TimeH2O at work: three forecasts, three decisions

TimeH2O, a time-series foundation model from H2O.ai, forecasts each value with a range, and the range is where the decision lives: a team chooses how far into it to act. The three stories below show how a treasury team (FX hedging), a claims team (staffing and reserves), and a network team (anomaly detection) turn that range into a decision. They use sample data made for these examples, not results from a customer.

1. You give: a table of past rows. Weekly payments, weekly claim counts or network readings. You do not clean them into a special format.
2. TimeH2O reads: no training step. TimeH2O reads the rows when you ask. You do not build or train a model first.
3. You get: a forecast with a range. The middle value and how far it can move. Your team uses both to make a decision.

Know how much foreign cash will arrive, and lock in today's rate for it.

Banking · Corporate treasury

A US company sells to clients in Europe, the UK, Japan, Canada and Mexico. The clients pay in their own currency. If the euro falls before the cash arrives, the company loses money. Treasury can lock in today's rate (an FX hedge), but only for cash it is sure will come.

• Who uses it: a treasury analyst
• TimeH2O reads: 2 years of weekly payments
• TimeH2O forecasts: cash for the next 13 weeks

What TimeH2O reads, and how much to lock in

The input is a plain table with one row for each week and currency. TimeH2O fills in three new columns for the weeks to come: a low value, a middle value, and a high value. There is an 8-in-10 chance that the real cash falls between the low and high values.

The lock-in policy sets where, between the low and middle, the analyst locks in. Move the slider to see the trade-off. If you lock in more, you protect more cash from a rate move. But you also get more short weeks, and in a short week the company must buy the missing currency.

Know next week's claim volume and cost, and staff for it.

Insurance · Claims operations

A claim can take 40 days to close. Most of that time is spent waiting in a queue because nobody knows how many claims will arrive next week or how much they will cost. With too few adjusters, the queue grows. With too many, people sit idle. Finance must also set money aside for claims that have not closed yet (claims reserving).

• Who uses it: a claims operations manager
• TimeH2O reads: 5 years of weekly claim counts and costs
• TimeH2O forecasts: claims and cost per week, by class, for the next 8 weeks

What TimeH2O reads, and how much to staff for

The input is a plain table with one row for each week and class. Each row has the number of claims and their cost. TimeH2O fills in three new columns for future weeks: a low value, a middle value, and a high value. There is an 8-in-10 chance that the real number falls between the low and the high value.

The manager chooses which side to plan for. If you staff for the high side, claims rarely wait, but some adjusters are free in a quiet week. If you staff for the middle, you spend less, but in a busy week claims wait. Finance uses the same cost range to plan the reserve. The actuary still sets the final reserve. Change the plan and see the trade-off.

Find network faults hours before a cell site fails, not when customers call.

Telco · Network operations center

A cell site sends a reading every 15 minutes for traffic, dropped calls, errors, power, and battery heat. A fixed alarm limit fires late, when the site is already down. A fault starts earlier, as a reading that leaves its normal range (anomaly detection).

• Who uses it: a network operations center (NOC) operator
• TimeH2O reads: 30 days of readings
• TimeH2O forecasts: the normal range of each reading

What TimeH2O reads, and how wide the normal range is

Each cell site sends one row every 15 minutes. TimeH2O reads 30 days of these rows and forecasts a low and a high value for each reading in the next 24 hours. A reading outside that range twice in a row triggers an alert.

The operator chooses how wide the range is. A narrow range finds slow faults earlier but gives more false alerts. A wide range gives fewer false alerts, but a slow fault shows up later. A fast fault, like the battery heat at Site 117, leaves any range quickly. The chart shows Site 089, where the power output falls a little at a time.

One idea, three teams

The three stories use the same steps. TimeH2O reads the past rows. It forecasts a value with a range. A business rule uses the forecast to act early.

The range matters as much as the middle value. A treasury team locks in below the middle value. A claims team staffs for the high side. A network team alerts when a reading leaves the range.

The same pattern fits other forecasts. TimeH2O can be used for:

• Thousands of series, such as demand for every product in every store. One model covers them all, so nobody builds a model for each series.
• New items with a short history, such as a new store, a product launch, or a new machine.
• Slow-moving items. Each extra series costs almost nothing, so they get a forecast too.
• Capacity and cost metrics, such as cloud cost and call volume.
• Planning with a range, such as safety stock and staffing.
• Known drivers, such as promotions, holidays, and weather. These go in as extra series that help predict the target.

TimeH2O is not the right tool for every forecast. Before you start, ask three questions: how many series do you have, how much history, and must a regulator see the reasons for each decision? Many series, little history, and no regulator points to TimeH2O. If a regulator must see the reasons, use H2O Driverless AI. If the problem is a table of rows with many drivers, use TabH2O. If the question is what happens when you change something, use a causal or simulation model.

Open the apps

All three apps are in the TimeH2O use-case catalog, where you set up a run, step through what TimeH2O does, follow each step from the forecast to the result, and download the source code as a sample to get started.

TimeH2O forecasts the cash that will arrive in each currency. Treasury locks in the right amount.

TimeH2O forecasts weekly claims and cost with a range. The manager staffs adjusters for the high side.

TimeH2O forecasts the normal range of every reading. A fault shows up hours before the site fails.

To use TimeH2O in Python, install the h2ofm package. TimeH2O's open weights are on Hugging Face under the H2O Model Non-Commercial License; production use needs a license from H2O.ai, so talk to H2O before you build on them.

pip install h2ofm

A minimal example, adapted from the h2ofm quick start:

import numpy as np
from h2ofm import TimeH2O

model = TimeH2O.from_pretrained("h2oai/timeh2o-v1")
forecast = model.forecast(np.sin(np.arange(200) / 10), horizon=24)

forecast.point          # (24,), median forecast; also forecast.median
forecast.quantiles      # (24, 9), default levels 0.1 through 0.9
forecast.quantile(0.9)  # (24,)

You need a Hugging Face login, and to accept the model's terms, to download the weights. The package needs Python 3.12 or later.

The stories and apps use sample data made for these examples. The numbers show how the steps work. They are not results from a customer.

For published results, TimeH2O-v1 ranks #1 of 32 on fev-bench by win rate (91.9%), with a skill score of 50.8% (non-commercial models included), as of 24 October 2026; see the launch post. TimeH2O is a tool in H2O AI Cloud, the trusted Sovereign AI Platform, and its open weights run in your own environment under the H2O Model Non-Commercial License; production use needs a license from H2O.ai. Try it on your own series with the Python example above, or talk to H2O about production use.