Food Demand Forecasting Using Machine Learning

M. Jayamma, R. Akib Hussain, P. Hemanth Kumar, M. Anil Kumar,

A. Prasanth and S. Prasharshavaradhan

Department of Computer Science and Engineering, Santhiram Engineering College, Nandyal, Andhra Pradesh, India

Keywords: Food Demand Forecasting, Machine Learning, Time Series Analysis, Predictive Analytics, XGBoost, LSTM,

Supply Chain Management, Data Visualization.

Abstract: Food Demand forecasting is a process that has to be done by food and beverage companies to ensure that they

manage their supply chains efficiently and have minimum food wastage and optimum inventory levels.

Accurate demand prediction can help businesses to meet customer expectations, reduce operating expenses,

and prevent stockout or overstocking. The final universal machine learning based system is developed from

the systemic insights gained through the study of past sales data, weather, seasonality, promotion and holidays.

There are some data preparation steps, cleaning, normalization, and some other statistical knowledge is used

to extract dignified features for the prediction (called feature engineering). For demand prediction different

machine learning algorithms are applied such as LSTM (Long Short-Term Memory networks), XGBoost,

ARIMA (Auto Regressive Integrated Moving Average) and Linear regression. These models will be trained

on past data, and evaluated on metrics like MAE (Mean Absolute Error), RMSE (Root Mean Square Error),

and R2 Score. Thus, XGBoost is good in prediction by fitting to the data in a less continuous manner while

LSTM effectively captures time series dependencies.

1 INTRODUCTION

Food demand forecasting is essential in catering

customer satisfaction, decreasing waste and

maximizing inventory control. Well-informed

forecasting allows businesses to regulate their

inventory levels, anticipate changes in demand, and

reduce operational expenses. Incorrect projections

can cause overstocking, leading to lost sales and

disgruntled customers, or understaffing, causing

increased storage costs. The complex and, sometimes,

volatile nature of food demand often goes beyond the

scope of traditional forecasting techniques such as

rule-based systems and statistical models. Factors

outside the business such as weather, special events,

seasonal trends, and promotional activities greatly

affect demand patterns Mahammad, Suman and

Sunar (2024). Machine learning algorithms are a

robust approach as they look for undiscovered

patterns in large amounts of data and make accurate

predictions Suman et al., (2023).

In this project, we build a prediction model using

historical sales data and contextual knowledge. To

choose the best machine learning algorithm for real-

world use, we compare a number of them. This work

aims to close the gap between classic statistical

methods and modern AI-driven forecasting

techniques Chaitanya et al., (2022).

2 LITERATURE REVIEW

Several studies have researched machine learning for

demand forecasting. The most common methods of

dealing with time series data include linear

regression and ARIMA, which rarely perform well on

non-linear data Hyndman et al., (2024). Interestingly,

tree-ensemble methods such as extreme gradient

boosting (XGBoost) have proven effective for large

and highly dimensional datasets Chen et al., (2018).

Specifically, Long Short-Term Memory (LSTM)

networks are preferred since they are shown to be

capable of capturing seasonal trends and long-term

dependencies in time series data Hochreiter et al.,

(2015). The existing literature show that by

considering external variables (weather, holidays,

promotions, etc.) the forecast accuracy could be much

higher Wei (2016).

Research has shown that ensemble methods are

effective when models such as XGBoost obtain

Jayamma, M., Hussain, R. A., Kumar, P. H., Kumar, M. A., Prasanth, A. and Prasharshavaradhan, S.

Food Demand Forecasting Using Machine Learning.

DOI: 10.5220/0013886500004919

Paper published under CC license (CC BY-NC-ND 4.0)

In Proceedings of the 1st International Conference on Research and Development in Information, Communication, and Computing Technologies (ICRDICCT‘25 2025) - Volume 2, pages

565-569

ISBN: 978-989-758-777-1

565

predictions from various base models to improve

accuracy Chen et al., (2017), while hybrid models

have shown potential in increasing prediction

accuracy through pooling a statistical method such as

ARIMA and deep learning models like LSTMs

Hochreiter et al., (2018).

Hyperparameter tuning techniques and automated

feature selection have made the model creation

process more efficient Athanasopoulos, G., &

Hyndman, R. J (2011). Old models have fared poorly

in situations where data is missing, and demand is

instantaneously fluctuating; however, modern

machine learning algorithms may be able to resolve

those issues by searching for latent relations in the

multidimensional data Shinde et al., (2018). This

project takes these developments and builds them into

a robust and reliable forecasting system.

3 METHODOLOGY

3.1 Data Collection

Accurate demand predictions are made by gathering

data from different sources. Historical Sales Data –

This data is derived from retail store management

systems and contains vital information such as store

ID, product ID, quantity sold, and date of sale. We

also collect Weather Data from external APIs to

account for how weather conditions affect food

consumption patterns. Temperature, precipitation,

and other relevant climate factors. Event and Holiday

Data: Data is pulled from public records to account

for sales spikes during holidays and other events. The

Integration of Promotions and marketing campaigns,

which allows to see how discounts and offers affect

demand. These datasets combine to make a complete

input for the forecasting model.

