How automation can forecast uncertainty, drive demand

Uncertainty is nothing new for hospitality. Hoteliers have struggled against shifts in unknowable economic forces and inexplicable trends for years, and revenue managers have been on the front lines trying to cut losses and improve forecast accuracy. Despite advances in automated decision-making systems, a lack of trust in technology has traditionally held the industry back from forecasting the unexpected. 

If the past two years have shown us nothing else, it has forced the industry to do more with less, including the need for sophisticated systems that consider a comprehensive data set. Best-in-class revenue management systems can measure uncertainty by creating an unconstrained demand forecast, which in turn produces a more accurate holistic forecast overall.

With the proliferation of hotel technology that targets forecasting and revenue management, it is essential to consider the pitfalls for hoteliers when looking at rules-based systems versus automated systems. What should they look for, and how do they avoid unintentionally putting their decision-making into a silo by choosing the wrong optimization type?

Let’s break down the differences in today’s RMS systems. 

Rule-Based Systems versus Artificial Intelligence

We need to begin by defining the rule-based system. Without getting too technical, these are logical programs that use pre-defined rules to make deductions and choices that perform automated actions. How does a rule-based system work? Rule-based systems conduct the work based on rules. These rules outline triggers and the actions that should follow (or are triggered). 

A trigger might be the occupancy reaching a certain threshold. These rules most often take the form of “If/Then” statements. The “If” statement outlines the trigger. A “Then” statement specifies the action to complete. For example, if occupancy reaches 20 percent, then increase the rate by 10 percent. 

So, if you want to create a rule-based system capable of handling 100 different actions, you’d have to write 100 different rules. If you’re going to update the system and add actions, you would need to write new rules. 

In short, you use rules to tell a machine what to do, and the machine will do exactly as you tell it. From there, a rule-based system will execute the actions until you tell it to stop. One caveat to remember is that if you tell it to do something incorrectly, it will do it incorrectly.
Now let’s look at what a rule-based system is not. Due to early use in the fields, rule-based systems are commonly confused with artificial intelligence and machine learning. However, they are neither. 

It’s easy to confuse the two as they can look very similar. Both involve machines completing tasks, seemingly on their own. The difference is that AI can determine the action to take; it can learn and adapt. Meanwhile, rule-based systems do exactly as instructed by a human. In other words, unlike artificial intelligence and machine learning, the actions carried out by rule-based systems (the rules that they follow) are determined by a human. 

The system doesn’t work it out for itself or intelligently make decisions. A rule-based system won’t change or update on its own and won’t "learn" from mistakes. Instead, rule-based systems follow the rules laid out by humans. They may work for simple problems where there is sufficient human expertise to enable the system’s design, as is the case with medical diagnosis or offshore drilling, for instance.

From a property-level perspective, consider the following. In a rule-based system, knowledge is sitting in a revenue manager’s head and is translated into the computer via actionable rules. 

Results are highly variable and depend on the quality of the revenue manager, their knowledge of the marketplace, diligence exercised in creating rules and time investment on a regular basis in creating and adjusting rules. The system will fall short of its purpose when revenue managers go on holidays or are indisposed due to other priorities. 

Remember, rule-based systems require a series of nested If … Then … Else statements to create a scenario for decision making. This mechanism can never encompass the actual situation. More importantly, if one revenue manager leaves, the new revenue manager is incapable of understanding and adopting the rules left by the departing manager. They will need to start from scratch. 

In a rule-based system, the ROI constraint does not justify how much time a revenue manager would take to use knowledge from their head and marry it with all the dynamic information coming in to translate into adjustments needed to the rules. Unadjusted rules leave significant money on the table. 

Forecasting Using Automated Decision Systems in Revenue Management

In reality, revenue management and price optimization decisions constantly being made in the hotel world encompass too many variables for the human mind to analyze in real time. Decision systems need to model real-life situations and provide a decision-making capability. 

This is based on the merits and demerits of individual cases as viewed in the context of the hotel’s global business objectives. These systems use forecasting and proactive mathematical optimization through artificial intelligence. 

Why is that important? Let’s look at the complexity of a hotel’s stay structure. 

From a data perspective, a hotel’s stay structure is a very complex time-based network built around multiple interacting arrival days and length of stay combinations. It must consider qualified and unqualified guest types, each with its own distinct demand. Mix in the uncertainty with respect to booking patterns, length of stay patterns, day of week patterns, seasonality, no-show rates, etc., and you have complexity beyond human analysis capabilities. 

The objective of revenue management is to accept or deny rooms to various qualified guest types for both transient and group business based on their different arrival days and length of stay combinations. Then finding the optimal price for the unqualified guest in a manner that maximizes the overall profit from the time-based network.

All decisions (even that of selling one room for a single night) impact all the other decisions since that room could be sold in combination with other nights for more overall revenue. There is a significant combinatorial complexity arising from the network nature of the problem. For example, stay itineraries that are a combination of adjacent busy and non-busy nights may yield more overall revenue as opposed to stay itineraries that are a combination of only busy nights.

Decisions of accepting or rejecting reservation requests must be made continuously and up to a year or more in advance for future days. This introduces uncertainty and a lack of complete knowledge at the time those decisions need to be made. And they need to be made in real-time, in an uncertain environment, based on the forecast of the demand at the arrival day and length of stay and how far in advance the booking request is made while accounting for the possibility of no show and cancelations.

In AI/ML-based optimized RMS systems, all the knowledge is sitting in the computer in terms of data and highly optimized models. Decision-making is highly dynamic, sophisticated, and economical due to a computer’s ability to process large amounts of data and take hundreds of inputs while accounting for thousands of scenarios in a matter of minutes. 

Currently, known decision-support solutions based on a rule-based approach have had limited success in dealing with the network nature of the problem in an uncertain environment. Hoteliers need an automated system that produces globally optimal decisions in real-time in a network-wide sense while considering the uncertain nature of demand.

This is only possible with forecasting techniques that produce not only a forecast of average demand but also the uncertainty of demand. We need such a forecasting system to feed the decision optimization system based on ML as well as mathematical optimization to maximize global profit while managing the risk associated with the uncertain nature of the demand and the marketplace.

When Good Enough is NOT Good Enough

In general, there is room for both traditional rule-based approaches and ML-based approaches. The selection depends on many factors, including the complexity of the problem you are trying to solve, the amount of data you have access to, the structure of the underlying system delivering value, and how often the decisioning system needs to be updated. 

For revenue management and price optimization problems, it is self-evident that the complexity of the decision problem in an uncertain environment calls for an automated decision system based on mathematical optimization instead of rules based on human expertise. 

Some hoteliers have decided that rule-based systems are “good enough,” especially as they restaff and retrain. But have they considered the most critical gaps, known or unknown, while understanding math and behavior-related anomalies? Commercial leaders must consider how to rectify these gaps while keeping in mind the staffing impacts and internal decisions they must make. 

The primary goal of forecasting is to identify the full range of possibilities, not a limited set of imagined certainties. Whether a specific forecast turns out to be accurate is only part of the picture—even a broken clock is right twice a day. 

Above all, the forecaster’s task is to map uncertainty, for in a world where our actions in the present influence the future, uncertainty is opportunity. At the end of the day, knowing if the revenue management system they select is a rule-based or an automated decision system will make the difference in a better understanding of the unknown.

Ravi Mehrotra is co-founder and chief scientist for IDeaS.