boylan, J. E. and Rostami-Tabar, B. (2022). Demand forecasting in supply chains: a review
of aggregation and hierarchical approaches, International Journal of Production Research. 60(1), 324–
348.
Bai, C. and Sarkis, J. (2020). A supply chain transparency and sustainability technology appraisal model
for blockchain technology, International journal of production research. 58(7), 2142–2162.
Blackburn, J. and Scudder, G. (2009). Supply chain strategies for perishable products: the case of fresh
produce, Production and Operations Management. 18(2), 129–137.
Borucka, A. (2023). Seasonal methods of demand forecasting in the supply chain as support for the
company’s sustainable growth, Sustainability. 15(9), 7399.
Catal, C., Fenerci, A., Ozdemir, B. and Gulmez, O. (2015). Improvement of demand forecasting models
with special days, Procedia Computer Science. 59, 262–267.
Chatfield, D. C., Kim, J. G., Harrison, T. P. and Hayya, J. C. (2004). The bullwhip effect—impact of
stochastic lead time, information quality, and information sharing: a simulation study, Production and
operations management. 13(4), 340–353.
Chen, F., Ryan, J. K. and Simchi-Levi, D. (2000). The impact of exponential smoothing forecasts on
the bullwhip effect, Naval Research Logistics (NRL). 47(4), 269–286.
Chen, L. and Lee, H. L. (2012). Bullwhip effect measurement and its implications, Operations research.
60(4), 771–784.
Chen, Q., Zhang, X., Liu, S. and Zhu, W. (2024). The inventory bullwhip effect in the online retail
supply chain considering the price discount based on different forecasting methods, The Engineering
Economist. 69(2), 108–128.
Chiadamrong, N. and Sarnrak, N. (2021). Meta-prediction models for bullwhip effect prediction of a
supply chain using regression analysis, International Journal of Information Systems and Supply Chain
Management (IJISSCM). 14(4), 36–71.
Chopra, S. and Meindl, P. (2001). Strategy, planning, and operation, Supply Chain Management.
15(5), 71–85.
Chopra, S. and Meindl, P. (2007). Supply chain management. Strategy, planning & operation, Springer.
Christopher, M. (2022). Logistics and supply chain management, Pearson Uk.
Chu, H., Zhang, W., Bai, P. and Chen, Y. (2023). Data-driven optimization for last-mile delivery,
Complex & Intelligent Systems. 9(3), 2271–2284.
Costantino, F., Di Gravio, G., Shaban, A. and Tronci, M. (2015). Spc forecasting system to mitigate the
bullwhip effect and inventory variance in supply chains, Expert Systems with Applications. 42(3), 1773–
1787.

27
Dejonckheere, J., Disney, S. M., Lambrecht, M. R. and Towill, D. R. (2003). Measuring and avoiding the
bullwhip effect: A control theoretic approach, European journal of operational research. 147(3), 567–
590.
Diaz, M. F. L., Ehm, H. and Ismail, A. (2022). Simulated-based analysis of recovery actions under
vendor-managed inventory amid black swan disruptions in the semiconductor industry: A case study
from infineon technologies ag, 2022 Winter Simulation Conference (WSC), IEEE, pp. 3513–3524.
Feizabadi, J. (2022). Machine learning demand forecasting and supply chain performance, International
Journal of Logistics Research and Applications. 25(2), 119–142.

Folkerts, H. and Koehorst, H. (1997). Challenges in international food supply chains: vertical co-
ordination in the european agribusiness and food industries, Supply Chain Management: An Interna-
tional Journal. 2(1), 11–14.
Forrester, J. W. (1997). Industrial dynamics, Journal of the Operational Research Society. 48(10), 1037–
1041.

Forrester, J. W. (2012). Industrial dynamics: a major breakthrough for decision makers, The Roots of
Logistics, Springer, pp. 141–172.
Haberleitner, H., Meyr, H. and Taudes, A. (2010). Implementation of a demand planning system using
advance order information, International Journal of Production Economics. 128(2), 518–526.

Hadley, G. and Whitin, T. M. (1963). Analysis of inventory systems, (No Title). .

Hochreiter, S. and Schmidhuber, J. (1997). Long short-term memory, Neural computation. 9(8), 1735–
1780.
Hwang, J. and Kim, H. (2020). The effects of expected benefits on image, desire, and behavioral intentions
in the field of drone food delivery services after the outbreak of covid-19, Sustainability. 13(1), 117.

