Reducing human error in revenue management decision-making Recent revenue management (RM) research has demonstrated what appear to be systematic deviations from optimization models of decision-making. It turns out that decision complexity exacerbates specific elements of biased RM decision-making. There are ways, however, in which managers can use this information to improve the quality of RM decision-making in their companies. More fundamentally, companies can learn from these predictable patterns to design programs that minimize RM decisionmaking error. 02 Volume 5 Issue 4
Authors Elliot Bendoly Associate Professor Goizueta Business School Emory University, US Michael Alan Sacks Associate Professor Goizueta Business School Emory University, US 03
Reducing human error in revenue management decision-making RM is the science of maximizing the sales and price per unit of inventory. Have you noticed recently that airline flights almost always fill to capacity? If you re a seasoned traveler, you may have also observed that full flights are more common today compared with even only a few years ago. The reason: sophisticated RM systems that maximize the yield for seat sales. RM is the science of maximizing the sales and price per unit of inventory. The airlines industry has developed highly sophisticated dynamic computer models for its pricing structure. These models make real-time pricing decisions based on the number of seats available at the time of purchase, amount of time remaining prior to the flight, and past purchase history for that specific flight. With a significant amount of meaningful data available, these systems optimize revenue and seat capacity, largely eliminating the role of human judgment in the process. However, the airline industry is an extreme case where all such RM decisions can be made via optimization models. For most other industries, a mix of computer models and human judgment are required for RM decision-making. The hotel industry serves as an ideal example. RM packages designed for the hotel industry allow for real-time monitoring of capacity utilization, integration with forecasts, and advanced analytics. However, individual RM judgment is still viewed as essential to appropriately manage the nuances confronted in realworld RM settings. For example, complex large group bookings, last-minute purchases, cancelations, weather-related changes and trip adjustments all complicate the RM system. Hence, the effectiveness of many of the applications of these systems, and their associated operating policies, is still very much human driven. Predictable human errors in RM judgment Given the central role of human decisionmaking in RM decisions, one might question the extent to which people make optimal versus suboptimal RM decisions. A wealth of research shows that human beings make common, predictable errors in general decision-making. However, little is known about decision-making errors in RM judgment specifically, and how to mitigate them moving forward. Just how good are people at making challenging RM decisions? It turns out that the answer is complex, yet understanding decision-making errors in RM can make a big difference in optimizing RM outcomes. A recent study by Bendoly 1 sheds light on this very question. The author sought to measure the extent to which human decision-making in an RM context differs from optimal levels, then study the reasons for such decision errors. The author carefully designed a controlled experiment to test pricing decisions across variations of resource capacity and time urgency. In the experiment, a revenue manager had S total units of capacity available for allocation to clients over a time frame of T discrete periods. The goal of the experiment was to maximize total revenue by selling as much of S as possible, and at the highest possible price levels, prior to the end of the total time T available (think of a hotel manager selling rooms for a specific date in the future). The experiment began at time T=40 (the full amount of time) then slowly advanced until T=0 (no time remaining), with subjects charged to sell five units at the highest possible levels of revenue. Subjects would receive computerized bids along the way, which they had to either accept or reject. The actual decision faced in each period is, therefore, whether to allocate the requested capacity, thus generating an associated level of revenue, or reject the request, assumedly with the hope of allocating that capacity to a higherpaying customer later (with the associated risk that such an opportunity might never surface). To address how complexity affects optimal RM decision-making, three experimental contexts were designed (see Figure 1). Experiment A was the most straightforward and experiment C was the most complex. Thus, the goal was to assess human decision-making on RM across three levels of increasing complexity, then compare the results to the same decisions made by optimization models. The final twist in the experiment is perhaps the most intriguing. The author used the latest theories within behavioral operations to assess how, and why, subjects make distinct types of decisionmaking errors, specifically looking at the relationship between motivational levels and performance. The first motivational dynamic might be categorized as indifference, in which a lack of sufficient 1. E. Bendoly, Linking task conditions to physiology and judgment errors in RM systems. Production and Operations Management 20(6), pp. 860-876, 2011. 04 Volume 5 Issue 4