3.2 Data Preprocessing

After collection, the data goes through several pre-

processing steps to ensure quality and consistency.

We need to remove the duplicate data, fill the missing

values using interpolation and detect the outliers

using statistical methods like Interquartile Range

(IQR). Feature engineering is applied after cleaning,

generating meaningful variables from the data. These

features like season, month, day of the week and lag-

based features.

3.3 Data Splitting

To make the model learn and test efficiently, the pre-

processed data is split into three parts: Training Set,

Validation Set, and Test Set. In general, the training

process uses 70% of the data while validation and

testing use 20% and 10%, respectively. The test set

assesses the performance of the model on an unseen

dataset, the validation set is used for hyperparameter

tuning and the training set is used for building and

fitting the model. This separation ensures that the

model generalises and reduces the chances of

overfitting.

3.4 Model Selection

Multiple machine learning models are implemented

to come up with the best accurate forecasting

solution. We use Linear Regression (LR) as baseline

model, because it is very interpretable and easy to

implement. ARIMA (Auto Regressive Integrated

Moving Average) → This is used for time series and

for detecting the trend and seasonal patterns from the

data. XGBoost was selected due to its excellent

performance in working with large datasets and non-

linear interactions. Another reason for this use is that

Long Short-Term Memory (LSTM) networks have

the capacity of recognizing long-term dependencies

in sequential data. After comparing the results of

these models, the model that performed the best gets

deployed.

3.5 Hyperparameter Tuning

This means using hyper-parameter tuning techniques

such as Grid Search and Random Search to optimize

the performance of the model. These techniques

systematically explore multiple combinations of

parameters to identify the optimal setup. Evaluation

metrics used for comparison of model performance

are Mean Absolute Error (MAE), Root Mean Square

Error (RMSE), and R2 Score. Lower error values and

a higher R2 value indicate better prediction accuracy.

3.6 Model Evaluation

After training and tuning the models, the test dataset

is used to evaluate the models. Expected v/s Actual

values are plotted using tools like Matplotlib or

Seaborn. That enables a clear comparison of the

accuracy of the models. If a model shows consistent

performance on many data contexts and has only low

prediction errors, it is considered fit for deployment.

3.7 Deployment

The performance of the best model is integrated into

a framework, such as Flask or Fast API. To achieve

ICRDICCT‘25 2025 - INTERNATIONAL CONFERENCE ON RESEARCH AND DEVELOPMENT IN INFORMATION,

COMMUNICATION, AND COMPUTING TECHNOLOGIES

566

near real-time forecasting, we can create an API

endpoint that accepts input data and returns expected

demand. This API can be used in the supply chain

management systems of the companies to take

decisions automatically. Moreover, the demand

forecasts are presented through dashboards

developed with industry-leading technologies such as

Tableau or Power BI to allow stakeholders to quickly

track and analyse demand behaviour.

3.8 Continuous Feedback and

Relentless Improvement

models: Quality and consistency the deployed models

are continuously monitored for quality and

consistency. It focuses on incorporating real-time

feedback and prediction error monitoring via

solutions like Prometheus and Grafana. If the model

begins to wane over time, it is retrained with the most

recent data. With model re-optimisation, your model

remains relevant to changing demand patterns.

Data and minimising the redundancy of their

phrases, this very in-depth process lays the

foundation to an efficient food demand prediction

model that assists companies with inventory

management, minimizing wastage and meeting

consumer demand effectively.

3.9 System Architecture

Figure 1: Historical Sales Data Processing Flowchart.

Figure 1 Shows the Historical Sales Data Processing

Flowchart.

4 EXECUTION AND OUTCOMES

4.1 Model Evaluation

The proposed machine learning models were

implemented using Python-based frameworks such as

TensorFlow and Scikit-Learn. The dataset was split

into training (70%), validation (20%), and test (10%)

sets to assess model generalization. During training,

we performed multiple iterations with different

hyperparameter settings to optimize model

performance.

4.2 Comparison of Model Performance

The robustness, accuracy, and efficiency of the

models were evaluated. ML methods such as Random

Forest boosted predictions much higher around 85%

whereas traditional methods like ARIMA only

reached an average of 75% accuracy. Deep learning

models, particularly LSTMs and CNN-LSTM

hybrids, achieved 92% accuracy.

4.3 Computational Efficiency

High processing power was needed to train deep

learning models. Training time was greatly shortened

by using an NVIDIA GPU, enabling quicker

iterations and model adjustment

4.4 Error Analysis and Fine-Tuning

To find forecasting mistakes, residual analysis was

done. The robustness of the model was enhanced by

methods including hyperparameter optimization and

dropout regularization. Across various datasets, the

final optimized model showed excellent

generalization and little overfitting.

5 RESULTS

In terms of error analysis, residual analysis was

performed to identify forecasting mistakes, and the

model's robustness was improved using techniques

such as hyperparameter optimization and dropout

regularization. After fine-tuning, the optimized

model exhibited excellent generalization with

minimal overfitting across various datasets.

