Development of an AI-based restaurant menu demand prediction model utilizing sales and meteorological data

Abstract

Accurate demand forecasting in the restaurant industry is critical for optimizing inventory management, minimizing food waste, and enhancing operational efficiency. This study developed an AI-based system that predicts menu-specific daily sales using historical sales and meteorological data collected from 2021 to 2023. Approximately 384 menu items were individually modeled using deep neural networks configured for multi-class classification. The system achieved strong predictive performance with a mean Pearson correlation coefficient of 0.7945. Additionally, flexible visualization options were implemented to sort predictions by expected or actual sales volumes. The results demonstrate the feasibility of AI-driven demand prediction systems and their potential to transform food service operations toward greater sustainability and efficiency.

Introduction

The global food system faces unprecedented challenges. According to the Food and Agriculture Organization (FAO), approximately 828 million people worldwide were affected by hunger in 2021, a figure exacerbated by climate change, economic instability, and geopolitical conflicts (Falk et al., 2023). Climate change, in particular, has emerged as a critical driver, reducing agricultural productivity through extreme weather events such as droughts, floods, and heatwaves (Yuan et al., 2024). Staple crops such as wheat, maize, and rice are projected to experience significant yield declines in vulnerable regions, threatening food security for billions (Ray et al., 2019).

Meanwhile, stark inequalities persist across the globe. Low- and middle-income countries grapple with chronic undernourishment and food scarcity, while affluent societies produce enormous volumes of food waste. It is estimated that one-third of all food produced globally—approximately 1.3 billion tons—is wasted annually (Brennan and Browne, 2021), representing not only a moral failure but also a considerable environmental burden. In developed nations, wastage primarily occurs at the retail and consumer levels, whereas in developing regions, post-harvest losses are more common due to infrastructural deficiencies. One promising strategy to address food waste is food upcycling, where surplus or byproduct materials are converted into value-added products (Bangar et al., 2024). Although food upcycling initiatives promote sustainability and resource efficiency, they address only the symptoms, not the root causes, of systemic inefficiencies. Ultimately, the most sustainable solution lies in preventing waste generation altogether by optimizing production, distribution, and consumption systems. Conventional restaurant analytics often capture symptoms—such as sales drops during weekends—without identifying root causes like adverse weather conditions or nearby events. Accurate demand forecasting is essential to achieving this goal, particularly in the restaurant industry, where operational inefficiencies contribute significantly to food wastage.

In this context, demand forecasting serves a dual purpose: promoting environmental sustainability and enhancing business profitability. Traditional heuristic or statistical forecasting approaches often fall short in capturing complex, nonlinear relationships influenced by dynamic external factors such as weather conditions and public holidays. Poor forecasts may result in spoilage, stockouts, and reduced customer satisfaction. Recent advances in artificial intelligence (AI) and big data analytics have shown considerable promise in improving forecasting accuracy. Meteorological factors are increasingly recognized as significant determinants of consumer food purchasing behavior. Yao et al. (2023) demonstrated that adverse weather events like heavy rain and poor air quality correlated with increased food delivery sales in Chinese cities. Similarly, Liu et al. (2022) found that fluctuations in temperature and rainfall substantially affected online takeaway consumption patterns. Furthermore, Zhang et al. (2021) identified a U-shaped relationship between ambient temperature extremes and restaurant dining frequency, with both heatwaves and cold spells leading to reduced customer traffic. Temporal events such as weekends, national holidays, and festive seasons also profoundly influence consumer dining behaviors. As shown by Güler et al. (2023), integrating weather conditions and special days into sales forecasting models improved predictive performance.

Chae et al. (2024) further emphasized that the fusion of external data sources into machine learning and deep learning frameworks significantly enhances forecast accuracy in large restaurant chains. Schmidt et al. (2022) and Zhang et al. (2025) highlighted that sophisticated architectures, such as hybrid CNN-LSTM-Attention models, can effectively capture non-linear and time-dependent demand patterns. In addition to general environmental influences, individual menu preferences are highly sensitive to daily conditions such as weather and holidays. For example, consumers may favor light, refreshing dishes during hot, humid periods, whereas they may prefer hearty, warming meals on cold, rainy days. Similarly, special holidays or weekends often shift consumer preferences towards more indulgent, celebratory menu items.

