Inside the model

Lapbar isn't a list of wait times with a plan stuck on top. It's a planning model we built from scratch, fed by our own data pipeline that watches nearly 200 parks around the clock, and it marks its own homework every night. Right now we're building the global model: one model that learns from every park, every day, and from years of history. The counts and accuracy record below update themselves from our database.

The headline number, live

17quintillion possible vectors for the model to draw on

A vector is one data point read through one combination of factors. We hold 8.1 million data points going back to 2015, and each can be read through any combination of the 41 factors we log or test, from weather and holidays to exchange rates: 2.1 trillion combinations. Every new data point adds trillions more vectors, and every new factor doubles the lot. This is the space the model searches when it decides which factors earn their place. It's a theoretical maximum: a factor is only used if it improves forecasts on days held back for checking.

20.7 milliondata points
70.3 trillionways to combine 46 factors
1.4 sextillionpossible vectors
199
parks logged around the clock, around the world
5min
between snapshots of every ride's posted wait
1min
between checks of whether each ride is up or down
6
forecast horizons, from now to four hours ahead, scored every 30 minutes
100
test visitors run through yesterday's park day again, every night
0
look-ahead: a replayed plan only sees what was known at the time
Live counts

Everything logged so far

Straight from our database and refreshed every 10 minutes, so these only ever go up.

670k
wait-time readings logged
2,500
breakdowns, closures and reopenings recorded
82k
weather forecasts and readings logged
8,900
wait forecasts checked against what really happened

What we collect

Most planners read today's wait times and stop there. Lapbar keeps everything, because a forecast is only as good as the history behind it. Our servers log all of this for every park we track, and keep the raw responses too, so the model can always be rebuilt from what was actually seen.

The data, drawn

How deep the data goes

Live from our database: how much history is loaded, what we're testing, how honest our chances are, and the two decades of background the global model learns from. All the charts.

History loaded so far

Loading live data…

What moves the queues

Loading live data…

Does "90% likely" mean 90%?

Lapbar gives each ride a chance it'll be running when you get there. Here's how those chances held up.

Loading live data…

The shape of the year

Loading live data…

Orlando since 2003: flights, hotels and tickets

Loading live data…

Work in progress

We're building the global model

Lapbar is training one model across every park it logs, so what it learns in one park helps all the others. It learns from two kinds of data at once: what's happening in the parks today, and years of history about how busy parks get and why.

Why build the most complete model we can?

Because a queue is never just a queue. How long you'll wait at 2pm depends on far more than the number on the sign at 10am: whether it's a school holiday, and not only where the park is but where its visitors fly in from; whether a storm is coming; whether a big ride nearby has just broken down; whether flights into town are full this month, or money is tight and families are staying home.

Every one of those things changes how busy a park gets. The more of them Lapbar understands, the better it can tell you when to ride, what to skip and where to go next. Better predictions mean shorter queues, less walking back and forth, and more of the rides you came for. That's the whole point.

Today, live

Every ride's wait and whether it's running, the weather, opening hours, shows and events, logged around the clock for nearly 200 parks. This is what today's plans use, and the counts above grow every few minutes.

Years of history

Over a year of past wait times for the biggest parks is loading now, so each park's model starts with seasons it hasn't lived through yet. Beneath that sit two decades of travel, tourism and attendance trends.

Tested before it's trusted

One past day in five is locked away and never used for tuning. New information only changes your plan if it makes the forecasts better on those locked days. Anything that doesn't help stays switched off.

What goes into it, and what it's doing today

"In plans" means today's plans use it. "Learning" means the model is learning from it now. "In testing" means it has to prove it makes forecasts better before it can change one.

