3.1 Hotel revenue management demand forecasting

Demand forecasting is the process of making predictions or estimates about future observations of a target variable based on historical and present data. Hotel demand forecasting is the process of predicting, depending on the data available, reservations or occupancy rates for hotel rooms and other services but can also be used to predict other variables, like cancellations (^{Antonio et al., 2019}) or no-shows, and to support pricing optimisation through dynamic rates (^{Zhang et al., 2019}), maximising the revenue and profit. In the hospitality industry context, forecasting the demand is important for hotel managers and RM team to make inventory allocation and pricing decisions to maximise their revenues (^{Kaya et al., 2022}; Zhang et al., 2019). Forecasting is a critical element in hotel planners' and managers' decision-making process, which involves the strategic allocation of hotel resources to maximise profitability. Indeed, it has been recognised that the use of accurate forecasting methods is vital to ensure better application of dynamic pricing to popular strategies like RM because errors in pricing will negatively affect a hotel's financial performance if the forecasting is not precise (^{Caicedo-Torres & Payares, 2016}). In recent years, advances in ML and other AI-based methods have revolutionised the field of demand forecasting, allowing hotels to predict demand more accurately and adjust their pricing in real time (^{Huang & Zheng, 2023}). Despite these advances, gaps remain in our understanding of how best to use these techniques.

3.2 Hotel demand forecasting methods based on artificial intelligence

Several forecasting methods associated with AI solutions were identified in the literature review. A mix of statistical methods and AI models was also found. The traditional statistical methods used to forecast hotel demand are time series analysis and econometric models. The methods within the time series analysis category include the moving average, which consists of calculating averages of demand over a specified period to smooth out fluctuations, and exponential smoothing, which assigns exponentially decreasing weights to past demand observations to give more importance to recent data. In the exponential smoothing subcategory, we can find a wide range of methods, such as the Holt and Holt-Winters method. Another traditional method is the autoregressive integrated moving average (ARIMA), which combines autoregressive, differencing, and moving average components to model and forecast time series data. The ARIMA model is often implemented through the Box-Jenkins approach, also known as the ARIMA Box-Jenkins methodology. This methodology provides a systematic framework for identifying, estimating, and validating a suitable ARIMA model for a given time series. SARIMA is commonly used to extend the ARIMA method to account for seasonality in data analysis (^{Antonio et al., 2019}; ^{Kaya et al., 2022}; ^{Pereira & Cerqueira, 2022}; ^{Phumchusri & Ungtrakul, 2020}; ^{Salamanis et al., 2022}; ^{Schwartz et al., 2021}; ^{Wang & Duggasani, 2020}). In the econometric category of models, we can find, for example, the autoregressive distributed lag model, error correction model, and vector autoregression (VAR) (^{Viverit et al., 2023}).

