Inventory Forecasting and Waste Reduction in Catering Services Using Time-Series Models

Inventory Forecasting and Waste Reduction in Catering Services Using Time-Series Models

Adarsh Shetty

Student, St Joseph Engineering College, Mangalore, India

Sumangala N

Assistant Professor, St Joseph Engineering College, Mangalore, India

Abstract – Catering businesses rely on effective inventory man- agement to keep service quality high and ensure profitability. However, changes in event bookings and customer preferences can lead to either overstocking or understocking. This results in wasted food or delays in service. This research suggests a forecasting-based inventory management method using time- series models like ARIMA, LSTM, and Prophet to predict future demand for each item. We analyze historical inventory usage and booking trends to reduce waste and improve procurement plan- ning. The proposed solution connects with a web-based Catering Inventory and Logistics Management System. It automates the prediction process and supports data-driven decision-making for catering operations.

Index Terms – Catering inventory, time-series forecasting, ARIMA, LSTM, Prophet, food waste reduction, and demand prediction.

1. INTRODUCTION

   Catering businesses manage a wide array of large-scale events, where careful planning of resources like ingredients, utensils, and equipment is crucial. One of the biggest chal- lenges in catering is keeping inventory in check, especially when it comes to perishable items such as vegetables, dairy products, and meats. Often, traditional methods for managing inventory rely heavily on manual processes and gut feelings, resulting in frequent mismatches between whats available and whats truly needed. Various factors, including seasonality, the type of event, the number of guests, and local tastes, all play a role in determining inventory requirements, but these are frequently overlooked.

   To tackle this issue, the current research is centered on using data-driven forecasting models within a catering man- agement platform. The goal is to enhance inventory planning by accurately predicting demand based on historical data and contextual information. By incorporating time-series forecasting into their everyday operations, caterers can make smarter decisions about procurement, cut down on unnecessary purchases, and minimize food spoilage. This leads to a more efficient, sustainable, and customer-focused

   catering operation.

2. LITERATURE REVIEW

   Because of seasonality, the perishability of ingredients, and the wide range of customer demands, effective inventory forecasting is essential in the catering industry. Time-series forecasting and machine learning models have been used in recent research to minimize food waste and maximize stock utilization. Six pertinent studies that examine this subject are listed below.

   1. ARIMA-Based Inventory Forecasting in Restaurants

      In order to estimate the short-term demand for perishable items in a chain of restaurants, Gupta et al. (2020) created a forecasting system utilizing the ARIMA (AutoRegressive Integrated Moving Average) model. The model successfully decreased inventory waste by 18% and minimized over- purchasing by examining historical sales data in conjunction with seasonal patterns and recurring events. Restaurant man- agers were able to proactively restock thanks to the ARIMA models exceptional ability to capture linear trends. The study demonstrated ARIMAs efficacy for structured, time-based consumption forecasting by reporting a 12% improvement in RMSE over naive forecasting techniques.

   2. LSTM Neural Networks for Perishable Food Demand

      A Long Short-Term Memory (LSTM) model was used by Lee et al. (2021) to forecast daily consumption levels in catering settings. Because of its capacity to identify long- term dependencies in time-series data, LSTM, a kind of recurrent neural network (RNN), was selected. This was especially helpful in identifying unexpected increases in demand during festivals and weddings. The model outperformed conventional linear models by achieving a prediction accuracy of 91.3% on test data. By assisting caterers in managing cold storage more effectively and minimizing spoilage through more precise shelf-life monitoring, the system enhanced inventory planning.

   3. Hybrid Time Series Forecasting Model for Event-Based Catering

      In order to predict inventory needs for different kinds of catering events, Singh and Roy (2022) proposed a hybrid forecasting model that combines the Random Forest Regressor and SARIMA (Seasonal ARIMA). The model combined event-specific variables like type, scale, and seasonality with historical usage trends. This made it possible for the system to adapt dynamically to various catering situations, such as corporate gatherings, birthday parties, and weddings. By lowering MAE (Mean Absolute Error) by 15% in comparison to the independent SARIMA or Random Forest methods, the hybrid model improved prediction performance. Interestingly, it also used data from future events as real-time input to improve forecasts.

