🔬 physics

An instance-based learning approach for evaluating the perception of ride-hailing waiting time variability

This study employs an instance-based learning approach on stated preference survey data to reveal that ride-hailing users' perception of unexpected waiting time is 2–3 times their value of time, that recent experiences heavily outweigh fading memories in decision-making, and that service cancellations significantly drive provider switching.

Original authors: Nejc Geržinič, Oded Cats, Niels van Oort, Sascha Hoogendoorn-Lanser, Michel Bierlaire, Serge Hoogendoorn

Published 2026-03-20

📖 5 min read🧠 Deep dive

CC BY 4.0

Original authors: Nejc Geržinič, Oded Cats, Niels van Oort, Sascha Hoogendoorn-Lanser, Michel Bierlaire, Serge Hoogendoorn

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

Imagine you are waiting for a ride-share car (like Uber or Lyft) on a rainy evening. The app tells you the driver will arrive in 10 minutes. You step outside, and the car arrives in 5 minutes. You feel great! But what if the car arrives in 20 minutes? You feel annoyed. What if the app says "Driver found," but then cancels the ride entirely? You feel furious.

This paper is a deep dive into how our brains remember these waiting experiences and how those memories change the next time we open the app. The researchers didn't just ask people what they think they would do; they made them play a "ride-share simulator" game 32 times to see how they actually learned and adapted.

Here is the breakdown of their findings using simple analogies:

1. The "Freshness" of Memory (The Ice Cream Analogy)

The researchers wanted to know: Do we remember every bad ride we've ever taken, or do we only remember the last few?

They tested a theory called Instance-Based Learning. Think of your memory like a bowl of ice cream.

The Top Scoop: The most recent ride you took is the scoop right on top. It is fresh, cold, and very distinct. It has the biggest impact on your decision to order again.
The Melting Scoops: As you go deeper into the bowl (older rides), the ice cream melts and loses its shape. The second-to-last ride matters, but only about one-third as much as the last one. By the time you get to the 7th ride ago, it's basically a puddle of water.

The Takeaway: We are very "short-term" thinkers when it comes to service reliability. If a company messes up today, you are likely to switch. But if they mess up last week, you've probably already forgotten it, especially if they've been good since then.

2. The "Early Bird" Surprise vs. The "Late" Penalty

Usually, we think being late is worse than being early is good. But this study found something surprising: We love being early even more than we hate being late.

The Analogy: Imagine you are promised a gift at 5:00 PM.
- If it arrives at 4:30 PM, you feel a rush of excitement (a "bonus").
- If it arrives at 5:30 PM, you feel annoyed (a "penalty").
- The study found that the "happiness" of getting it early is worth 1.5 times more than the "anger" of it being late.

The Takeaway: Ride-hailing companies might be better off promising a slightly longer wait time (e.g., "15 mins") and then arriving early (e.g., "10 mins"). This creates a "positive surprise" that makes customers happier than if they promised "10 mins" and arrived exactly on time.

3. The "Ghost Ride" (Cancellations)

What happens when you book a ride, and the driver cancels?

The Analogy: This is like ordering a pizza, waiting for the delivery guy to arrive, and then having the restaurant call and say, "We lost your order, sorry."
The Finding: This causes a massive spike in anger. It's not just a small annoyance; it feels like a huge betrayal. The study found that a cancellation is so annoying that a company would need to give you a significant discount on your next ride just to get you to try them again.

4. The "Habit Loop" (Inertia)

Why do we keep using the same app even if it's not perfect? Because of Habit.

The Analogy: Imagine you always buy coffee at the same shop on your way to work. Even if the shop next door has slightly better coffee, you keep walking past it. Why? Because your brain is lazy. It doesn't want to do the mental math of comparing prices and wait times every single morning.
The Finding: The more times you pick the same company in a row, the "heavier" your habit becomes. It's like a snowball rolling down a hill; it gets bigger and harder to stop.
The Twist: The moment you switch to a different company, that snowball melts instantly. Your habit resets to zero. So, if you have a bad experience, you are very likely to switch, but once you switch, you have to build a new habit from scratch.