  • In plans
    Live queue times. Every ride's posted wait every 5 minutes, and whether it's running, at nearly 200 parks around the world.
  • In plans
    The wait you'll really have. Posted times usually run long, so Lapbar corrects each one to how long people actually queue, and keeps learning from the queue times visitors choose to share. The most precise of those will come from the Live Activity's "I'm in the queue" and "I'm on the ride" taps, which time a queue to the minute. Coming soon
  • In testing
    Real queues and walks from visitors. With location sharing on, visitors confirm on their Lock Screen when they join a queue and come off a ride, anywhere in the park, so real waits arrive for every ride people visit, not just the ones they remember to tap. Their walks between rides teach real walking speeds: once a park has 300 shared walks over 20 days, its plans use the speed people really walk there (within 15% of our own estimate, since plans already slow walks for crowds). Nothing that identifies anyone is shared. Collecting starts once the app update with it is out. Coming soon
  • In plans
    Weather. Forecasts as they were made and what actually happened: rain, heat, wind, thunder and severe weather warnings.
  • In plans
    The park's own calendar. Opening hours, early entry, evening events, ticketed parties and showtimes, and every change to them.
  • In plans
    Holidays. Public and school holidays wherever there's a park, region by region.
  • In plans
    Height limits. Each park's published minimum height for its rides at almost every park Lapbar plans, so a plan for a family leaves out a ride a child is too short for, or has the grown-ups take turns with Rider Switch. Where we haven't got a ride's limit yet, the plan says to check the sign there.
  • In plans
    How long each ride takes. Not just the queue: the pre-shows, the ride itself and getting back out. Rise of the Resistance is about 20 minutes from the front of the queue and Flight of Passage about 13, so a plan never squeezes a second headliner into time that isn't there. Lengths are checked by hand from published ride times and rounded to the minute.
  • In plans
    New rides and closures, so a ride's busy opening months or a favourite closed for repairs don't throw the forecasts off.
  • Learning
    Past wait times. More than a year of history for over 1,200 rides at the biggest parks, loading now.
  • In testing
    Past storms. Years of airport weather reports and satellite lightning, matched minute by minute to past queue times: which rides close for lightning, how long after the last thunder they reopen, how queues move indoors, how many people leave afterwards, the rush when rides reopen, and the hours storms usually arrive in each month. Past storm days are also relived with and without live lightning warnings, to measure how much an early move is worth. The year-long replay of queue forecasts now also runs with and without the weather as it was known at each moment, rain and thunder at the airports and storm cells tracked from satellite lightning, so we can measure how much they sharpen forecasts one to four hours ahead and how well they predict which rides will close. In Orlando's summer that's mid-afternoon to early evening. Each one changes the storm plan only once it beats the current setting on the locked days.
  • In testing
    Storm closures and breakdowns, predicted. Two models, each learned from past days. One predicts which running rides lightning will shut in the next two hours, from live satellite lightning (how close and how recent), storms on their way, which way the upper winds are carrying the storm (from NOAA's HRRR forecast, in the US), the usual storm hours, how each ride has behaved in past storms, and how many outdoor rides have already closed. The other predicts breakdowns on top of each ride's own record: a ride that has just reopened, or has already broken down today, is likelier to stop again, and the first hour, the time of day, heat, rain, how busy it is and how much of the park is down are each tested too. Every input is tested on its own, on days the models never learned from, and storm hours and the summer 2026 Florida storms are scored separately. A model reaches plans only after it beats today's rule on those days, then on the year-long replay, the locked days and live park days.
  • In testing