Building on these insights, this study hypothesizes that integrating meteorological data—specifically temperature and precipitation—with historical sales records can capture nuanced shifts in menu preferences driven by external factors. Accordingly, we developed an AI-based forecasting system that not only predicts overall demand but also anticipates dynamic changes in item-specific popularity. Given these considerations, this study proposes an AI-driven restaurant menu demand prediction model that leverages 3 years of historical sales and meteorological data to enable precise item-level forecasting. By aligning production with actual demand more closely, the model aims to minimize food waste, enhance operational efficiency, and contribute to global efforts toward building a more resilient and equitable food system.

Materials and methods

Data collection

The dataset spans July 1, 2021 to September 30, 2023, corresponding to the period when consistent digital records were available and fully validated by the restaurant’s POS system. Due to confidentiality agreements, the specific name and location of the restaurant cannot be disclosed. However, the company consistently achieves annual revenue exceeding approximately 100 million USD, ensuring the data represents a large-scale, high-traffic commercial food service environment.

Meteorological data, including daily average temperature (°C) and total precipitation (mm), were sourced from the Korea Meteorological Administration (KMA). National holidays and weekends were annotated based on the Korean official calendar. A new feature, holiday indicator, was generated based on whether the date corresponded to a national holiday. Additionally, a weekend indicator was created to capture whether the date fell on a Saturday or Sunday. The holiday indicator and weekend indicator were treated as independent binary variables to enrich the feature space. All datasets were synchronized by date without missing entries and combined into a unified dataset.

Data preprocessing

The preprocessing procedure included several steps to ensure data quality and suitability for modeling (Table 1):

• Datetime Feature Extraction: Each date was encoded into cumulative day-of-year (pDay) and a binary holiday indicator (1 for holidays, 0 otherwise).

• Missing Value Handling: A thorough audit confirmed there were no missing entries in either sales or weather data.

• Sales Aggregation: Daily sales counts were aggregated by menu item, resulting in one value per item per day.

• Rolling Features: A 7-day rolling cumulative sales feature was engineered to capture recent sales momentum and short-term trends.

• Feature Normalization: All numerical input features (temperature, precipitation, pDay) were normalized to the [0,1] range using Min–Max scaling to facilitate neural network training convergence.

• Label Encoding: Menu item sales were discretized into five categories based on historical quantiles (very low, low, medium, high, very high) to frame the task as a classification problem.

The pDay feature, representing the day-of-year, was computed using Python’s datetime library, which accurately accounts for leap years (e.g., February 29 in leap years is included as day 60).

Table 1.

Structure of the final dataset

Feature	Description
Date	Date of record
pDay	Cumulative day number within the year
Temperature	Daily average temperature (°C)
Rain	Daily total precipitation (mm)
Holiday Indicator	Binary indicator for national holidays (1 = holiday, 0 = non-holiday)
Weekend Indicator	Binary indicator for weekends (Saturday or Sunday) (1 = weekend, 0 = weekday)
Menu0 ~ Menu399	Daily sales counts for each menu item

Model development

Each menu item was modeled individually using a customized feedforward deep neural network (DNN), which was configured in Keras with a categorical cross-entropy loss function and the Adam optimizer. Sales volumes were discretized into predefined categories based on the historical data distribution.

The neural network architecture for each menu item was composed as follows:

• Input Layer: The number of neurons equals the number of input features (pDay, temperature, rain, holiday, and weekend).

• First Hidden Layer: 100 neurons; activation function: Rectified Linear Unit (ReLU); weight initialization: He normal initializer (named after Kaiming He—not a pronoun—specifically designed for networks using ReLU to maintain variance during training).

• Second Hidden Layer: 100 neurons; activation function: ReLU.

• Output Layer: The number of neurons was dynamically determined by the label encoding process for each menu item, corresponding to the number of sales categories. The activation function was set to softmax, producing a probability distribution over the predicted sales categories.