5. The "First Date" Barrier

Trying a brand new ride-hailing app feels risky.

The Analogy: It's like going on a first date with a stranger. You are nervous. You don't know if they are nice, if they will show up, or if they will be weird.
The Finding: People need a "discount" (about €2.00 or $2.00) just to overcome that nervousness and try a new service for the first time. Once they try it and it goes well, that barrier disappears.

Summary for the Real World

If you are a Ride-Hailing Company:

Under-promise and over-deliver: Tell people it will take 15 minutes, and try to get there in 10. The "early arrival" bonus is powerful.
Fix mistakes fast: If you cancel a ride, offer a big discount immediately. The bad memory is fresh and heavy, but it fades quickly if you fix it.
Don't let them switch: Once a customer is in your "habit loop," they are loyal. But if they have one bad day, they might switch, and you have to win them back from zero.

If you are a Passenger:

Your brain is playing tricks on you. You are judging your next ride based almost entirely on your last ride, forgetting the dozens of good ones you had before that.
You are actually happier when a service surprises you with speed than you are upset when it's slightly late.

This study essentially maps out the "emotional GPS" of a passenger, showing that our decisions aren't just about math (time and money); they are about how fresh our memories are and how much mental energy we want to spend making a choice.

1. Problem Statement

The rise of ride-hailing services (e.g., Uber, Lyft, DiDi) has introduced significant uncertainty regarding service reliability, particularly in waiting times and trip cancellations. While previous studies have analyzed mode choice, there is a lack of understanding regarding how users perceive and react to service variability over time. Specifically, existing literature often relies on explicit reliability information (stated upfront) rather than experiential learning, where users learn reliability through repeated trial-and-error.

The study addresses two main gaps:

Valuation of Variability: How do travelers value unexpected waiting times (early vs. late) and service cancellations in a ride-hailing context?
Learning Mechanisms: How do memory decay (forgetting past experiences) and inertia (habit) influence future choices? The authors aim to test novel theoretical formulations of these mechanisms using an Instance-Based Learning (IBL) framework.

2. Methodology

A. Survey Design (Stated Preference Experiment)

Data Source: A Stated Preference (SP) survey administered to 1,304 members of the Dutch Mobility Panel (MPN) in August 2021. 936 valid responses were retained after filtering for completion time (3–30 minutes).
Experimental Structure:
- Respondents faced 32 repeated choice tasks between two anonymous ride-hailing companies (Company A and Company B).
- Attributes: Expected waiting time, travel cost, and travel time (fixed at 20 mins for both).
- Variability: Waiting times were drawn from log-normal distributions. Company A was cheaper but more variable; Company B was more expensive but more reliable.
- Feedback Loop: After each choice, respondents received feedback on the actual experienced waiting time and whether the trip was cancelled.
- Cancellation: Company A had a 10% cancellation probability; Company B had 2%.
- Context: The scenario was set as returning home from a leisure activity in the evening.

B. Model Estimation

Model Type: Panel Mixed Logit (ML) model was used to handle the panel structure (repeated choices per individual) and account for unobserved heterogeneity.
Utility Function: The systematic utility ( $U$ $U$ ) was decomposed into three parts:
1. Descriptive Information: Static attributes (Cost, Estimated Waiting Time).
2. Experiential Information: Dynamic attributes based on past feedback (Actual vs. Expected waiting time difference, Cancellation status).
3. Error Term.
Instance-Based Learning (IBL) Formulations:
- Memory Decay: The authors tested four combinations of decay functions:
  - Time-based vs. Event-based (ordering experiences by time elapsed vs. sequence of occurrence).
  - Relative weights (averaging weights so they sum to 1) vs. Absolute weights (summing weights, allowing cumulative disutility to grow).
  - Best Fit: An Event-based formulation with Absolute weights performed best, suggesting that the most recent experience dominates, but the cumulative weight of multiple negative experiences matters more than a single one.
- Inertia (Habit): Five specifications were tested to capture the tendency to stick with a chosen option:
  - "One" (binary, last choice only).
  - "Sum" (cumulative count of choices).
  - "Reset" (cumulative count, but resets to 0 upon switching).
  - Logarithmic versions of "Sum" and "Reset".
  - Best Fit: The "Reset" specification performed best, indicating that habit builds with consecutive choices but is broken immediately upon switching.