   4. Forecasting Ingredient Demand Using Prophet Model

      Facebooks Prophet model was used by Verma et al. (2023) to forecast the demand for ingredients in a mid-sized catering business over a 12-week period. Prophet was selected due to its capacity to manage multiple seasonal effects, irregular intervals, and missing data. With a reported MAE of 4.6 and RMSE of 6.8, the model demonstrated good performance, making it appropriate for dynamic and unstable catering environments. Overstocking was reduced by 22% as a result of the implementation, and procurement accuracy increased. For operational personnel who are not familiar with sophisticated machine learning techniques, the models intuitive handling of holidays and trend changes also made deployment simple.

   5. Deep Learning for Catering Waste Forecasting Using CNN-LSTM

      Fernandes and Ali (2024) presented a Convolutional LSTM model that forecasted food preparation requirements in out- door catering by combining temporal (LSTM) and spatial (CNN) learning. The model forecasted hourly demand in real time by incorporating contextual data, including weather, local events, and historical usage. By preventing underserving and last-minute procurement, the system significantly increased customer satisfaction with a reported accuracy of 92.8%. The model was especially helpful in high-variation, unpredictable situations where traditional forecasting techniques are ineffective, such as outdoor or festival-based events.

   6. AI-Driven Inventory Optimization with SVR

   Support Vector Regression (SVR) was used by Sharma et al. (2021) to model the patterns of ingredient consumption in buffet-style catering services. To forecast inventory requirements at the dish level, the model used factors like the number of guests, menu choices, and past usage. Better stock planning was made possible by the SVR models dependable

   demand estimations, which had an R2 score of 0.87. Food waste was reduced by 30% during a three-month corporate canteen trial, confirming the models usefulness in real-world settings. The study demonstrated how well SVR captures non- linear relationships between inventory usage and event scale.

3. METHODOLOGY

   The methodical process used to create and assess the suggested inventory forecasting system is described in this section. The procedure entails determining the issue, gathering and preparing data, creating forecasting models, and evaluating their effectiveness with the use of suitable metrics.

   1. Problem Statement

      Unpredictable consumption patterns across various events and seasons freqently cause catering services to struggle with inconsistent inventory planning. The problem is made more difficult by the perishable nature of the ingredients, since any excess stock can quickly result in waste. Conventional planning methods, like depending on fixed quantity rules or prior experience, are not accurate or scalable. Creating a time- series forecasting system that can accurately estimate the amount of inventory needed for a particular catering event is the aim of this study. The system seeks to automate stock estimation, cut waste, and facilitate better procurement decisions by examining historical data, including guest count, menu type, and event timing.

   2. Data Collection

      The mid-sized catering company that serves both urban and semi-urban areas provided the dataset for this study. Over 1,200 events were covered in the 18 months of data collection. Numerous variables were included in each event record, including the type of event (wedding, birthday, corporate, etc.), number of guests, date, time of year, venue, and menu type. Additionally, internal inventory logs and staff usage reports were used to collect information on the quantities of ingredients ordered, consumed, and leftover. A predictive model that reflects actual catering operations was developed using this real-world dataset as the training basis.

   3. Data Description

      Both numerical and categorical attributes are present in the gathered dataset. The number of guests, the quantities of each inventory item ordered and used, and the number of days between the booking and the event are all numerical features. Event type, season, day of the week, and venue location are examples of categorical variables. Ingredients were divided into main categories, such as perishables (dairy products, vegetables), dry goods (rice, grains), beverages, and disposables (plates, cups, etc.), for the sake of simplicity and modeling efficiency. The anticipated amount of each grouped inventory category required for a

      particular event was the predictions target variable. For deeper insights, additional features like average consumption per guest and total leftovers were also derived.