Khashei, M. and Bijari, M. (2011). A novel hybridization of artificial neural networks and arima models
for time series forecasting, Applied soft computing. 11(2), 2664–2675.
Kiuchi, A. (2024). Recurrent neural network based reinforcement learning for inventory control with
agent-based supply chain simulator, 2024 IEEE 20th International Conference on Automation Science
and Engineering (CASE), IEEE, pp. 1903–1909.
Kleinemolen, I. (2024). Inventory Optimization and Simulation Analysis for Supply Chain Disruption
Events, PhD thesis, Massachusetts Institute of Technology.
Kumar, A., Shankar, R. and Aljohani, N. R. (2020). A big data driven framework for demand-driven
forecasting with effects of marketing-mix variables, Industrial marketing management. 90, 493–507.

Lee, H. L., Padmanabhan, V. and Whang, S. (1997a). The bullwhip effect in supply chains, . .
Lee, H. L., Padmanabhan, V. and Whang, S. (1997b). Information distortion in a supply chain: The
bullwhip effect, Management science. 43(4), 546–558.
Lee, H. L., So, K. C. and Tang, C. S. (2000). The value of information sharing in a two-level supply
chain, Management science. 46(5), 626–643.
Li, D., Wang, X., Chan, H. K. and Manzini, R. (2014). Sustainable food supply chain management.
Li, S. H. and Dörfler, F. (2024). Mitigating transient bullwhip effects under imperfect demand forecasts,
arXiv preprint arXiv:2404.01090. .

Liu, D., Wang, W. and Zhao, Y. (2022). Effect of weather on online food ordering, Kybernetes. 51(1), 165–
209.
Maheshwari, P., Kamble, S., Pundir, A., Belhadi, A., Ndubisi, N. O. and Tiwari, S. (2021). Internet
of things for perishable inventory management systems: an application and managerial insights for
micro, small and medium enterprises, Annals of Operations Research. .

28
Min, H. (2010). Artificial intelligence in supply chain management: theory and applications, International
Journal of Logistics: Research and Applications. 13(1), 13–39.
Mohsin, A. and Yellampalli, S. S. (2017). Iot based cold chain logistics monitoring, 2017 IEEE Inter-
national Conference on Power, Control, Signals and Instrumentation Engineering (ICPCSI), IEEE,
pp. 1971–1974.
Moroff, N. U., Kurt, E. and Kamphues, J. (2021). Machine learning and statistics: A study for assessing
innovative demand forecasting models, Procedia Computer Science. 180, 40–49.
Nair, A. and Vidal, J. M. (2011). Supply network topology and robustness against disruptions–an
investigation using multi-agent model, International Journal of Production Research. 49(5), 1391–
1404.
Negi, S. and Wood, L. C. (2019). Transportation lead time in perishable food value chains: an indian
perspective, International Journal of Value Chain Management. 10(4), 290–315.
Osman, S. A., Xu, C., Akuful, M. and Paul, E. R. (2023). Perishable food supply chain management:
Challenges and the way forward, Open Journal of Social Sciences. 11(7), 349–364.
Pal, S. (2023). Optimizing the last mile: Advanced predictive analytics for delivery time estimation
in supply chain logistics, International Journal for Innovative Research in Multidisciplinary Field.
9(11), 55–61.
Paruthipattu, S. P. (2021). Demand Forecasting based on External Factors using Clustering and Machine
learning, PhD thesis, Dublin, National College of Ireland.
Peng, R. and Xiao, Y. (2014). How to manage the bullwhip effect in the supply chain: A case study on
chinese haier group.
Potočnik, P., Šilc, J., Papa, G. et al. (2019). A comparison of models for forecasting the residential
natural gas demand of an urban area, Energy. 167, 511–522.
Rahman, A., Srikumar, V. and Smith, A. D. (2018). Predicting electricity consumption for commercial
and residential buildings using deep recurrent neural networks, Applied energy. 212, 372–385.
Ramanathan, U. (2014). Performance of supply chain collaboration–a simulation study, Expert Systems
with Applications. 41(1), 210–220.
Ramírez, F. G. (2021). Neural multi-quantile forecasting for optimal inventory management, arXiv
preprint arXiv:2112.05673. .
Ravichandran, N. (2008). Managing bullwhip effect: two case studies, Journal of Advances in Manage-
ment Research. 5(2), 77–87.
Rezaeefard, M., Pilevari, N. and Faezy Razi, F. (2024). Present a mixed approach of neural network
and bat algorithm to predict customer demand in the supply chain to reduce the bullwhip effect,
International Journal of Nonlinear Analysis and Applications. 15(4), 65–78.
Rezki, N. and Mansouri, M. (2024). Deep learning hybrid models for effective supply chain risk manage-
ment: mitigating uncertainty while enhancing demand prediction, Acta Logistica. 11(4), 589–604.
Sak, H., Senior, A. W., Beaufays, F. et al. (2014). Long short-term memory recurrent neural network
architectures for large scale acoustic modeling., Interspeech, Vol. 2014, pp. 338–342.
Sarıcıoğlu, A., Genevois, M. E. and Cedolin, M. (2024). Analyzing one-step and multi-step forecasting
to mitigate the bullwhip effect and improve supply chain performance, IEEE Access. .
Shahrabi, J., Hadavandi, E. and Asadi, S. (2013). Developing a hybrid intelligent model for forecasting
problems: Case study of tourism demand time series, Knowledge-Based Systems. 43, 112–122.
Siami-Namini, S., Tavakoli, N. and Namin, A. S. (2019). A comparative analysis of forecasting financial
time series using arima, lstm, and bilstm, arXiv preprint arXiv:1911.09512. .
Sodhi, M. S., Son, B.-G. and Tang, C. S. (2012). Researchers’ perspectives on supply chain risk man-
agement, Production and operations management. 21(1), 1–13.

