EE Holdings is opening Machine Learning to its first teams. It is for anyone with a spreadsheet and a question, such as which customers will leave or what a house will sell for, who wants a model they can check before they rely on it.
A model is a program that learns a pattern from examples so it can make a guess about new ones. The hard part is not getting a model to produce a number. It is knowing whether that number is any good. Machine Learning is built around answering that question honestly.
Every score next to a simple guess
Every result is shown next to a baseline: the simplest possible guess, like always picking the most common answer or always predicting the average. A model has to beat it to be worth using.
Scores are measured on held-back rows, examples hidden from the model while it learns and used only at the end. The platform also checks for overfitting, which is when a model does much better on rows it studied than on new ones, like a student who memorized last year's test.
A written contract before any training
Each model lives in a container with a contract. Before anything trains, the contract states the inputs, the exact thing being predicted, its units, how far ahead it looks, the baseline to beat and how it will be scored. No contract, no training.
Splits are chosen to fit the data. Forecasts are tested in time order with a gap, so the model never sees the future. Lab data is split by batch, so near-duplicate samples cannot inflate the score.
What we found on 19 public datasets
We ran the same training engine a container uses on 19 public datasets and published every result, including the weak ones. A few examples:
- Handwritten digits: a simple guess gets 7.5 out of 100 right. Our model gets 98.
- Blood cell microscope images: a simple guess gets about 11 out of 100 right. Our model gets about 82.
- California housing prices: a simple guess misses by about $90,500 on average. Our model misses by about $29,500.
- Bike sharing demand: a simple guess misses by 175 rentals an hour on average. Our model misses by 43.
- Brent crude oil, 20 days ahead: our model is only about 1% better than a simple guess.
The last result is in the list on purpose. Short-term commodity prices are close to unpredictable, and a platform that hid that would not be worth trusting with anything else. These are single test runs on public data, and results on your own data will differ. The full list, with sources and fit checks, is on the datasets page.
What is in it
- Several model types, one fair comparison. EE VZN, our gradient-boosted engine, runs alongside random forests, linear models, image models and sequence models, all scored the same way. The best one is chosen on practice tests, never on the final test.
- Train on your own hardware. A small Python Worker Kit runs on a laptop, GPU box or lab server. Your data goes only to your own workers, and results come back with checksums so they can be reproduced.
- Versions with a paper trail. Each version records its dataset and settings. Approving a new one needs a written reason, and every change goes into an audit log.
- An API for your other apps. Keys with limited, revocable access let other platforms upload data, start a run and read the results.
- A guide for every container. Each container has an avatar you can ask what is missing from the contract, what a score means or what to try next, in plain words.
Built first for three kinds of data
- Life science and lab data: protein panels, DNA sequences, clinical measurements and microscope images.
- Markets and commodities: forecasts with ranges, tested in time order.
- Manufacturing and operations: pass/fail and defect prediction from sensor data, and demand forecasting.
How to get access
Machine Learning is opening to a small group of teams first. The Workspace plan is free during early access. Paid plans for teams that connect their own hardware are arranged directly while we set prices, and published prices will come before anyone is asked to pay self-serve. Managed cloud training and an enterprise plan are planned, not available yet.
Bring one question and a sample of your data. Request early access and tell us what you want to predict, and we will help set up your first container.
Links
- Machine Learning
- All 19 test runs
- Glossary of terms
- Plans
- EE Holdings companies directory
- More EE Holdings news
Machine Learning is part of the EE Holdings portfolio. Test results are from engine v7, run Nov 8, 2026, on public datasets. They are not a promise of results on your data. Features described as planned are not yet available. Published Nov 8, 2026.