With too little challenge and plenty of time, people are more likely to reject offers they otherwise should accept. challenge in a task results in a low state of motivation for the individual worker. Typically, such a lack of motivation is ultimately associated with lower measures of objective performance in a task. The second dynamic revolves around increasing the level of challenge associated with a task by a sufficient, but not an excessive, amount. Challenging, yet attainable, tasks can raise a decision-maker s motivational level and increase attentiveness and arousal regarding the task, generally resulting in enhanced performance. The qualifier not an excessive amount is critical here, as excessive challenge raises the third dynamic: distraction, stress and a general reduction in motivation driven by a kind of hopelessness from perceptions of being overwhelmed by unattainable work expectations. Together, these findings imply an inverted-u relationship between challenge and performance outcomes (see Figure 2). Because the concurrent management of multiple blocks of capacity (in experiments B and C) can generally be viewed as more challenging than the management of a single block (in experiment A), one would expect to see differences in the dynamics of arousal, indifference and stress, as well as their impact on performance. The number of blocks simultaneously managed by an RM is, of course, only one possible contributor to the level of challenge associated with a task. Also ostensibly relevant to the management of these tasks, is the comparison of time remaining to capacity yet to sell. For example, it may be that individuals faced with seemingly high levels of capacity still unallocated close to the deadline may feel particularly challenged. In an attempt to fill this capacity and, hence, avoid non-occupancy, individuals might operate with lower than normatively modeled price thresholds. Figure 1. Assessing human decision-making on RM across three experiments of increasing complexity Experiment A B C Units of capacity available for sale by revenue manager One block of five units Two independent blocks of five units Four independent blocks of five units Example One hotel with five rooms Two hotels in different cities with five rooms Four hotels in different cities with five rooms City 1 City 1 City 3 Source: Authors own. City 1 City 2 City 2 City 4 05
Reducing human error in revenue management decision-making Given the central role of human decision-making in RM decisions, one might question the extent to which people make optimal versus suboptimal RM decisions. Alternatively, those with limited capacity remaining near the deadline may feel that earlier efforts to use ambitious thresholds yielded the result desired, and are sufficient in securing their net performance over the horizon considered. They may, in turn, be less likely to seek out ambitious revenue levels for the remaining capacity, either by virtue of a fundamental lack of motivation (indifference) or the view that pursuing ambitious revenues on such capacity is simply less realistic (overwhelming), given their own assumptions regarding the pool of demand. The result that follows might be a tendency to use less than normatively modeled thresholds at these low capacity levels, irrespective of the time remaining. The overall expectation was that low levels of time remaining, coupled with high levels of capacity yet to allocate, should be particularly associated with feelings of stress related to overwhelming workload. Additionally, such effects should be particularly relevant in increasingly complex contexts (experiments B and C) where multiple blocks of capacity are managed concurrently. These compounded issues, if in fact serving to generate stress rather than simply arousal, are expected to be associated with increases in accept errors (agreeing to bids priced at lower than necessary rates) more so than reject errors (taking a pass on bids that otherwise should be accepted). That is, people will accept bids that they should objectively reject due to the aforementioned biases. Conversely, with high amounts of time remaining and no imminent decision necessary, people may be more likely to make reject errors due to low motivation caused by under-stimulation. This would be most likely in the least complex situations (experiment A) as compared to more complex scenarios (experiments B and C). Measuring motivational state The only way to determine whether certain types of deviations from anticipated RM (i.e., certain errors) are being driven by either a lack of arousal or a sense of being overwhelmed, is to simultaneously study hypothetical markers of such emotional reactions more directly. Fortunately, objectively observable physiological markers of emotional and cognitive states, such as arousal (or lack thereof in the case of indifference) and stress, exist. Markers that have been studied in the past, range from neuroelectrical activity in the brain, as measurable by electrodes, or functional magnetic resonance imaging to heart rate variance. Physiological responses that require less obtrusive means for measurement, such as those measurable through the video-monitoring of the eye, have also proved useful and increasingly cost-effective. Eye tracking offers the advantage of allowing the simultaneous observation of multiple potentially idiosyncratic markers (e.g., pupil size, blink rate, x-y fixation). Pupil size has consistently been shown to increase upon heightened levels of mental workload and arousal. A typical range of pupil diameters for humans is 2mm 8mm, with relative variations from an individual s relaxed state (i.e., when not engaged with work) typically used to denote increases in arousal. In contrast to the observed linkage between pupil size and arousal, blink frequency is largely thought to reflect negative mood states, such as nervousness, stress and feeling overwhelmed by a task. An often-cited reference for this marker is the Nixon effect, referring to the former president s increased blink frequency (50 times per minute) during a discussion of his removal from office. Conversely, blink rate is thought to slow down when individuals engage in successful and comfortable problem-solving endeavors. Thus, the author utilized pupil size to measure arousal and motivation, and blink frequency to measure stress and anxiety. 06 Volume 5 Issue 4