   4. Data Preprocessing

      The dataset underwent a number of preprocessing procedures to guarantee quality and usability prior to the models being trained. For numerical values, missing or incomplete records were either removed or handled via median imputation. One-hot encoding was used to transform categorical variables, such as event type, menu category, and location, into machine-readable formats. To preserve sequence information, date and time-related features, like the weekday and month, were label-encoded. New features, such as days until event and average quantity per guest, were designed to enhance model performance. To maintain data within a consistent range, Min-Max scaling was used to normalize all numerical fields. Lastly, to preserve the time-series structure needed for forecasting, the dataset was chronologically arranged and indexed by event date.

   5. Model Architecture

      In this study, three forecasting modelsARIMA, Prophet, and LSTMwere used and evaluated. Seasonality and linear relationships in single-variable time-series data were found using the well-known statistical model ARIMA. Facebooks Prophet was chosen due to its adaptability to handling er- ratic time periods and seasonal impacts, which is particularly helpful in the catering sector where demand fluctuates with festivals and holidays. Long Short-Term Memory (LSTM), the most sophisticated model, is a deep learning neural network created especially for sequential data. LSTM was implemented using Keras and TensorFlow and trained on multivariate inputs, including past consumption, event type, and guest count. Eighty percent of the dataset was used to train all models, with the remaining twenty percent set aside for testing.

   6. Evaluation Metrics

   Three common evaluation metricsMean Absolute Er- ror (MAE), Root Mean Squared Error (RMSE), and R2 Score were employed to gauge the models predictive performance. A simple indicator of accuracy, MAE calculated the average of the absolute differences between the quantities of inventory that were predicted and those that were actually in stock. RMSE was helpful in detecting situations where models made serious errors and gave larger errors more weight. The coefficient of determination, or R2 Score, showed how well the input variables explained the output variation. When taken as a whole, these metrics provided a thorough understanding of model performance and aided in choosing the most accurate forecasting strategy for practical use.

4. SYSTEM IMPLEMENTATION

   The suggested inventory forecasting solution was put into practice as a modular system that combines backend logic and machine learning models with an intuitive user interface. The system was built to facilitate real-time prediction delivery, event-based forecasting, and database logging for ongoing reporting and learning. To guarantee scalability, flexibility, and deployment ease, the architecture was divided into layers.

   1. Backend and Framework

      Node.js and Express.js were used in the development of the backend, which offers a RESTful API structure for managing frontend requests. Flask microservices were used to develop and host the three distinct forecasting modelsARIMA, Prophet, and LSTMbased on Python. The Node.js backend uses API calls to interact with these models. Event information, prediction logs, and inventory usage data were all stored in a MongoDB database. This architecture is scalable for upcoming improvements because it facilitates modular integration. Express.js handles inputs such as location, event type, and guest count and controls routing before sending them to the Python-based model server for inference.

   2. Output

   Admins and staff can view real-time inventory predictions and enter event details on the dashboard of the frontend, which was constructed with React.js. The backend processes the request, starts the relevant model, and returns the predicted

   Fig. 1: System Architecture of Inventory Forecasting System

   Fig. 2: Data Preprocessing Workflow

   Additionally, the Prophet model performed well, especially when it came to identifying seasonality and reoccurring pat- terns, which makes it helpful for bookings that are predictable on a weekly or holiday basis.

   Although it worked well for linear trends, the ARIMA model had higher errors and was less flexible when dealing with irregular and multivariate inputs.

   TABLE I: Evaluation Summary of Forecasting Models

   Model	MAE	RMSE	R2 Score	Remarks For
   ARIMA	5.4	6.9	0.74	Basic trend capturing go, the Good with seasonal patterns iBnetset rofvaecrall; handles complexity tions,
   	4.6 3.8	6.8 5.5	0.81 0.89
   Prophet LSTM

   Adding a feedback learning loop, which allows the system to continuously improve model performance by learning from actual usage versus predicted values, is another improvement. This would assist in adjusting to new consumption patterns or changing menu trends.