With the evolution of technology and computer capabilities, ML algorithms have been used to mix traditional statistical models with ML, building a new approach to forecasting hotel demand (^{Doborjeh et al., 2022}). This approach involves the use of large amounts of data to train algorithms, allowing them to learn patterns and make predictions with greater accuracy. By combining traditional statistical methods with machine learning (^{Huang & Zheng, 2023}), hotels can improve their forecasting accuracy and make more informed pricing and inventory management decisions. Some of the ML models identified are ANNs (^{Ampountolas, 2021}; ^{Caicedo-Torres & Payares, 2016}; ^{Huang & Zheng, 2021}; ^{Kaya et al., 2022}; ^{Sánchez et al., 2020}; ^{Sánchez-Medina & C-Sánchez, 2020}), used to simulate the human brain system by using artificial layers (Doborjeh et al., 2022). ANNs have demonstrated good results as a model for managing nonlinear data; however, they do not demonstrate the same results in capturing long-term time dependence (^{Huang & Zheng, 2023}). To support the weakness of ANNs, long short-term memory (LSTM) (Huang & Zheng, 2021; ^{Salamanis et al., 2022}; ^{Wang & Duggasani, 2020}; ^{Wu et al., 2022}) is a type of recurrent neural network (RNN) architecture that can capture the temporal dependencies in hotel reservation data (^{Huang & Zheng, 2023}). A support vector machine (SVM) (Sánchez et al., 2020; Sánchez-Medina & C-Sánchez, 2020), used with the support vector regression (SVR) extension (^{Pereira & Cerqueira, 2022}; ^{Phumchusri & Ungtrakul, 2020}), allows users to map the data into a higher-dimensional space and identify the hyperplane that best fits the training data, permitting nonlinear relationships between predictors and demand. The random forest (RF) (Pereira & Cerqueira, 2022; Sánchez-Medina & C-Sánchez, 2020) is used to construct an ensemble of decision trees and aggregate their predictions to forecast the demand. These trees can capture nonlinear relationships and interactions among predictors, making them suitable for complex demand patterns. The gradient boosting machine (GBM) (Ampountolas, 2021; Sánchez et al., 2020) creates an ensemble of weak prediction models iteratively, learning from previous mistakes. Models like XGBoost (^{Antonio et al., 2019}) or LightGBM are popular implementations of gradient-boosting algorithms. Linear regression (LR), as a foundational statistical model, contains various extensions for flexible and strong modelling. Ridge regression, kernel ridge regression (Caicedo-Torres & Payares, 2016), and lasso regression (Pereira & Cerqueira, 2022) are used for data regularisation. Due to some limitations of the traditional forecasting models, other AI-based statistical approaches have been used to complement them and improve their accuracy. The Cross-Industry Standard Process for Data Mining (CRISP-DM) (^{Antonio et al., 2019}), clustering booking curves (^{Viverit et al., 2023}), deep factorisation Machine (DeepFM), Seq2Seq (^{Zhang et al., 2019}), K-means clustering (^{Kaya et al., 2022}), generic algorithm (GA) (Sánchez-Medina & C-Sánchez, 2020), multi-layer perceptron, and radial basis function networks (Caicedo-Torres & Payares, 2016) are other ML models mentioned in the literature reviewed. Some articles propose extra background methods to fine-tune the data collected. Several methods and models are identified that can be used for different purposes, and the choice of method should depend on the specific research question and available data. It is important to consider each approach's strengths, limitations, and results carefully before selecting a method or model for analysis. A more in-depth analysis will be applied to this study later.

3.3 Hotel data set

As mentioned earlier, the forecasting method used will depend on the kind of variable that should be predicted and the forecasting horizon (^{Koupriouchina et al., 2014}). Several variables can be set to forecast, such as occupancy, reservations, prices, and cancellations, which can be forecasted in the hotel business to assist the revenue manager's decision-making. For each variable, a set of data should be collected and processed to apply the corresponding method or model. ^{Caicedo-Torres and Payares (2016}) argue that occupation forecasts can be grouped into two categories: historical reservation models and advanced reservation models. The historical reservation model focuses on the time series modelling problem to understand historical patterns and predict the future. The advanced reservation model uses the concept of "pick-up", adding a certain number of future bookings to existing reservations on the books in a certain advanced period of days. According to the literature review, six data types are used to perform forecasting: room occupation (^{Ampountolas, 2021}; ^{Ampountolas & Legg, 2021}; ^{Antonio, de Almeida, et al., 2019}, 2019; Caicedo-Torres & Payares, 2016; ^{Chen et al., 2023}; ^{Huang & Zheng, 2021}; ^{Kaya et al., 2022}; ^{Pereira & Cerqueira, 2022}; ^{Phumchusri & Ungtrakul, 2020}; ^{Sánchez et al., 2020}; ^{Schwartz et al., 2021}; ^{Viverit et al., 2023}a; ^{Wang & Duggasani, 2020}; ^{Webb et al., 2020}), room reservations (^{Ampountolas, 2021}; ^{Antonio, de Almeida et al., 2019}; ^{Caicedo-Torres & Payares, 2016}; ^{Kaya et al., 2022}; ^{Salamanis et al., 2022}; ^{Sánchez-Medina & C-Sánchez, 2020}; ^{Viverit et al., 2023a}; ^{Wang & Duggasani, 2020}), room cancellations (^{Antonio, de Almeida, et al., 2019}, ²⁰¹⁹; ^{Rakesh et al., 2022}; Sánchez-Medina & C-Sánchez, 2020), online reviews (^{Wu et al., 2022}), social media keywords (^{Ampountolas & Legg, 2021}), and prices (^{Kaya et al., 2022}; ^{Wang & Duggasani, 2020}; ^{Zhang et al., 2019}). Due to the interconnection between data sets, some scholars and researchers use a mixed-data approach to achieve more accurate forecasting results. This involves combining multiple data types and analysing them together to gain a more comprehensive understanding of the factors influencing hotel demand. Additionally, ML algorithms can be applied to these mixed data sets to improve the accuracy of forecasting models further (Viverit et al., 2023b). A hotel data set used in forecasting demand comparisons is presented in Table 2 to offer a better understanding.

