Some setups don't fit in a formula. You know a clean stage 2 base when you see one, but try writing it as moving averages and ranges and you either miss half of them or drown in noise. The Model Trainer lets you skip the formula: show Scanr.ai a few charts of the pattern, and it scans the market for more.

It shipped this week for Super and Super Pro, and I recorded a 15-minute walkthrough of building one from scratch, above. Here's the short version.

Your model book

Model Trainer sits in the sidebar with two pages. My Models is your model book: every pattern you've trained, with the example charts behind it. Train is where you build one.

A model is a chart pattern you've taught by example. It needs three examples at minimum, and more examples make a sharper model. Once it's trained it lives in your account, and you can scan with it like any other filter.

Pick three examples

In the video I build a stage 2A model from three bases that broke out:

Tesla (TSLA) in late October 2019, after months of going sideways, ending on a big green candle
Seagate (STX) in mid-November 2023
Nvidia (NVDA) in October 2019, a long base

Load a ticker, slide the window over the stretch of chart that shows the pattern, and press Add as example. For a stage 2A breakout, the window should hold a flat, quiet base and the candle that breaks out of it.

Models train on daily bars. To cover a longer base, raise the Scale rather than looking for a weekly chart: a stretched daily window does the same job. The model won't follow stage analysis rules to the letter, and some matches will be a little off. That's expected.

Train, then judge what it finds

With three examples in, press Train model. It takes a few seconds, then the Candidates tab lists today's charts that look like yours. The threshold slider sets how close a chart has to be to make the list. On October 29, 48 stocks cleared it.

To look further back, Sweep a range and choose Last 12 months. The sweep scores every daily window in the past year against your model, so it turns up far more candidates than today's market alone.

Then go through the candidates:

Accept the ones that are clearly your pattern. Spotify (SPOT) was a good stage 2A breakout that day, so it went in.
Reject the clear misses, like a chart that's breaking down.
Skip anything you're unsure about.

Be strict in both directions. The model learns from a handful of charts, so one bad accept or one wrong reject moves it a lot. When in doubt, skip.

Each verdict becomes an example for the next Retrain. Retrain, review the new candidates, and repeat until the list looks like the charts you'd trade. Then Save model.

Scan the market with it

Start a new scan, add a group, and add the Pattern model filter. Pick your model, run the scan, and you get stocks whose charts match it today.

To see the closest matches first, open the cog and sort by Pattern score. In the video, Spotify, one of the charts we'd just accepted, came up near the top, which is what you want from a model trained on it. You can also ask the AI terminal to scan with one of your models by name.

Start from a preset

If you'd rather not start from nothing, Model presets has four we trained: stage 2A, stage 1, flat base and a volatility contraction pattern (VCP). Scan with one as it is, or Save a copy and keep training it on your own examples. Presets also show up in the scanner's Pattern model filter, grouped by version, and we'll release new versions as we tune them.

Super keeps 5 models and Super Pro keeps 20. You can train them in the iOS and Android apps too. The full list of what shipped with the Model Trainer is in the Oct 28 changelog.

Jump to a part of the video

0:33 My Models
2:02 Adding examples
3:19 Timeframe and scale
5:17 Training and the Candidates tab
7:49 Accept, reject or skip
10:48 Scan with your model
12:34 Preset models

What's next

The Alerts Hub lands next week, and custom indicators, written as your own scripts, follow a couple of weeks after that. There's a quick look at the end of the video.

Train your first model →

Tell us what you train, or what you'd like the trainer to do next, at feedback.scanr.ai or [email protected].

— Eoin & The Scanr.ai team