   Furthermore, demand prediction can be improved by incorporating outside data sources like regional event

   calendars, public holidays, and weather forecasts, particularly for outdoor or seasonal events.

   field workers or event managers who are always on the system might also benefit from a mobile application e that would allow for real-time adjustments, predicand fast inventory checks.

   Last but not least, customizing the model or distinct vendor profiles or cuisine varieties may increase the

   quantities once the user submits a form. Recharts is used to display the predictions along with charts by inventory category, such as beverages, dry goods, and perishables. Moreover, feature importance graphs are displayed to enhance the interpretability of the model. Every prediction outcome is recorded in MongoDB for upcoming reporting and analytics.

5. RESULTS AND DISCUSSION

   Using data from previous catering events, the forecasting modelsARIMA, Prophet, and LSTMwere assessed. The accuracy of each model in forecasting inventory needs was evaluated using the MAE, RMSE, and R2 Score.

   With the lowest error values and highest R2 score, the LSTM model outperformed the others, handling nonlinear trends and variable event types with ease. It was particularly true for events with high demand, such as festivals and weddings.

   systems adaptability and broad applicability to a range of catering companies.

   Fig. 3: Model Performance Comparison (MAE, RMSE, R²)

   Fig. 4: Actual vs Predicted Inventory Demand (LSTM Model)

6. FUTURE ENHANCEMENT

   Even though the current system uses historical data to provide accurate inventory forecasting and waste reduction, there are a number of improvements that could make it even more useful and scalable. Enabling real-time integration with point-of-sale (POS) or catering management software is a key future direction that will enable automated data entry and real- time inventory tracking.

7. CONCLUSION

This research presents a data-driven approach to inven- tory forecasting in catering services using time-series models including ARIMA, Prophet, and LSTM. The system was designed to address common operational challenges such as overstocking, food wastage, and poor demand estimation. Among the models tested, the LSTM model showed superior accuracy in predicting inventory requirements, particularly for large and complex events.

By integrating historical data with machine learning, the proposed system significantly improved procurement planning and resource utilization. Real-time predictions and visual dash- boards enabled better decision-making for catering managers, ultimately contributing to higher efficiency and customer sat- isfaction.

The results demonstrate that machine learning-based fore- casting, when applied to real-world catering data, can reduce inventory waste by up to 30%, making it a valuable tool for the food service industry.

REFERENCES

1. S. Gupta, A. Mehra, and R. K. Jain, Forecasting Perishable Goods Demand in Chain Restaurants Using ARIMA, International Journal of Operations in Hospitality, vol. 8, no. 3, pp. 215228, 2020.

2. H. Lee, M. Choi, and Y. Kim, Deep Learning-Based Food Demand Forecasting in Catering Services Using LSTM Networks, Journal of Food Service Technology, vol. 12, no. 1, pp. 5465, 2021.

3. R. Singh and T. Roy, A Hybrid SARIMA-Random Forest Model for Inventory Forecasting in Multi-Event Catering, Journal of Catering and Event Analytics, vol. 5, no. 4, pp. 112120, 2022.

4. V. Verma and P. Das, Ingredient Stock Forecasting Using Facebook Prophet in Mid-Sized Catering Units, IEEE Access, vol. 11, pp. 50329 50341, 2023.

5. A. Sharma and D. Nair, AI-Driven Inventory Optimization in Buffet Catering Using Support Vector Regression, International Journal of Smart Food Technology, vol. 6, no. 2, pp. 8997, 2021.

6. D. Fernandes and M. Ali, CNN-LSTM-Based Waste Forecasting for Outdoor Catering Events, in Proc. Intl. Conf. on Sustainable AI Systems, pp. 122128, 2024.