Presenter: Yifan Feng, Assistant Professor, NUS Bussiness School

Discussant: Kimon Drakopoulos, Associate Professor of Data
Sciences and Operations, USC Marshall School of Business

Learning to Rank under Strategic “Brush Wars”

Abstract:
We consider a dynamic learning and ranking problem of a digital
platform. Uninformed of the products’ intrinsic qualities, the platform
strives to design a sequential ranking policy that learns from historical
traffic data while accounting for potential manipulation by sellers to
inflate their performances, which we refer to as “brushing.” Are there
effective yet simple ranking algorithms to combat manipulative ranking
brushing?
We provide a positive answer by proposing a simple ranking algorithm,
termed Experiment-Then-Commit (ETC). We study the sellers’
strategic responses to the ranking algorithm by formulating an
N-player-T-period “brush war” game. We first show that the dynamic
game admits a static reduction through a dominant-strategy argument.
Then for every fixed N, we characterize the game’s asymptotic
behavior when T is large. We show the nonexistence of pure strategy
equilibria, shedding light on the possibility of efficiency loss. However,
for a large market where N is infinite, a different pattern emerges. We
formulate a novel non-atomic game with a continuum of sellers and
characterize a “self-reinforcing” market equilibrium. Under this
equilibrium, the seller’s brushing amount increases in the product’s
quality. In other words, the sellers’ strategic responses “reinforce”
complete learning of the platform. As a result, ETC can be highly
effective even under sellers’ manipulative ranking brushing. We also
discuss the managerial implications.

Presenter: Yvonne Huijun Zhu, PhD Student, NUS IORA

On the Value of Flexibility in Adaptive Experiments

Abstract:
Adaptive experimentation refers to the practice of changing (improving)
the experiment configurations on the fly based on observed data.
While it can potentially increase sample efficiency, it potentially needs
strong flexibility as an experimenter and thus may be difficult to
implement in practice.
To explore the interplay between sample efficiency and flexibility in
adaptive experiments, we formulate a Bayesian sequential hypothesis
testing problem. The goal is to minimize the sum of sampling and
penalty costs under different levels of flexibility. Here, flexibility is
defined as the number of times one can change the experiments. We
show that there is a notable gap between “No Flexibility” (where
experiments cannot be changed) and “Full Flexibility” (where
experiments can be freely changed adaptively). Nevertheless, the gap
can be (asymptotically) closed by just allowing to change the
experiment once. That is achieved by an experimentation rule we
develop, termed Nested SPRT, which is both simple to derive in closed
form and easy to implement.

Presenter: Zhi Chen, Assistant Professor, NUS Business School

Discussant: Pascale Crama, Professor of Operations Management,
SMU Lee Kong Chian School of Business

The Value of Private Feedback in Trial-and-error Innovation
Contests

Abstract:
Firms have increasingly turned to innovation contests as a means of
procuring complex industrial innovations from their supplier base. At
the start of the contest, the buyer announces the innovation challenge
and the award in the form of a valuable supply contract. To address the
challenge, it is common that suppliers adopt the trial-and-error
approach to develop solutions, and then submit them to the buyer for
evaluation as the contest progresses. In some contests, the buyer
privately reveals the interim performance of developed solutions to the
suppliers during the trial-and-error process (private feedback), but in
others, the buyer withholds such information (no feedback). Motivated
by such divergent practices, we seek to understand whether providing
private feedback (or not) results in higher profits for the buyer. We find
that when there is no urgent need for faster time-to-market, providing
private feedback is more profitable for the buyer when either there are
many evaluation rounds or when the cost of a trial is high. We further
uncover two sources of the value of private feedback: it incentivizes
suppliers to conduct more trials (“quantity effect”) in a more efficient
way (“quality effect”). In contrast, no feedback results in higher profits
for the buyer when there are few evaluation rounds and when the cost
of a trial is low. On the contrary, when the faster time-to-market is a
major concern for the buyer, we show that private feedback becomes
less attractive, and the buyer prefers suppliers to conduct parallel trials
(no feedback) particularly when the cost of a trial is very low or high, or
when the competition is intense with a large number of suppliers. Our
results have direct managerial implications and help shed light on
various feedback policies used in practice.

