Behind the scenes
How does it work?
DLPTime is live wait times and AI forecasts for Disneyland Paris. Here's, no fluff, how it's built.
Hi, I'm Benjamin
I'm the publisher of DLPTime. I originally built this site for my own days at Disneyland Paris: I was tired of jumping between five apps that all show the same queue without telling me whether it's about to rise or drop.
DLPTime tries to answer one simple question: "is it the right moment to go, or should I wait?" Here's, honestly, how I do it.
Where the data comes from
All sources are public or open. No personal data is used to train the models.
Wait times
collected in real time from the public queue-times.com API, which aggregates the official Disneyland Paris feeds.
Weather
via the free Open-Meteo API (temperature, precipitation, cloud cover, wind) at the park's location.
School holidays
French, British and Belgian school holidays (the three main visitor markets), kept in sync with official sources.
Premier Access
the dynamic price of Premier Access is a precious signal: Disney adjusts it based on expected crowds, which makes it a leading indicator.
Special events
Halloween and Christmas seasons, park anniversaries, new openings, known refurbishments.
Wait times are stored over time. The database currently holds several months of history per attraction, which is what lets the models learn recurring patterns (days, hours, seasons).
Why machine learning?
An attraction doesn't follow a simple rule like "calm in the morning, packed in the afternoon". Patterns depend on the day of the week, holidays, weather, novelty, and the attraction itself. Way too many variables for hand-written rules.
Machine learning shines on exactly this kind of problem: it learns by itself which signals matter and how much. My job is to feed it the right data and watch that it doesn't go off the rails.
An attraction doesn't follow a simple rule. Way too many variables for hand-written rules, and exactly where ML shines.
The signals taken into account
Each attraction has its own behaviour. The models look at several families of signals in parallel:
Recent queue history
01the latest readings, their trend and their volatility.
Time
02hour of day (cyclically encoded), day of week, weekend or not.
Park load
03the average wait across the whole park right now, to place each attraction in its context.
Local weather
04temperature, rain, clouds, wind. Rain changes everything, especially for outdoor rides.
School holidays
05France, UK, Belgium. A school day and a Tuesday during half-term aren't the same park.
Premier Access
06the current PA price reflects the crowd Disney is anticipating. A precious indicator.
Events & seasons
07Halloween, Christmas, anniversaries, new attractions. Every season has its fingerprint.
The trick is less "which signals" than "how they interact and with what weight per attraction". That part stays in-house.
Short term: 15 to 60 minutes
For short-term predictions (15 min, 30 min, 1 h) I use gradient boosting (LightGBM), trained per attraction and per horizon. Fast, robust, and readable when a case goes wrong.
The day-ahead profile
For the day-ahead profile, I blend a statistical model on historical patterns with TimesFM 1.3 (Google's time-series foundation model) running in shadow mode for comparison. Selection is driven by daily quality measurement (MAE).
Limits, because honesty matters
Forecasts are indicative. An attraction can break down, a soft opening can be announced without notice, a day can be atypical for a thousand reasons.
The further the horizon, the less precise. The best predictions are at 15-30 minutes; the day-ahead profile is more of a silhouette than a precise script.
Recent attractions (less than 6 months old) have little history. The models are still learning.
I don't predict the human experience, only the wait time. A packed day can be wonderful, a quiet day can be rainy.
An indie project, no ads
DLPTime is a personal project, with no affiliation to The Walt Disney Company. No ads, no commercial tracking beyond anonymous audience measurement. The code runs on an OVH VPS, the models retrain every night.
How to support the project
If you find the site useful, the best way to support it is to share it. And to tell me what works or doesn't, I read everything.
GO DEEPER
The attraction-by-attraction guide
For every ride: best window, peak hour and reliability, refreshed every night.