Additional information is used to improve the forecasting results in some of the mixed methods mentioned.^{Wu et al. (2022}) combine customer reviews from two popular online platforms with the historical hotel demand for specific dates to analyse the impact on forecasting demand using the LSTM and ARIMAX methods. ^{Antonio et al. (2019}b) combine hotel reservations and cancellations using the Cross-Industry Standard Process Model (CRISP-DM), a data-mining method, to predict future hotel cancellations. Another complex example is the dynamic price prediction research from ^{Zhang et al. (2019}), in which 759 hotels in China are utilised and divided into two datasets to forecast the rates in the off-season and peak season using a comparative competitor analysis base price setting. ^{Salamanis et al. (2022}) research the impact of weather on hotel reservation cancellations, combining historical time series hotel data on reservations, cancellations, and weather conditions in a DL methodology using an LSTM layer. These are some examples showing how complex a forecasting model can be and the kind of data used to perform the model.

3.4 Target forecast objectives

As mentioned earlier, the data collected from hotel databases or external sources can be used to improve the accuracy of forecasts of variables of interest for maximising revenue. In the literature review, two targets were identified. i) Cancellation prediction was found in six studies (^{Antonio et al., 2019}; ^{Chen et al., 2023}; ^{Rakesh et al., 2022}; ^{Sánchez et al., 2020}; ^{Sánchez-Medina & C-Sánchez, 2020}): along this research line, the aim of the investigation focuses on finding the best forecasting method to predict future cancellations since this event can contribute negatively to more accurate forecasts and revenue performance. ii) Occupation forecasts were found in six studies (^{Ampountolas, 2021}; ^{Caicedo-Torres & Payares, 2016}; ^{Huang & Zheng, 2021}; ^{Kaya et al., 2022}; ^{Viverit et al., 2023}; ^{Wu et al., 2022}). Occupation forecasts for better price prediction were found in one study (^{Zhang et al., 2019}). As suggested by the author, due to the development of digital and computer reservation systems, forecasting for dynamic pricing is another big challenge in hotel RM. To resolve this issue, a new forecasting model is proposed using DeepFM and Seq2eq, two ML models, one combining DL and factorisation and the other a type of NN architecture for sequence-to-sequence learning tasks, and a data-driven dynamic pricing methodology to recommend a more rational price than revenue managers' rule-based technique. In the role of RM, occupation forecasting is the most important for better operational and tactical decisions (^{Wang & Duggasani, 2020}). Therefore, it is crucial for hotels to invest in advanced technology and data analysis tools to forecast the hotel demand accurately and adjust their prices accordingly. Additionally, RM strategies should consider external factors, such as competition and market trends, to stay ahead of the game.