For model development, approximately 384 distinct menu items were considered. Each menu item was treated as an independent learning task. Thus, 384 individual models were sequentially trained, each optimized for the specific demand characteristics of its corresponding menu item. This approach allowed the model to adapt finely to menu-specific consumption patterns rather than enforcing a single unified prediction model across diverse menu categories.

The architecture consisting of two hidden layers with 100 neurons each was determined through preliminary grid search experiments, testing various combinations of layer depths and neuron counts. This configuration offered optimal performance while maintaining training efficiency across 384 menu-specific models. The classification labels were created by applying labels encoding to the historical sales records, segmenting sales into predefined intervals. No regularization techniques such as dropout or batch normalization were applied in the baseline configuration, ensuring consistent model architecture across different menu items. This architecture was selected to balance computational efficiency with the ability to capture nonlinear relationships between environmental factors (weather, calendar attributes) and menu-specific demand patterns.

Model training

Model training was implemented using custom Python scripts developed specifically for this project. The primary training pipeline utilized TensorFlow and Keras libraries to construct, compile, and train the deep learning models for each menu item independently. GPU acceleration was used where available to expedite training times. In cases where computational resources were limited, training was parallelized across menu items to maximize throughput.

Key training steps included:

• Data Splitting: For each menu item, the dataset was split in chronological order, with the first 80% used for training and the remaining 20% for validation. No random shuffling was applied to preserve the temporal dependency of the data.

• Model Compilation: Models were compiled using the Adam optimizer with a fixed learning rate of 0.001, leveraging Keras's implementation. The loss function was categorical crossentropy, suitable for multi-class classification problems.

• Batching and Epochs: Training was conducted with a batch size of 32 samples and a fixed schedule of 50 epochs. No dynamic learning rate scheduling or early stopping was applied, to maintain uniform training cycles across the 384 individually trained menu models. To mitigate overfitting risks, dropout layers (rate = 0.3) were inserted after each hidden layer, and a relatively shallow architecture (two hidden layers with 100 neurons each) was adopted. While dropout and weight decay helped reduce overfitting, the lack of dynamic stopping criteria may have impacted generalizability, particularly for low-volume menu items. This limitation offers a potential area for future improvement.

• Monitoring and Logging: During training, both training and validation loss values were recorded epoch by epoch. These logs were saved automatically to CSV files for later analysis. Validation accuracy trends were visually inspected to confirm model convergence and detect signs of overfitting.

The training script also included functionality to dynamically normalize input features, encode target labels using one-hot encoding, and save the best model weights after training for each menu item individually. Overall, the training strategy focused on stability, reproducibility, and scalability, enabling efficient model development across hundreds of individual menu categories. The training was performed under the following settings:

• Training/Validation Split: Data were split chronologically, with 80% used for training and 20% reserved for validation. Temporal order was preserved to prevent information leakage.

• Batch Size: 32 samples per batch were processed to ensure stability and speed during training.

• Optimizer: The Adam optimizer was employed with a learning rate of 0.001, utilizing default beta parameters (β₁ = 0.9, β₂ = 0.999).

• Loss Function: Categorical Cross-Entropy was used due to the multiclass nature of the output.

• Epochs: Each model was trained for 50 full passes over the training dataset.

• Early Stopping: Not applied to maintain a consistent training regime across all menu items.

Training loss and validation loss were monitored throughout the process to verify convergence.

Evaluation metrics

Model performance was evaluated using multiple quantitative metrics, each capturing different aspects of prediction quality:

• Mean Absolute Error (MAE):

  $$MAE=\frac{1}{n}\sum _{i=1}^n\left|y_i,-,{\widehat{y}}_i\right|$$

  MAE measures the average magnitude of errors between predicted values () and actual values () without considering their direction. It provides a straightforward interpretation of the average deviation. A lower MAE indicates that, on average, the predictions are closer to the actual observed values.

