Mistral's new model predicted hotel room prices a year ahead with 88.6% accuracy

1Why this matters

Room price is the biggest lever a hotel has. Every night that passes unsold is revenue gone for good, and every room sold too cheap on a busy night is money left on the table.

• The decision repeats every day. A hotel prices every night of the coming year, and the right price for each night moves as bookings come in.
• It repeats across every hotel. A group with dozens of hotels makes this call for thousands of nights at once.
• People cannot look at all of it. So they look at what they can, and the rest waits.

The tools are good at collecting numbers. What they leave to people is the judgement: is this price right, for this hotel, on this night?

2The words, explained once

This section assumes you have never built a forecast. If you have, skip to section 3.

The hotel side

• Revenue manager. The person who sets room prices for one or several hotels.
• Room price for a night. What a guest pays for one standard room, one night. Here we use the average price actually paid for each arrival date.
• Revenue management software. A tool that recommends prices. People can follow its advice or override it.
• Competitor rates. The prices of nearby hotels for the same night, collected by a separate tool.

The forecast side

• Forecast. A prediction of future numbers from past ones, like a weather forecast for prices. Here, the price each night should carry, up to a year ahead.
• Accuracy. 100% minus the average gap between the forecast and the real price, as a share of the real price. A forecast that is off by 10% on average is 90% accurate.
• Range. A good forecast gives a best guess and a range: "probably 120 euros, and 8 times in 10 between 100 and 140". We call that the 80% range. Its width tells you how sure the forecast is.
• An honest range. The range is honest when about 8 real prices in 10 land inside it. Too narrow, and it is often wrong. Too wide, and it tells you nothing.
• Factor. Extra information you give the forecast: the day of the week, a public holiday, a trade fair in town.
• Known in advance. A factor you can know for a future date when you make the forecast, such as a holiday. Exchange rates or air traffic next year are not known in advance, so they cannot help a year ahead.

The models

• Timestral. Mistral AI's new forecasting model, released 15 Oct 2026. You give it a price history and it returns a forecast and a range. It works out of the box: no training on your data, free to download and run on your own computers.
• LightGBM. A popular machine learning tool from Microsoft. It has to learn from your own history before it can forecast.
• TimesFM and Chronos-2. Ready made forecasting models like Timestral, from Google and Amazon.
• ARIMA, ETS and Prophet. Classic statistical tools, the standard way to forecast for decades. Prophet comes from Meta.

3How a room is priced today

We worked with a large European hotel group on a tool that flags badly priced nights. Their own scoping work describes the job today.

Three sources, read side by side

• The booking system holds today's price and sales conditions.
• The revenue management software holds its own forecast and a recommended price.
• A rate shopper shows what competitors charge for the same night.

The group's data lead put the daily question plainly: is my price right for the 30th? And for the 31st? A revenue manager asks it for every hotel and every date.

The time it takes

• 10 hotels per person is a typical portfolio for one revenue manager.
• 2 hours a day, and often half of Monday, go to reviewing that portfolio to find the critical nights, according to the group's own scoping.
• The check against competitors is manual. It is done to make sure the price is not out of line, and it takes time.

The rules written by hand

To find problem nights faster, teams write rules. Each rule is an if then check, run for each hotel, each night and each standard room:

• If bookings in the last 7 days are weak and the forecast occupancy is low, flag the night as likely too expensive.
• If the price sits 15% or more under the competitors, flag it as likely too cheap.
• If the price is 20% or more away from the software's recommendation, ask the revenue manager to apply it or explain.

In the version we built, 24 rules were written down. Every threshold in them, 15%, 20%, 55% occupancy, is a judgement call. None was measured on the group's data. And a rule only fires on the case its author thought of.

4What we tested

The data

We searched the main data hubs and research papers for an open set with years of daily prices for many hotels, and found none. So we used the best two public sources we could find.

• Two hotels in Portugal, day by day. The Hotel Booking Demand dataset holds the bookings of a city hotel and a resort hotel. We took the average price paid for each arrival date: 793 days, from 1 Jul 2015 to 31 Aug 2017.
• 175 Spanish hotel markets, month by month. Spain's statistics office publishes the average hotel room price per region, province and tourist area. We used 175 of these series, some going back to 2008.

The task

• Forecast every night's price for the next 365 days. We also scored 90 days (3 months) and 180 days (6 months) ahead.
• Replay the past. We pretended to stand on 6 to 29 different past dates per hotel, forecast the following months using only what was known on that date, then compared with what really happened.
• At least a year of history. Below 365 days a tool has not seen a full season, so no forecast starts before that.

The fair play rules

• Only factors known in advance. Calendar, school and public holidays, events, sport fixtures, daylight hours and normal weather for the season. No exchange rates, no air traffic, no web search trends.
• No booked price. The average price of rooms already booked for a night reads the hotel's own price back to the model. We dropped it entirely.
• No peeking. When a tool had several possible setups, it picked one on older dates and was scored only on later ones.
• Nothing tuned by hand. Each tool ran as it comes out of the box.

5A year ahead

A year ahead is where a hotel plans its budget, signs group contracts and prices its big events. It is also the hardest forecast to get right.