    Skip-the-line return times. A year of Lightning Lane return times at Walt Disney World, from the ThemeParks.wiki history. When a ride's returns are booked further into the day than usual, its standby queue tends to grow over the next few hours, so the model is testing whether that sharpens forecasts one to four hours ahead. Prices for the day turned out to add little once weekdays, seasons and holidays are known. It changes forecasts only once it beats the current model on the locked days.
  • In testing
    How the crowd evens out queues. Do people follow the crowd into long queues? Not on average. Across 170 past days at our test parks, a ride whose queue was well over its usual length for that time came in shorter half an hour later than the forecast expected, and a short one longer, at every park: people see a long queue and go elsewhere, and short ones fill. The rides around a ride added nothing beyond that. The model is testing whether allowing for it sharpens forecasts 30 minutes to an hour ahead. It changes forecasts only once it beats the current model on the locked days.
  • In testing
    Past show and parade times. More than a year of published showtimes for the parades, fireworks and stage shows, loading now, so the model can learn ride by ride which queues empty out while each show is on and test that against the same model with shows switched off.
  • In testing
    Holiday weekends, day by day. Every day of every holiday gets its own effect, learned from past queues: the Wednesday before Thanksgiving, Memorial Day Saturday, the Monday after, Christmas week, UK half-terms, Golden Week in Japan and China, Chuseok in Korea and more, with the run-up and the drop-off. US holidays are checked against crowd calendars back to 2014. Spring break is also modelled state by state: which weeks Texas, Florida, Georgia, New York and the rest are off, mixed by where each park's visitors come from. Each country's holidays change plans only once they beat the current holiday rule on held-out holidays.
  • In testing
    Where visitors come from. Which countries and US states each resort's visitors travel from, month by month, measured from public arrival records and airline ticket samples back to 2005, and the school breaks and holidays there. A half-term in Britain or a school break in Brazil can fill Orlando too.
  • In testing
    Travel and tourism trends. How many people are flying in, how full the hotels are and how many visited each park in years gone by, going back as far as 2002.
  • In testing
    Economic signals. Exchange rates, petrol prices, jobs and how confident families feel about spending, which all shape who can afford a trip and when.
  • In testing
    What each ride is. Its type, who built it, how old it is, whether it's indoors and its height limit, from open ride databases. Rides of the same kind tend to break down in similar ways, so a ride with little history of its own can start from what similar rides do, in the same climate. It only reaches a forecast if it makes breakdown forecasts better on the locked days.
  • In testing
    A learned fine-tune for the next few hours. A small set of decision trees, trained only on past days, nudges each forecast from how that queue has moved over the last hour, how long its sign has been stuck and the time of day. It is still maths: it has no words in it, and the same moment always gives the same nudge. The nudge is capped, so the model's own forecast and its reasons still come first. On test days it cut the average miss by 3 to 6% half an hour to four hours ahead. It changes forecasts only once it wins on the locked days and in a live shadow run.
  • In testing
    Closure ripples. When one ride closes, how much longer every other queue in the park gets, ride by ride, and whether it matters why it closed: a breakdown, the weather, several rides at once, or a planned closure for a parade. It also tracks the second wave: how long a broken ride's queue is usually held before the park empties it, and where those guests go next. Each effect is measured against quiet moments when nothing closed, so only real ripples count, and the forecast only uses them if they prove themselves.
  • In testing
    Big events. Hurricanes, the pandemic closures, recessions, big ride openings and price changes going back to 2001, so the model can tell a one-off from a pattern.