• Root Mean Squared Error (RMSE):

  $$RMSE=\sqrt{\frac{1}{n}\sum _{i=1}^n{\left(y_i,-,{\widehat{y}}_i\right)}^2}$$

  RMSE squares the errors before averaging and then takes the square root, penalizing larger errors more severely than MAE. This makes RMSE sensitive to outliers, providing insight into cases where the model may significantly underperform for certain instances. A lower RMSE reflects both overall accuracy and control over large deviations.

• Mean Absolute Percentage Error (MAPE):

  $$MAPE=\frac{100}{n}\sum _{i=1}^n\left|\frac{y_i-{\widehat{y}}_i}{y_i}\right|$$

  MAPE expresses prediction errors as a percentage of the actual values, making it easier to interpret relative error irrespective of scale. Lower MAPE values signify better predictive proportional accuracy. MAPE is particularly useful in commercial applications, where stakeholders prefer to understand prediction accuracy in percentage terms.

• Pearson Correlation Coefficient (r): The Pearson correlation coefficient measures the strength and direction of the linear relationship between predicted and actual sales:

  $$r=\frac{{\sum }_{i=1}^n\left(y_i,-,\overline{y}\right)\left({\widehat{y}}_i,-,\overline{\widehat{y}}\right)}{\sqrt{\sum _{i=1}^n{\left(y_i,-,\overline{y}\right)}^2\sum _{i=1}^n{\left({\widehat{y}}_i,-,\overline{\widehat{y}}\right)}^2}}$$

  where and represent the mean values of actual and predicted sales, respectively. An value close to 1 indicates a strong positive linear correlation, while a value near 0 suggests no linear relationship.

Overall system workflow

The complete workflow of the demand prediction system is illustrated in Fig. 1. First, raw sales and weather data were collected and subjected to preprocessing steps including datetime feature extraction (e.g., day of year encoding), menu item encoding, and basic cleaning. After preprocessing, an integrated dataset was saved in Excel format (Result.xlsx). Next, basic data distributions were analyzed using bin plots to visualize the density of sales counts across menu items and seasons. Feature variables—including day number, average temperature, precipitation, and holiday indicator—were selected as inputs for the deep learning models. Each menu item was handled independently. The input data were processed through two hidden layers, each comprising 100 neurons, followed by a softmax output layer for classification.

Fig. 1.

Overview of the data preprocessing, modeling, and prediction service workflow

After training, each menu’s model was saved and prepared for prediction service deployment. Prediction services were built so that when a user inputs target conditions (e.g., date, forecasted temperature, precipitation, and holiday status), the corresponding menu-wise sales predictions are generated immediately. The final results are visualized as bar graphs comparing predicted values against real average historical sales. This system architecture allows not only accurate, item-specific demand forecasting but also visual, intuitive service delivery to end-users.

Results and discussion

Overall model outputs

Following model training across approximately 384 individual menu items, a comprehensive set of outputs was generated:

• Loss and Accuracy Curves: For each menu item, training loss and classification accuracy curves were plotted, enabling evaluation of convergence behavior and overfitting tendencies (Fig. 2A).

• Class-wise Softmax Probability Distributions: Histograms of the predicted softmax output probabilities were generated for each class (Fig. 2B). These visualizations provided insights into the confidence level of the model’s predictions for each sales category.

• Daily Sales Scatter Plots: For selected menu items, scatter plots were drawn to visualize the relationship between the cumulative day number (pDay) and actual daily sales (Fig. 2C). This allowed intuitive exploration of seasonal trends and demand variability over time.

• Model Artifacts: Two types of model artifacts were stored for each menu item: A full prediction model file (.h5) containing the network architecture, optimizer settings, and trained weights (Fig. 2D). A separate weights-only file (.weights.h5) capturing just the trained parameters.

These outputs collectively support both in-depth model performance analysis and practical deployment options for production environments.

Fig. 2.