29
Sterman, J. D. (1989). Modeling managerial behavior: Misperceptions of feedback in a dynamic decision
making experiment, Management science. 35(3), 321–339.
Suali, A. S., Srai, J. S. and Tsolakis, N. (2024). The role of digital platforms in e-commerce food supply
chain resilience under exogenous disruptions, Supply Chain Management: An International Journal.
29(3), 573–601.
Syntetos, A. A. and Boylan, J. E. (2005). The accuracy of intermittent demand estimates, International
Journal of forecasting. 21(2), 303–314.
Szul, T., Nęcka, K. and Mathia, T. G. (2020). Neural methods comparison for prediction of heating
energy based on few hundreds enhanced buildings in four season’s climate, Energies. 13(20), 5453.
Thecharmsofs, O., BILITY—NORMAL, V. R. R. B., LOG, O. and AT, T. (2001). Log-no rmal distri-
butions across the sciences: Keys and clues, BioScience. 51(5).
Tjen, S., Gozali, L., Kristina, H. J., Gunadi, A. and Irawan, A. P. (2023). Demand forecasting using
time series and ann with inventory control to reduce bullwhip effect on home appliances electronics
distributors, Mobile Computing and Sustainable Informatics: Proceedings of ICMCSI 2023, Springer,
pp. 285–300.
Tkachuk, S., Łukasik, S. and Wróblewska, A. (2025). Consumer transactions simulation through gener-
ative adversarial networks under stock constraints in large-scale retail, Electronics. 14(2), 284.

Udenio, M., Vatamidou, E. and Fransoo, J. C. (2023). Exponential smoothing forecasts: taming the
bullwhip effect when demand is seasonal, International Journal of Production Research. 61(6), 1796–
1813.
Viloria, A., Matos, L. F. A., Gaitán, M., Palma, H. H., Guzmán, Y. F., Vásquez, L. C., Mercado,
C. V. and Lezama, O. B. P. (2019). Demand forecasting method using artificial neural networks,
Dependability in Sensor, Cloud, and Big Data Systems and Applications: 5th International Conference,
DependSys 2019, Guangzhou, China, November 12–15, 2019, Proceedings 5, Springer, pp. 383–391.
Wangphanich, P., Kara, S. and Kayis, B. (2010). Analysis of the bullwhip effect in multi-product,
multi-stage supply chain systems–a simulation approach, International journal of production Research.
48(15), 4501–4517.

Wu, Y., Yuan, M., Dong, S., Lin, L. and Liu, Y. (2018). Remaining useful life estimation of engineered
systems using vanilla lstm neural networks, Neurocomputing. 275, 167–179.
Yang, N., Ding, Y., Leng, J. and Zhang, L. (2022). Supply chain information collaborative simulation
model integrating multi-agent and system dynamics, Promet-Traffic&Transportation. 34(5), 711–724.

Yang, Y., Lin, J., Liu, G. and Zhou, L. (2021). The behavioural causes of bullwhip effect in supply
chains: A systematic literature review, International Journal of Production Economics. 236, 108120.
Yao, W., Zhao, H. and Liu, L. (2023). Weather and time factors impact on online food delivery sales: a
comparative analysis of three chinese cities, Theoretical and Applied Climatology. 153(3), 1425–1438.
Yao, Y. and Dresner, M. (2008). The inventory value of information sharing, continuous replenishment,
and vendor-managed inventory, Transportation Research Part E: Logistics and Transportation Review.
44(3), 361–378.
Zhang, C. and Wei, W. (2022). Iot network for international trade cold chain logistics tracking based on
kalman algorithm, Computational Intelligence and Neuroscience. 2022(1), 1608167.

Zhang, G. P. (2003). Time series forecasting using a hybrid arima and neural network model, Neurocom-
puting. 50, 159–175.
Zhu, X., Ninh, A., Zhao, H. and Liu, Z. (2021). Demand forecasting with supply-chain information and
machine learning: Evidence in the pharmaceutical industry, Production and Operations Management.
30(9), 3231–3252.

30