Today's plans use the live layer and Lapbar's own per-ride models built from it. The rest joins the plans only once it proves itself on the locked days, and the accuracy record above will show it when it does. See the live counts.

Maths, not a chatbot

A proper scientific model

Lapbar's forecasts and plans come from a mathematical model we built ourselves: statistics fitted to millions of real recordings, the odds of each ride breaking down, measured walking times, and a search through thousands of possible days. It isn't ChatGPT, Claude or any other AI chatbot, and no chatbot sits between the data and your plan.

The scientific method, every day

  1. 1

    Measure

    Every few minutes we record every ride's wait and whether it's running at nearly 200 parks, plus the weather.

  2. 2

    Predict

    Every 15 minutes the model writes down what each wait will be, up to four hours ahead, before anyone knows the answer.

  3. 3

    Check

    Every 30 minutes those predictions are scored against what really happened. The scores are on this page, good or bad.

  4. 4

    Keep only what wins

    The model is a set of named parts (wait times, ride breakdowns, storms, crowd ripples and more), and a change touches one part at a time. It goes live only if it beats the current model on days it has never seen, and a change that makes plans worse is turned down. Any part can be put back on its own.

Where the answers come from

A chatbotWrites whatever sounds likely, from text it has read.

LapbarWorks it out from real recordings: queue times, breakdowns and weather, logged every few minutes.

Ask it twice

A chatbotCan give you a different answer each time.

LapbarSame park, same moment, same live data: exactly the same plan, every time.

Does it make things up?

A chatbotCan invent things that sound right but aren't.

LapbarCan't. Every ride, wait and walk comes from real park data and our map of the paths, and every forecast is checked against what really happened.

Same situation, same answer

Give Lapbar the same park, the same moment, the same live data and the same wish list, and it gives exactly the same plan, down to the minute. Run it a hundred times and you get the same answer a hundred times. Your plan only changes when something real changes: a queue, a ride, the weather, or where you are.

Our tests work the same way. The test visitors we play through past days make their choices from fixed, numbered starting points, so a rerun is identical move for move. That's how we compare one version of the model with the next fairly.

Every stop explains itself

Because it's maths, every recommendation can say why in plain words, worked out from the numbers behind it, like "In the morning: it often goes down in the rain, and rain is likely from about 14:00." If a call doesn't pan out, you can still see why it was made.

Where AI does appear: behind the scenes, an AI assistant helps our team suggest experiments to test overnight. It can only pick from a fixed list of settings, never sees anyone's plans or personal details, never forecasts a queue or makes a plan, and nothing it suggests reaches the app unless it passes the same tests as every other change.

The model, built in-house

The planning engine is our own work, written from the ground up for theme parks. It runs on our servers, and these are the parts that make up every plan.

Queue forecasts

Each ride's wait for every minute of the day, built from the day's crowd level, the ride's usual shape through the day and learned hour-by-hour corrections. A live "nowcast" blends in the latest readings, every forecast comes with a range, and posted times are corrected to how long people really wait.

Breakdown maths

Every ride has its own track record: how often it stops, how long repairs usually take and how queues surge when it reopens. When a big ride goes down, the model knows which nearby queues will swell, learned from past breakdowns. Every night it also compares rides across every park by type, maker, age, climate and weather, to see which kinds of ride stop most, and when. The results go in the app's Stats tab. Coming soon When one ride's own record shows it stops far more in the rain or at one part of the day, the day plan moves it to when it's more likely to be running and says why on that stop. Coming soon

Weather that matters

Lightning risk closes outdoor rides, rain and heat change who's queuing where, and every plan carries a storm version ready to go. Lapbar measures how each ride's queue moves in rain and heat, rather than guessing: on a hot afternoon, outdoor queues thin out most and water rides hold up.

A real walking map

Magic Kingdom alone is 3,523 points and 4,088 path segments, with every ride's entrance and exit mapped separately. Main Street and the hub slow down as crowds build, parade routes are blocked while a parade runs, and the railroad counts as a way to get around when it's quicker. Next, it learns real walking times from walks visitors choose to share anonymously, to see how much crowds really slow each path. Coming soon

A proper optimiser

The planner simulates your whole day minute by minute, then searches thousands of possible orders in about a second (a beam search, then simulated annealing, then a final polish) for the one that gets you the most of what you ranked highest. It re-optimises from wherever you are whenever the park changes, without flip-flopping on noise.

The rope drop window

For each park, how long queues stay short after you get in, measured from our own five-minute queue records (with Early Entry measured separately), and which rides often open late. Every day plan uses each ride's own opening record: a ride that usually opens late is planned for after its usual start, and the stop says why. Coming soon Frenzy mode uses it to put your must-do whose queue builds fastest first, then short queues close by, so you don't cross the park for a short line. Coming soon The windows are early measurements from a few weeks of days, and they're refitted as the history grows.