The result In the single-block experiment (A), reject errors were more commonly made with high levels of capacity and time remaining as compared to the optimization model. In other words, subjects took a pass on bids that they otherwise should have accepted when they had plenty of time and capacity remaining. Conversely, accept errors were more commonly made with low amounts of time remaining and resources yet to allocate. In this case, subjects settled for lower offers than they needed to accept. The next step was to assess the extent to which motivational levels drove these behaviors. As predicted, pupil-dilation levels (indicative of arousal levels) rose as capacity decreased and as the number of blocks increased. In contrast, blink rates (measures of stress and discomfort) increased with higher levels of remaining capacity, combined with lower levels of time left. This shows that the low levels of arousal at the early stages of the experiment (where high levels of time and capacity remain) are associated with reject errors, and higher levels of stress in the latter stages (where time is limited and capacity remains) contribute to the accept errors. In experiments with multiple concurrent blocks to manage (B and C), reject errors appeared much less common, or at least smaller in magnitude on average. Pupil dilation rates suggest that the greater complexity in managing multiple blocks appears to mitigate the lower levels of engagement. In contrast, accept errors appear much more common in the more complex experiments (B and C), when compared with those observed in experiment A. Thus, in the comparatively more complex context, subjects were less likely to reject offers they should accept but more likely to accept suboptimal offers they should reject. The impact of time remaining and capacity appear to have the greatest impact on blink rate in experiment C, and, conversely, to be significantly reduced in experiment A. In experiments B and C, blink rate seems consistently on the rise, with decreasing time and increasing capacity levels. Interestingly, the point of this inversion seems to map to the point at which individuals appear to shift from a tendency to commit reject errors to a tendency toward accept errors. The findings of this study shed important light on the biases that people may have in making RM decisions. With too little challenge and plenty of time, people are more likely to reject offers they otherwise should accept. Conversely, overwhelmed by limited time and capacity yet to allocate, people are more likely to accept offers they otherwise should reject. This effect is 07
Reducing human error in revenue management decision-making Knowing the types of errors that take place in RM decisions, and why they occur, helps to build actionable tools for preventing and overcoming such biases. magnified in conditions of complexity. In especially complex RM decisions, people are less likely to be bored by any decision, thus less reject errors take place. The concern in complex scenarios is higher rates of accept errors, where the combination of complexity and time scarcity with resources yet to allocate adds significant levels of stress. Knowing the types of errors that take place in RM decisions, and why they occur, helps to build actionable tools for preventing and overcoming such biases. Implications for practice As a starting point, managers should be aware of the conditions under which RM decision-makers will make suboptimal decisions, reject or accept errors specifically. By doing so, they can begin to monitor for the root causes of such behaviors and reduce their frequency. The onset of stress due to excessively challenging work (to the detriment of performance) may be dealt with simply through creating mechanisms to bolster self-confidence commensurate with difficulty level. A broad approach to achieving this state is suggested in the form of increased training and resource availability to bolster awareness Recommendations of these important dynamics. If successful, these Business culture programs can reduce the negative effects check of stress, and consequently reduce costly Cost optimization accept errors. check Aside from these general tactics, Growth strategy check however, the present work suggests that certain more nuanced approaches to managing workload might be applied. For example, if the amount of capacity to be filled in a limited amount of time appears stressful, one solution may be simply to artificially lower the apparent amount of capacity that needs filling. Such an adjustment might involve a reallocation of some of that capacity task to another revenue manager or an individual crosstrained sufficiently to deal with it, or even to a more automated, if imperfect, artificial intelligence (AI) mechanism. Alternately, portions of capacity might be held in buffer, beyond the purview of RMs, until other capacity units are allocated, and the threat Figure 2. Motivation and performance: inverted-u relationship between challenge and performance outcomes Pupil dilation Indifference Source: Authors own. Arousal Level of challenge Performance Arousal Cognitive limit? Few errors Blink rate Errors of some kind are made Level of challenge Visual engagement Stress Level of challenge Negative stress effects Data validit 08 Volume 5 Issue 4
of stress-based complications in judgment are mitigated. By selectively reducing capacity, RM decision-makers will feel less stress and thus have a reduced tendency to make accept errors. Having access to accurate measures of work arousal (through pupil dilation measures as in this study) and stress (through blink frequency) could deliver a huge impact in real-time adjustments to work conditions. On first thought, it is difficult to imagine a typical workplace adopting the type of software necessary for such measurements today. However, it is conceivable that at some point certain workers themselves may be willing to have stress levels monitored with an interest in allowing management to take action to reduce it. If acceptable, such statemanipulations might be applied more effectively. Based on the views of revenue managers participating in a related followup study, this manipulation tactic appears to be a tenable prospect. Finally, we can use the above tools to help pinpoint thresholds where stress levels affect the quality of RM decisions across specific industries. The challenges of RM decisions within the airline industry, for example, may be distinct to those within the bed mattress industry. Examining the role of psychological motivation in RM decisions within industries can help solidify measures specific to that industry. From there, interventions can be designed specific to employee needs within industries and even within unique firms. These efforts can help maximize human decision-making in an RM context and improve working conditions at the same time, a win for companies and their employees. These efforts can help maximize human decision-making in an RM context and improve working conditions at the same time, a win for companies and their employees. 09