3. Key Results

A. Value of Waiting Time and Cancellations

Unexpected Waiting Time: The value of unexpected waiting time is approximately 2–3 times the Value of Time (VoT).
- Early Pick-up: Travelers place a high reward on arriving earlier than expected (WTP: €0.44/min or 26.59 €/h).
- Late Pick-up: The penalty for waiting longer than expected is lower than the reward for early arrival (WTP: €0.30/min or 18.14 €/h).
- Note: This counter-intuitive result suggests early arrivals are viewed as a positive surprise rather than late arrivals being the sole "dissatisfier."
Service Cancellation: A cancelled trip results in a significant disutility, equivalent to a price increase of €4.45. This represents a strong motivation to switch providers.
Barrier-to-Entry: Trying a new service for the first time incurs a perceived penalty of approximately €2.00, suggesting a discount is needed to entice new users.

B. Memory Decay and Inertia

Memory Decay: The decay parameter ( $d$ $d$ ) was estimated at ~1.67.
- The most recent experience is by far the most influential.
- The second most recent experience has only ~33% of the weight of the latest.
- By the 7th instance, the weight drops below 0.05.
- Implication: Bad experiences fade quickly if followed by good ones, but a single recent bad experience has a massive immediate impact.
Inertia: The "Reset" model confirmed that inertia accumulates with consecutive choices but resets to zero immediately after a switch. Each additional consecutive choice adds a marginal value of €0.09 to the utility of staying with the same provider.

4. Key Contributions

Novel IBL Application: This is one of the first studies to apply Instance-Based Learning theory to ride-hailing, specifically testing Event-based vs. Time-based memory decay and Absolute vs. Relative weighting.
Asymmetric Valuation: The study reveals that travelers value early pick-ups more highly than they penalize late pick-ups, challenging the traditional "dissatisfier" theory where only negative deviations matter.
Quantification of Cancellation: It provides the first quantitative estimate of the disutility of ride-hailing cancellations (€4.45), highlighting the severity of this failure mode.
Habit Dynamics: It refines the understanding of inertia in transport choice, showing that habit is fragile and resets instantly upon switching, rather than persisting as a long-term bias.

5. Significance and Implications

For Operators:
- Reliability Management: Since the most recent experience dominates decision-making, operators must address negative experiences (delays/cancellations) immediately (e.g., via compensation) to prevent churn. However, because memories fade quickly, a single bad experience does not doom a provider if subsequent rides are good.
- Communication Strategy: The high value placed on early pick-ups suggests providers might benefit from under-promising (setting conservative estimates) to consistently exceed expectations, though this must be balanced against the risk of users arriving late to the pickup point.
- Customer Acquisition: New entrants face a €2.00 "barrier-to-entry." Significant initial discounts are necessary to overcome this hesitation.
For Policymakers:
- Understanding the high sensitivity to cancellations is crucial for regulating on-demand services, as these events drive users away from specific platforms more aggressively than minor delays.
- The findings support the integration of ride-hailing with public transport, provided reliability is managed, as users are willing to switch based on recent performance rather than long-term brand loyalty.

In summary, the paper demonstrates that ride-hailing user behavior is driven by a short-term memory of the most recent experience, an asymmetric valuation of time deviations, and a fragile habit that resets upon switching.