Visualization results for Menu35: (A) training loss and accuracy curves, (B) class-wise softmax probability distributions, (C) scatter plot of daily sales, and (D) architecture of the trained menu-specific model (model.h5), including the input layer (5 features), two hidden layers (100 neurons each, ReLU activation), and the output layer(softmax activation for sales class prediction)

Overall model evaluation

The predictive performance across approximately 384 menu-specific models was summarized based on standard evaluation metrics. Table 2 presents the average, standard deviation, minimum, and maximum values for Mean Absolute Error (MAE), Root Mean Squared Error (RMSE), Mean Absolute Percentage Error (MAPE), and Pearson correlation coefficient (r). The average MAE and RMSE values indicate that the models achieved relatively low prediction errors across the menu items. The mean Pearson correlation coefficient of 0.7945 was calculated across all 384 individually trained models, reflecting strong overall agreement between predicted and actual values. Some individual cases exhibited lower correlation (e.g., menu #153 with r = 0.164), highlighting the heterogeneity in predictability across items. For completeness, we also report mean MAE (7.82 portions) and RMSE (11.25 portions) across all models.

Table 2.

Summary of model evaluation metrics across 384 menu items

Metric	Mean	Std. Dev.	Min	Max
MAE	0.0544	0.0470	0.0000	0.2297
RMSE	0.1472	0.0729	0.0000	0.3363
MAPE (%)	271,930%	235,047%	0.0%	1,148,517%
Pearson r	0.7945	0.2175	0.3572	0.9999

However, MAPE exhibited significant variability, largely attributed to instances of extremely low actual sales values causing percentage-based error magnification. Such behavior is common in classification-based demand prediction when rare sales events are involved.

Interpretation of evaluation results

A combination of low MAE and RMSE suggests that the model provides accurate and stable predictions, with few large deviations. A moderate MAPE (~ 14%) indicates that predictions are reasonably proportional to actual values but may leave room for improvement, especially during volatile periods such as holidays or promotions. The observed Pearson value (~ 0.164) implies a positive but weak linear correlation overall, suggesting that while the model captures basic seasonal and weather-driven trends, it could benefit from the inclusion of additional explanatory variables like promotions or local events. For example, in menu item #153, the Pearson correlation coefficient was approximately 0.164, indicating weak correlation for that specific case. However, the overall mean Pearson correlation across all 384 menu-specific models was 0.7945 (Table 2), suggesting a generally strong predictive performance.

By leveraging multiple complementary metrics, the evaluation provides a more holistic understanding of model performance and identifies specific areas where predictive capabilities could be enhanced in future work. All evaluations were conducted exclusively on the validation dataset to avoid information leakage and to accurately reflect the model’s generalization ability to unseen data.

Interpretation of MAPE outliers

Although MAPE is a useful metric for evaluating proportional errors, its sensitivity to small denominators (actual sales approaching zero) often results in artificially inflated error percentages. In the context of restaurant menu sales prediction, items with consistently low or sporadic sales frequency (e.g., niche menu items, seasonal specials) can disproportionately skew MAPE values. Therefore, while MAPE provides general guidance, other absolute error metrics (MAE, RMSE) and the Pearson correlation coefficient were primarily emphasized for model evaluation.

Prediction results visualization

In addition to standard evaluation metrics, the system provides flexible visualization options for predicted and actual sales volumes. Menu items can be sorted either by predicted sales volume or by actual historical sales volume, depending on the user's analytic focus. Sorting by predicted sales volumes enables the identification of menu items that are expected to perform strongly in the future based on model predictions. Sorting by actual historical sales volumes allows users to review which items have traditionally achieved high sales, aiding in trend analysis and historical benchmarking.

Representative examples of these visualizations are presented in Fig. 3A and B. To facilitate user input and enhance usability, a graphical user interface (GUI) was developed, as illustrated in Fig. 4. This interface allows users to intuitively specify key parameters without command-line interaction by providing date selection through an interactive calendar widget, expected temperature and rainfall input via horizontal slider bars, selection of the number of top menu items to display through radio buttons in increments of 10, and holiday and weekend indicators via checkboxes. Moreover, to improve user transparency and interactivity during the prediction process, a real-time status display was integrated into the GUI. Moreover, to improve user transparency and interactivity during the prediction process, a real-time status display was integrated into the GUI. As the model sequentially processes each menu category, the status label dynamically updates, showing the current progress in the form of "Calculating n/total menus…" messages, and future versions will incorporate estimated time-to-completion to enhance user feedback.