3.5 Comparative studies of forecasting methods

Seven papers comparing forecasting methods based on the traditional statistical approach or comparing hybrid models using traditional and ML models were identified. ^{Schwartz et al. (2021}) explore the interactions between two methods of accuracy enhancement: forecast combination and human learning. ^{Pereira and Cerqueira (2022}) analyse two short-term hotel demand forecasting methods for lead times up to 14 days ahead. The paper focuses on ML models, such as support vector regression, random forest, and ANNs, and compares them with traditional statistical methods, such as seasonal naive and exponential smoothing methods for double seasonality. The study aims to provide insights into the potential benefits of using ML models in hotel demand forecasting for RM. ^{Wang and Duggasani (2020}) propose two LSTM models for forecasting hotel reservations and compare their performance with that of traditional ML algorithms. The authors aim to address the limitations of the existing ML techniques in modelling the sequential dependence of bookings over time and to demonstrate the effectiveness of LSTM models in predicting the demand for hotel rooms. ^{Phumchusri and Ungtrakul (2020}) propose hotel daily demand forecasting models using both time series models and ML techniques. The article aims to investigate the use of the transformation of time series data as an input variable for forecasting hotel demand. It also compares the forecasting accuracy of different methods and provides insights into their strengths and weaknesses. ^{Salamanis et al. (2022}) propose introducing a DL-based approach for forecasting long-term tourism demand. The authors suggest two different models based on the long short-term memory network (LSTM), which can incorporate data from exogenous variables, such as weather-related parameters, to generate accurate long-term predictions. The proposed models are evaluated using real data from three hotels in Greece and compared with traditional statistical models. ^{Ampountolas (2021}) aims to compare the forecasting performance of three different models (SARIMAX, NN, and GARCH) in predicting the daily hotel demand. ^{Webb et al. (2020}), whose study focus is on evaluating the efficacy of advanced booking methods in RM for hotels located near national parks in the USA, aim to compare the accuracy and effectiveness of various statistical and ML models in forecasting the demand in the advanced booking environment and to identify best practices for RM in this context. These studies suggest that incorporating ML techniques into RM can lead to more accurate and efficient forecasting, ultimately resulting in increased business revenue and profitability. Therefore, companies need to consider implementing these hybrid models in their RM strategies.

3.6 Artificial intelligence forecasting advantages

ML is one of the most fascinating recent AI technologies. Many applications that we use daily involve learning algorithms, which allow the software to adapt and improve over time (^{Doborjeh et al., 2022}). By incorporating ML into RM, businesses can make more accurate predictions about customer behaviour and adjust their pricing accordingly. This approach can lead to increased profits and a competitive advantage in the market. However, several kinds of ML techniques are available, and it is important for businesses to choose the one that best fits their needs. Some ML techniques are supervised learning, unsupervised learning, and reinforcement learning (^{Das et al., 2015}). Each technique has its strengths and weaknesses, and businesses should carefully consider which one to use based on their specific goals and data. Additionally, they should ensure that they have high-quality data to feed into the ML algorithms to achieve the best results. As mentioned by scholars and researchers regarding the data used in forecasting hotel RM, some of these techniques, like demand forecasting and pricing optimisation, can be applied to predict and support revenue managers' decisions. Several models and methods have been applied. In the review chapter on forecasting results, we aim to identify the ML model that can best predict hotel demand. However, after screening the data, it is suggested that the best model depends on the data available and the problem being addressed. Nevertheless, ANNs, support vector machines, and linear regression are some of the most prominent hotel demand forecasting methods (^{Huang & Zheng, 2023}). The complex relationships between the input variables and the intended output can be captured effectively by ANNs (^{Ampountolas, 2021}; Ampountolas & Legg, 2021; ^{Caicedo-Torres & Payares, 2016}; Huang & Zheng, 2021; ^{Kaya et al., 2022}; ^{Pereira & Cerqueira, 2022}; ^{Phumchusri & Ungtrakul, 2020}; ^{Sánchez et al., 2020}; ^{Sánchez-Medina & C-Sánchez, 2020}). Support vector machines are a form of supervised ML that can efficiently identify the optimal linear boundary between distinct data classes (Pereira & Cerqueira, 2022; Phumchusri & Ungtrakul, 2020; Sánchez et al., 2020; Sánchez-Medina & C-Sánchez, 2020). Using historical data, linear regression is a straightforward but effective technique for forecasting demand (^{Salamanis et al., 2022}; Sánchez-Medina & C-Sánchez, 2020; ^{Webb et al., 2020}; ^{Wu et al., 2022}).