Presenter: Zhaoxuan Wei, PhD Student, NUS IORA

Partial Backorder Inventory System: Asymptotic Optimality and
Demand Learning

Abstract:
We develop a unified stochastic inventory model that not only captures
the impatience feature for unmet demands but also integrates the
classic backlogging and lost-sales inventory models. For such model,
when both demand and patience distributions are known, we establish
the uniform (asymptotic) optimality of the base-stock policy. While the
backlogged demands become unobservable (demand is partially
observable and patience is unobservable), we introduce a novel policy
family that operates without backlogged demands information, and
prove this proposed policy can approach the cost efficiency of the
optimal policy in the system when the demand and patience
distributions are known. Further we extend our analysis to an online
inventory control problem in which precise data on demand and
patience distributions are unobservable and only sales are observable
by developing a UCB-type algorithm that yields a near-optimal policy.
The regret bounds given by the algorithm are provably tight within the
planning horizon, and are comparable to the state-of-the-art results in
the literature, even in the face of partial and biased observations and
weaker system ergodicity.

Presenter: Long Zhao, Assistant Professor, NUS Business School

Discussant: Ilia Tsetlin, Professor of Decision Sciences, INSEAD

Predicting Tail Quantiles Through Aggregation of Medians: Model
and Analysis

Abstract:
Quantile forecasts are essential inputs for decision making under
uncertainty. The most useful quantiles are from the tails since tails
provide rich information about the underlying uncertainty. However,
directly predicting tail quantiles is challenging for human experts and
even quantitative models. On the other hand, non-tail quantiles such
as medians are easier to predict directly but are less useful in
uncertainty assessment. Motivated by this mismatch, we study how a
decision maker can predict tail quantiles through aggregation of
non-tail quantile forecasts (such as medians). Intuitively, we want to
shift the median forecasts by the true difference between median and
tail quantiles. However, this task is challenging because the underlying
randomness is unknown, and the median forecasts are subject to
biases and noises. Our method involves aggregation of median
forecasts to achieve a substantial reduction of noise and leveraging
past data to correct the biases and estimate the shift. We evaluate our
proposed method using the M5 uncertainty competition submissions
and find that our method outperforms established benchmarks in the
literature. We also offer a theoretical understanding of why our method
performs well empirically.

Presenter: Lan Wu, PhD Student, NUS Business School

Mitigating the Spiral Down Effect: Online Learning under
Mixed-fare Structure

Abstract:
In the dynamic landscape of airline revenue management, accurately
predicting customer behavior and adjusting ticket-selling strategies in
the face of fluctuating demand is critical. Traditional models frequently
overlook intricate customer buy-down patterns, resulting in a
detrimental downward spiral of revenue decline. This study combines
the adjusted fare concept with the Bayes Selector algorithm to create
an innovative approach for optimizing airline ticket-selling strategies
under mixed fare structure. By dynamically adjusting fare structures
based on purchase probabilities and expected profits, and utilizing
probabilistic estimates and learning capabilities of the Bayes Selector
algorithm, this integrated methodology enables airlines to adaptively
refine their seat allocation strategy for maximum profitability.
Our findings show that a constant regret is attainable for the problem
under the (independent) differentiated product demand setting. More
interestingly, we establish a logarithmic regret bound in the case of
mixed fare structure. To validate these results, we provide a simple
analysis of how estimation errors in the probabilities of the types (of
customers) is managed in the online seat allocation algorithm.
Furthermore, we illustrate the practical application of our approach
through a real-world airline scenario. This study concludes that by
learning and adapting to the nuances of customer behavior, airlines
can significantly enhance their revenue management capabilities,
leading to more accurate demand estimation and seat allocation
strategy in alignment with the existing environment.