Busyness patterns

Nightly, Lapbar works out how every ride behaves by weekday, month, heat, rain, holidays, ticketed parties and evenings. It even learns which rides empty out during the fireworks and parade, ride by ride, so the plan can send you there.

The learning loop

It checks its own homework. Every night.

Every 15 minutes the model writes down what it thinks each ride's wait will be in 15 minutes, half an hour, an hour, two hours and four hours. Every 30 minutes it compares those forecasts with what actually happened.

It also trains on each park's past year, the way it would have been used: at the start of every month a version is fitted only on the days before it and then scored on that month, so it never sees what happens next. It learns the crowd by month and weekday, each ride's usual shape through the day, and how long a live reading stays useful. Only if those versions win on the locked days does the fit go live, and the forecasts then keep being scored on new days.

Every day, and again the same day whenever a new experiment is added, it also replays the past 400 days at sixteen parks, one day at a time: the four Walt Disney World parks, the two at the Disneyland Resort, and ten in Asia and Europe (the Disney and Universal parks in Japan, China, Hong Kong and Singapore, both Disneyland Paris parks and Europa-Park). Each day is replayed as if it were live: refit on everything before the day, forecast it, score it against what happened and against simply assuming the wait stays the same, then step on to the next day. Every experiment says what it is for: better waits, better waits in the next half hour, earlier warning that a ride will close, or better plans. It has to win at that, without making anything else clearly worse, in three tests before it can change a plan: on that replay, on the locked days, then in a shadow run on live park days, where its forecasts are scored but never shown. With parks in Asia, Europe and the US in the tests, some park is open at every hour of the day, so live checks never stop. A park where it lost keeps its current settings. Every version of the model, and every experiment, gets its own two-word name, like Peculiar Pickle or Swooping Corkscrew, drawn at random by the system the moment it's made and kept for good. Every named model stays on our internal leaderboard for good, scored on each part of the tests above, so an old model that was the best at one thing can be found again. You can see which one is planning today on the live page.

The replay also plans real past days for a fixed panel of 96 test visitors, from young families and show lovers to coaster fans and foodies. Each visitor keeps the same settings in every test, so every version of the model is judged on the same people. Each one's day is also played without the app, by the same person with the same wish list, meals and hours. Like real people, they end the day in different ways: many stay for the fireworks, some head home once their list is done, and a fireworks fan spending several days at the park on a longer trip only stays for the show on some nights. Each visitor's habit is the same with and without the app. Without the app, while they're in the park they keep going, re-riding favourites and joining queues as they find them. Some follow a simple plan: their wish list in order, a loop round the park clockwise or anticlockwise, the back of the park first, the headliners first, or always the nearest ride. Others play it like enthusiasts, with a rope-drop plan from online advice or chasing the shortest posted wait. Days are graded from calm to disaster by how much ride time was lost to breakdowns and storms, so we can see how much Lapbar adds on a smooth day and how well it protects a bad one. An experiment that makes forecasts better but plans worse is turned down.

Changes to the planner itself go through a fixed test bench: the same frozen real days at the six Disney parks in Florida and California and the ten test parks in Asia and Europe, from calm to disaster, with all 96 visitors and their no-app days. Each change is scored the same way against the last accepted version and has to pass the same checks. Plans can't get worse overall or for any kind of visitor, day or park, and must-dos, meals and wasted trips are watched too. A second set of locked days, never used for tuning, has to confirm a win. The days where Lapbar still does worse than a typical visitor without the app are ranked by cost, and the biggest one is the next thing we fix. Each night yesterday's real day at every park is added as a fresh test. Real days also go wrong for human reasons, so each new version is tested a second way too: the same visitors meet the same seeded mishaps, like arriving late, an unplanned toilet or snack stop, a child refusing a ride at the door, going home early or a phone that dies for an hour. Lapbar and the same person without the app meet exactly the same ones, and we measure how much of Lapbar's head start survives them. The rates of those mishaps are our own estimates for now; this test reports, it doesn't decide.