Comparison of accuracy between the past and new

system models as shown in Figure 2 below.

Food Demand Forecasting Using Machine Learning

567

Figure 2: Comparison of Accuracy: Past Vs New System

Models.

6 CONCLUSIONS

The machine learning-based food demand forecasting

system that has been suggested shows notable gains

in prediction accuracy when compared to

conventional techniques. By considering outside

variables like the weather, sales, and holidays, the

model offers useful information that helps companies

minimize food waste and improve inventory control.

Using algorithms like LSTM, XGBoost, and ARIMA

guarantees accurate forecasting.

With the help of the system's API deployment,

real-time predictions facilitate quick decision-making

and lower the possibility of stockouts or overstocking.

Dashboards for visualization help stakeholders better

understand demand trends by offering lucid insights.

Future research into more sophisticated models, such

as transformers, and the integration of other data

sources can improve predicting accuracy even more.

All things considered, data-driven decision-making is

enabled by this solution for effective supply chain

management.

REFERENCES

Athanasopoulos, G., & Hyndman, R. J. (2011). The value

of feedback in forecasting competitions. International

Journal of Forecasting, 27(3), 845-849.

Brownlee, J. (2017). Introduction to Time Series

Forecasting with Python: How to Prepare Data and

Develop Models to Predict the Future. Machine

Learning Mastery.

Chaitanya, V. Lakshmi. "Machine Learning Based

Predictive Model for Data Fusion Based Intruder Alert

System." journal of algebraic statistics 13.2 (2022):

2477-2483

Chaitanya, V. Lakshmi, et al. "Identification of traffic sign

boards and voice assistance system for driving." AIP

Conference Proceedings. Vol. 3028. No. 1. AIP

Publishing, 2024

Chen, T., & Guestrin, C. (2016). XGBoost: A scalable tree

boosting system. Proceedings of the 22nd ACM

SIGKDD International Conference on Knowledge

Discovery and Data Mining, 785-794.

Devi, M. Sharmila, et al. "Machine Learning Based

Classification and Clustering Analysis of Efficiency of

Exercise Against Covid-19 Infection." Journal of

Algebraic Statistics 13.3 (2022): 112-117.

Devi, M. Sharmila, et al. "Extracting and Analyzing

Features in Natural Language Processing for Deep

Learning with English Language." Journal of Research

Publication and Reviews 4.4 (2023): 497-502.

Goodfellow, I., Bengio, Y., & Courville, A. (2016). Deep

Learning. MIT Press.

Hochreiter, S., & Schmidhuber, J. (1997). Long Short-Term

Memory. Neural Computation, 9(8), 1735- 1780.

Hyndman, R. J., & Athanasopoulos, G. (2021). Forecasting:

Principles and Practice. OTexts.

Mahammad, Farooq Sunar, Karthik Balasubramanian, and

T. Sudhakar Babu. "A comprehensive research on

video imaging techniques." All Open Access, Bronze

(2019).

Mahammad, Farooq Sunar, and V. Madhu Viswanatham.

"Performance analysis of data compression algorithms

for heterogeneous architecture through parallel

approach." The Journal of Supercomputing 76.4

(2020): 2275-2288.

Mahammad, Farooq Sunar, et al. "Key distribution scheme

for preventing key reinstallation attack in wireless

networks." AIP Conference Proceedings. Vol. 3028.

No. 1. AIP Publishing, 2024.

Makridakis, S., & Hibon, M. (2000). The M3-Competition:

Results, conclusions, and implications. International

Journal of Forecasting, 16(4), 451-476.

Mandalapu, Sharmila Devi, et al. "Rainfall prediction using

machine learning." AIP Conference Proceedings. Vol.

3028. No. 1. AIP Publishing, 2024.

Shinde, P., & Shah, S. (2018). A Review of Machine

Learning and Deep Learning Applications. Proceedings

of the International Conference on Computing

Methodologies and Communication (ICCMC), 264-

267.

Suman, Jami Venkata, et al. "Leveraging natural language

processing in conversational AI agents to improve

healthcare security." Conversational Artificial Intellig-

ence (2024): 699-711.

Sunar, Mahammad Farooq, and V. Madhu Viswanatham.

"A fast approach to encrypt and decrypt of video

streams for secure channel transmission." World

ICRDICCT‘25 2025 - INTERNATIONAL CONFERENCE ON RESEARCH AND DEVELOPMENT IN INFORMATION,

COMMUNICATION, AND COMPUTING TECHNOLOGIES

568

Review of Science, Technology and Sustainable

Development 14.1 (2018): 11-28.

Wei, W. W. (2006). Time Series Analysis: Univariate and

Multivariate Methods. Pearson.

Zhang, G., Eddy Patuwo, B., & Hu, M. Y. (1998).

Forecasting with artificial neural networks: The state of

the art. International Journal of Forecasting, 14(1), 35-

62.

Food Demand Forecasting Using Machine Learning

569