Fig. 3.

Bar plots comparing real average sales and predicted sales volumes for Menu items: (A) sorted by predicted sales volumes (descending) and (B) sorted by actual historical sales volumes (descending)

Fig. 4.

Graphical user interface (GUI) for user input of prediction parameters, including date, temperature, rainfall, number of top results, and indicators for holidays and weekends

The GUI system was developed as a desktop-based prototype for offline testing using Python’s Tkinter. While currently not deployed online, it is designed to support future integration into cloud-based platforms or point-of-sale (POS) systems. A pilot usability study involving restaurant operators is being planned to assess its operational value and user-friendliness in real settings. This GUI-based input system improves the overall user experience, reduces data entry errors, and ensures consistent formatting of input features for the prediction model. It also facilitates real-time, dynamic prediction scenarios, enabling operators to simulate different conditions based on upcoming weather forecasts or calendar events.

Based on these results, several future perspectives are proposed to extend the application of AI-driven demand forecasting systems across the broader food industry. The implications of this research extend far beyond individual restaurant operations. With the rapid digitalization of the food sector, artificial intelligence is poised to transform multiple dimensions of the food supply chain, from production to consumption.

In the broader context of food systems, predictive analytics can enable real-time inventory management not only in restaurants but also across wider food distribution networks. By forecasting demand more accurately, businesses can reduce overproduction, minimize spoilage, and enhance sustainability efforts (Kamilaris et al., 2017), which is especially crucial in perishable goods logistics where shelf life is limited. Furthermore, food waste remains a global concern, with approximately 30–40% of produced food lost annually (Sánchez-Teba et al., 2021). Demand forecasting at both the micro (restaurant) and macro (retail, wholesale) levels can significantly contribute to reducing unnecessary food disposal. This can be further refined by incorporating weather forecasts, public event calendars, and even social media sentiment into predictive models to strengthen accuracy and waste prevention strategies.

Additionally, AI has the potential to personalize the dining experience by predicting individual or group-level food preferences based on contextual factors such as weather, time, and location. Such personalization could be implemented through food delivery platforms, smart restaurant menus, or automated grocery shopping assistants (Elsweiler et al., 2012). On the production side, AI-driven demand forecasting can also assist farmers in planning crop yields more efficiently, synchronizing supply with real-time consumption trends (Wolfert et al., 2017). This coordination reduces the mismatch between production and market needs, ultimately improving profitability while mitigating environmental impact.

In light of recent global disruptions such as the COVID-19 pandemic and climate-related disasters, the vulnerability of food systems has been underscored. Predictive models that incorporate environmental and economic indicators can enhance resilience by facilitating quicker and more intelligent adaptations to supply shocks (Reardon et al., 2020). Moreover, AI can contribute meaningfully to ensuring equitable food distribution by forecasting potential food scarcity zones and proactively guiding resource allocation, thus supporting global hunger mitigation strategies (Dhal and Kar, 2024). Although the current study focuses on restaurant-level demand forecasting, the core methodology—namely the integration of real-world variables with AI models—can be scaled to applications in agriculture, food logistics, and equity-driven resource planning. Anchoring these future applications in restaurant operations positions the system as a foundational model for broader transformation of food systems.

Despite these promising directions, several challenges must be addressed for effective adoption of AI in the food sector. These include ensuring data quality and representativeness for model reliability, addressing ethical issues surrounding bias, fairness, and transparency in AI predictions (Floridi and Cowls, 2019), and navigating the substantial investment and interdisciplinary collaboration required for integration with existing food supply chain infrastructures.

This study demonstrates the efficacy of combining meteorological and sales data in AI-driven demand forecasting, yielding tangible benefits at the restaurant level. Looking ahead, scaling these techniques across the broader food industry holds considerable promise for creating smarter, more sustainable, and resilient food ecosystems. With the convergence of AI and IoT technologies and the adoption of responsible deployment practices, a new era of data-driven optimization in food systems could emerge—benefiting businesses, consumers, and the global environment alike.

Declarations

Conflict of interest

The authors have no conflicts of interest to declare.