• Timestral came first, at 88.6% accuracy. On average, its forecast was within 11.4% of the real nightly price, a full year out.
• LightGBM came second, at 84.6%.
• The two other ready made models, TimesFM from Google and Chronos-2 from Amazon, scored 79.5% and 79.0%.
• The classic statistical tools fell far behind. ARIMA scored 70.2%. ETS and Prophet were below 50%: a year out, their forecasts drifted away from real prices.

6Closer in: 3 and 6 months

Closer to the night, more is known and every tool improves. Here the result is a close race.

• 3 months ahead, LightGBM scored 91.3% and Timestral 90.6%. Chronos-2 was close behind at 89.7%.
• 6 months ahead, LightGBM scored 90.0% and Timestral 88.2%.
• A year ahead, the order flips: Timestral 88.6%, LightGBM 84.6%.

One likely reason: LightGBM learns from each hotel's own recent history, which pays off on the near months. A year out, that history says less, and its forecasts slip.

7A range you can trust

A single number is never exactly right. What a revenue manager needs is the range, and the confidence that the range means what it says.

We asked each tool for its 80% range, the band where it expects 8 real prices in 10 to land, and counted how often they did.

• Timestral's ranges held. 79% of real prices landed inside at 3 months, 82% at 6 months and 91% a year ahead.
• LightGBM's ranges were too narrow. Only 65%, 62% and 51% of real prices landed inside. Its ranges look precise and miss a third to half of the time.
• The classic tools went the other way. ETS and ARIMA caught 90% to 100% of prices, but their ranges were wider than the price itself, up to 6 times wider. A range from 0 to 600 euros is always right and never useful.

This is why accuracy alone can mislead. LightGBM's average is close at 3 months, but a revenue manager acting on its range would chase the wrong nights.

8175 Spanish markets

Two hotels are a small test. So we ran the same comparison on 175 Spanish hotel markets: regions, provinces and tourist towns, with the official average room price for each month, forecast 12 months ahead.

Here Timestral was told only two things about the future: the month, and whether Easter falls in it.

• It beat ETS on 165 of 175 markets and ARIMA on 160.
• It beat Chronos-2 on 156 and TimesFM on 131.
• It beat simply repeating last year's prices on 129.

9What helps a forecast

We tested 15 kinds of extra information, one at a time and together: the calendar, the season, normal weather for the time of year, school and public holidays in the countries guests come from, local events, football matches, daylight and Easter, and the rooms already sold for each future night.

• The time of year and normal weather do most of the work. With just those two, Timestral gave its best forecast a year ahead.
• Holidays, events, football and daylight help a little closer in. Adding them lifted Timestral by up to half a point at 3 and 6 months, and lowered it a year ahead, from 88.6% to 87.8%.
• Every factor at once scored worse. Giving Timestral every factor a year ahead dropped its accuracy to 77.0%. That far out, more inputs added more noise than signal.
• Information you cannot know in advance cannot help. We left out exchange rates, air traffic and web search trends, because nobody knows next year's values.

10Try it

• Drag the start date and watch the forecast and its range redraw.
• Switch on a factor, such as public holidays, and see how the forecast moves.
• Compare tools on the same nights, with each tool's accuracy shown.
• Upload your own file with at least 365 days of prices. It stays in memory for your visit and is never saved.

11What we learned

12Five moves for Monday

If you run pricing for a group of hotels, do these in order. Each takes days, not months.

1. Pull a year of clean daily prices for a handful of hotels, with the price actually paid per night.
2. Run the same test on your own history. Forecast a year ahead from past start dates and score it against what happened, next to your current software's recommendation.
3. Count how often reality lands inside the range. A tool whose 80% range holds 80% of prices is one you can build alerts on.
4. Flag only the nights outside the range. Let revenue managers review those, and track how many they act on.
5. Recalibrate your rules. Replace the judgement thresholds with ones measured against the forecast.

13The fine print

• Two hotels, public data. The daily result covers 2 hotels, with 12 one year forecasts, 44 at 180 days and 58 at 90 days on the scored start dates. Treat it as a strong early signal.
• The hotel group's own data was not used here. This note runs on public data only. The next step is the same test on a group's own history.
• Best setup per tool. Each tool is shown with the factor set that scored best for it at each horizon.
• Spanish markets are monthly averages. They test the long view across many markets. They do not test night by night pricing.
• Vendor claims are Mistral's. We did not run Mistral's own benchmarks again. On a laptop CPU, one Timestral forecast took 9 milliseconds with the smallest model and 85 milliseconds with the largest.
• Accuracy here is one minus the average gap, summed over all scored nights. It differs from the share of nights priced exactly right.

14Glossary

Accuracy

100% minus the average gap between forecast and real price, as a share of the real price.

Factor

Extra information given to a forecast, such as a holiday or an event.

Forecast

A prediction of future values from past ones.

Horizon

How far ahead the forecast looks: 90, 180 or 365 days here.

Known in advance

A factor whose future value is known when the forecast is made.

Pretrained model

A model trained once on many series, then used on new data with no further training.

Range (80%)

The band a forecast expects 80% of real values to fall into.

Rate shopper

A tool that collects competitors' prices for the same nights.

Revenue manager

The person who sets room prices for one or several hotels.

Revenue management software

A tool that forecasts demand and recommends room prices.