Each night it re-learns from the last 28 days, tests the new version on yesterday (which it wasn't allowed to see) and only switches over if the new version is at least as accurate. Then 100 made-up visitors, from young families to coaster fans, replay yesterday with the plan Lapbar would have given them, and each day is scored against the best plan possible with perfect hindsight.

15min
between rounds of fresh forecasts for every ride
30min
between checks of forecasts against reality
28days
of history in each nightly refit
1rule
a new model only goes live if it beats the old one on unseen data
The accuracy record

How accurate it is, and whether it's getting better

Every forecast is scored against the wait the park went on to post. This is the running record for Magic Kingdom, with nothing left out, and it grows by a day every night.

Straight from the pipeline

Today at Magic Kingdom, as it happens

These charts draw themselves from our own database each time you load the page. We started logging in November 2026, so they show what's happening now rather than long-run averages.

Honest about what it doesn't do yet

Lapbar plans days at 36 parks in the beta, and logs every other park it tracks so their models start with real history.

Since November 2026 a Magic Kingdom plan can also include Lightning Lane, meals and breaks, and trips of several days. The app gets these in its next update.

  • Lightning Lane. Coming soon Plans use the return times you've booked. If you have a Multi Pass, Lapbar suggests what to book first and when to book the next one, and says whether a Single Pass ride is worth it. On the day, it shows the next return time the park is offering right now. Times later in the day, and for other days, are Lapbar's own estimates, learned from what it has seen at the park. Those estimates are still young, and they get better as history builds up. Universal Express and Merlin Fastrack come with those parks.
  • Meals and breaks. Coming soon Lunch and dinner are in every plan unless you turn them off, plus a snack or a sit-down break if you want them, each timed for short counter queues and near your next ride.
  • Several days. Coming soon For Orlando and Southern California, Lapbar picks the quietest of the dates you could go and puts each ride on the day its queue should be shortest. Days beyond the weather forecast assume typical weather until the forecast reaches them.
  • Disney hotel transport. Coming soon From your Walt Disney World hotel, each bus, boat, Skyliner and monorail trip is riding time plus a modelled queue to board, longer at opening and closing and on busy days, given as a range. Disney publishes no transport times, so these are estimates.

Credits

Live wait times are powered by ThemeParks.wiki and Queue-Times.com. Includes weather data from MET Norway. Historical weather from the Iowa Environmental Mesonet and lightning from NOAA's GOES satellites. Storm outlooks from NOAA's Storm Prediction Center and hurricane tracks from the National Hurricane Center. Radar and lightning probability from NOAA MRMS, storm forecasts from NOAA HRRR, wildfire smoke from the NOAA Hazard Mapping System, and storm chances outside the US from ECMWF open data (Copyright ECMWF, CC BY 4.0). Rail disruption messages from SNCF Voyageurs (ODbL). Exchange rates: source ECB. Our research also uses Wikipedia page views from Wikimedia and news data from The GDELT Project. Walking paths and ride details © OpenStreetMap contributors. Ride types and makers also from Wikidata. Height limits from each park's published rider rules, via DeepArrival. Every provider keeps ownership of its own data. The model, the learning, the calculations and the plans are proprietary to Lapbar, a trading name of Aethus Limited.

This product uses the FRED® API but is not endorsed or certified by the Federal Reserve Bank of St. Louis. This product uses the Census Bureau Data API but is not endorsed or certified by the Census Bureau. BLS.gov cannot vouch for the data or analyses derived from these data after the data have been retrieved from BLS.gov.

Lapbar is independent. It is not affiliated with, endorsed by or sponsored by Walt Disney World, Universal, Merlin or any other park operator. Park and ride names are used only to describe where Lapbar works. Your location and personal details are never part of the data: the learning uses park data and anonymous plans with location removed. The one exception is walks you choose to share, which are kept anonymously and matched to the park's paths. Coming soon Privacy policy.

Get Lapbar How it works