Abstract
This research investigates data-driven assistants that provide congestion forecasts to users of shared facilities, such as roads or cafeterias, to enhance coordination and system efficiency. We analyze the theoretical conditions for "self-fulfilling prophecies"—forecasts that remain accurate even after users react to them—and introduce Expodamp, a simple yet effective algorithm designed to achieve these stable states. Our real-world campus experiment demonstrates that Expodamp successfully reduces prediction errors by accounting for the feedback loop between forecasts and user behavior.
Introduction: The Feedback Loop Challenge
Digital assistants like Google Maps provide "popular times" forecasts to help users avoid crowds. However, these forecasts create a feedback loop: if an app predicts a facility will be empty, many users may decide to go at that time, causing the very congestion the forecast aimed to prevent. This phenomenon, often called "flapping," makes accurate prediction difficult and can lead to system instability.
The authors address two primary challenges:
- Theoretical: Determining the conditions under which a "self-fulfilling prophecy" exists as a stable Bayesian Nash Equilibrium.
- Algorithmic: Designing an assistant that learns to reach this equilibrium using only aggregate behavioral data, without knowing individual user preferences.
Theoretical Framework and Methods
The interaction is modeled as a Bayesian game between an assistant and a population of users. The assistant provides a public forecast $A$, and users choose their actions $B_i$ based on this forecast and their private signals $W_i$.
Using the Leray-Schauder-Tychonoff fixed-point theorem, the authors prove that self-fulfilling prophecies exist even in large-scale settings where the assistant only observes aggregate headcounts.
The Expodamp Algorithm
To find these equilibria in practice, the authors propose Expodamp. Standard forecasting methods often fail because they treat past data as independent of the forecast. Expodamp uses a modified version of exponential smoothing that "dampens" oscillations. By accounting for the feedback loop, it stabilizes the system toward a fixed-point forecast where the predicted congestion matches the actual outcome.
Experimental Evaluation
The researchers conducted a 70-day interventional study at a large campus cafeteria serving approximately 400 users daily. A web app provided daily forecasts to a subset of users. The study compared the Expodamp algorithm against a standard Average baseline, which predicts future congestion based on historical averages without accounting for the feedback effect.
Key Results
The experimental results demonstrated that accounting for the feedback loop significantly improves prediction accuracy. Expodamp achieved a lower Mean Squared Error (MSE) compared to the baseline method.
Beyond numerical accuracy, the paper establishes that achieving a perfect prediction is formally equivalent to solving the coordination problem. By providing a trusted signal, the assistant "outsources" the complex task of social reasoning, allowing users to reach a stable equilibrium without needing to model each other's behavior directly.
Discussion and Future Work
While the study assumes users respond optimally to forecasts, real-world behavior may be more complex. Future research could explore "within-day" updates and mechanisms to steer users toward socially optimal outcomes (maximizing social welfare) rather than just stable equilibria. Additionally, the role of population diversity remains a key factor; highly homogeneous populations may still face coordination challenges even with predictive assistance.
Summary updated