AI Hospitality Alliance
Back to Research
Academic ResearchJanuary 1, 2024International Journal of Hospitality Management

Decoding the future: Proposing an interpretable machine learning model for hotel occupancy forecasting using principal component analysis

Most AI forecasting models are 'black boxes' that revenue managers don't fully trust — and tools you don't trust don't get used consistently. This paper proposes a machine learning approach that groups historical booking patterns into recognizable clusters (early corporate, last-minute leisure, event-driven spike, etc.) before applying a forecasting model to each, making the prediction explainable: not just 'occupancy will be 87%' but 'because booking pace matches the early corporate pattern, which historically reaches 87%.' The result is both more accurate than standard methods and more operationally useful, because revenue managers can see the reasoning and make better decisions around it.

Authors

Daniele Contessi, Luciano Viverit, Luis Nobre Pereira, Cindy Yoonjoung Heo

Article content

Hotel revenue managers are under consistent pressure to justify their forecasts — not just to get the number right, but to explain to GMs, owners, and asset managers why the model is saying what it's saying. This paper addresses one of the most persistent frustrations in AI-based revenue management: prediction accuracy has improved dramatically with machine learning, but most ML models are "black boxes" that produce good answers without being able to explain them. The paper proposes an approach that maintains the accuracy gains of machine learning while making the model's reasoning transparent and auditable.

The methodology is built around Principal Component Analysis (PCA), a statistical technique that reduces the complexity of booking pattern data into a smaller set of meaningful dimensions. Instead of a model that processes hundreds of individual variables simultaneously in ways that resist interpretation, the PCA approach first groups booking curve behaviors into recognizable typologies — clusters that revenue managers can actually recognize as familiar patterns from their own experience. Think: early corporate bookers, last-minute leisure compression, event-driven spikes, shoulder-period drag. These clusters are identified automatically from historical data rather than defined manually, which means they reflect the actual patterns in your property's booking history rather than generic industry templates.

Once booking patterns are organized into these interpretable clusters, a pickup model is applied to each cluster separately — generating forecasts that are not just more accurate than single-model approaches, but also explainable in terms revenue managers understand. When a forecast says occupancy will hit 87% on a specific date, the interpretable model can also say: "this is because booking pace to date matches the early corporate pattern, which historically reaches 87% by day-of." That's a different and more actionable piece of information than a black-box prediction of 87%.

The commercial benefits are documented in the paper with real data. Forecast accuracy improved meaningfully over both traditional methods and standard ML approaches. But the paper argues that accuracy improvement understates the full value of interpretability. Revenue managers who understand and trust their forecast models make better decisions — they're more willing to hold rate in the face of short-term booking pace pressure, more confident in their yield decisions, and more effective at communicating their strategy to stakeholders who don't have a forecasting background.

For technology buyers evaluating forecasting systems, this paper provides a useful evaluation framework. The key question to ask any ML-based forecasting vendor is not just "how accurate is your model?" but "can you show me why the model is making this specific prediction?" Vendors who can answer that question meaningfully are building toward interpretable AI. Those who can't — who say the model is too complex to explain — are optimizing for accuracy at the expense of operability. In practice, a model that revenue managers trust and use consistently will outperform a more accurate model that feels like a black box and gets overridden by gut feel in high-stakes decisions.

For asset managers and owners, the interpretability dimension has governance implications: AI forecasting systems that produce explainable outputs can be audited, debated, and improved. Those that can't become an organizational dependency without accountability.

Tags

Revenue ManagementOperationsEthics

Related research

Academic ResearchFebruary 3, 2023

Artificial Intelligence in the Tourism Industry: An Overview of Reviews

This paper is helpful because it shows where AI already has strong evidence behind it and where more caution is still needed. The clearest benefits are in efficiency, personalization, forecasting, and customer communication. At the same time, the research highlights issues that operators cannot ignore, including privacy, guest trust, staff impact, and the challenge of adopting AI in smaller or less data-rich businesses. For hotel leaders, the practical message is to move ahead in areas where the value is well established, while taking a more deliberate approach to ethics, employee transition, and whether a tool genuinely fits the business.

Academic ResearchNovember 1, 2025

Adoption of Artificial Intelligence (AI) Technology in Enhancing Tourist Experience: A Conceptual Model

This article proposes a conceptual model for adopting Artificial Intelligence (AI) technology to enhance tourist experiences. It highlights the potential of AI to transform tourism by personalizing services, improving operational efficiency, and enriching guest interactions, offering practical insights for hospitality professionals aiming to leverage AI in their strategies.

For Professors

Submit article for consideration

If you are a professor or researcher and would like to suggest a publicly available article for inclusion in the Research Hub, you can submit it for review and possible inclusion through our dedicated submission form.