3.7 Analyses of forecasting methods by models, results, and limitations

This chapter examines the application of various statistical methods and ML algorithms for forecasting hotel outcomes. By comparing and contrasting these methods, we hope to shed light on their advantages, disadvantages, and potential for improving forecasting accuracy in the hospitality industry. Forecasting is a critical activity for dynamic pricing strategies in RM, and errors in pricing will negatively affect hotel financial performance if forecasting is not precise (^{Caicedo-Torres & Payares, 2016}). Strategic planning and decision-making rely heavily on accurately forecasting hotel results in the highly competitive hospitality industry. Forecasting assists hotel managers in anticipating demand patterns, optimising pricing strategies, allocating resources efficiently, and maximising revenue and visitor satisfaction (^{Phumchusri & Ungtrakul, 2020}). Historically, statistical methods have been used to predict hotel performance; however, with the advent of ML techniques, new opportunities for more accurate and sophisticated forecasting models have emerged. From the literature reviewed, a comparison was made to establish a relationship between the methodology used (statistical and AI models), the targeted forecast, the results, and the limitations. To better understand and compare, the articles were divided by forecast target: demand, cancellations, and prices. Table 3 compares the studies that concern demand forecasting; Table 4 presents studies that forecast booking cancellations; and Table 5 contains studies that predict pricing recommendations.

3.8 Recommended future research

Grounded on the analysed articles, Table 6 presents in the main future lines of research proposed by the authors. The table summarises the topics mentioned in the various articles and the authors who suggested them. The identified future lines of research pave the way for new research possibilities and contribute to the advancement of knowledge in the field. The systematisation of these suggestions allows a broad and organised view of the directions that studies can follow.

4.1 Theoretical Implications

The study makes a theoretical contribution by conducting a systematic literature review to identify the current state of the art of AI methods for forecasting hotel demand. It provides insights into the effectiveness of AI-based solutions, including machine learning (ML) and artificial neural networks (ANNs), for demand forecasting in the hotel industry. The study highlights AI technology's potential to significantly improve forecasting accuracy and facilitate data-driven decisions for hotel management. By comparing traditional forecasting models with AI-based algorithms, the study sheds light on their advantages, disadvantages, and potential for improving forecasting accuracy in the hospitality industry. The study emphasises the importance of incorporating accurate hotel demand forecasts into revenue management (RM) to make better pricing decisions, leading to increased profits and a competitive advantage in the market. Overall, the study contributes to the theoretical understanding of AI-based demand forecasting in the hotel industry, highlighting the potential benefits and challenges associated with the use of AI techniques. It also suggests future research directions to enhance further the effectiveness of AI techniques in revenue management and forecasting in the hospitality industry.

4.2 Practical Implications

Good accuracy demand forecasting for RM is vital for all practitioners, scholars, and researchers. This result implies a daily decision driver for hotel managers and RM teams. More accurate demand forecasts support better decisions that improve both operational and financial success in the hotel business. An accurate forecasting model is necessary in a dynamic and competitive distribution environment where dynamic pricing relies on daily decision-making. Conversely, an accurate forecast of the occupancy rate allows hotel managers to plan better issues such as inventory, supplies, and workforce(^{Caicedo-Torres & Payares, 2016}). The results reveal that AI-based forecasting models tend to outperform traditional models. For this reason, we think that all practitioners, whether they are hotel managers, revenue managers, or revenue management system developers, should move one step ahead and start learning, implementing, and using AI-based forecasting models of hotel demand.