logo
    logo
    Log in
    Browse help/Pattern Models

    Pattern Models

    Requires SuperAvailable on: Web · iOS · Android

    On this page

    A pattern model learns a chart setup from examples you mark. You show it a few charts that look right to you, and it scores every stock by how closely its recent chart looks like them. Then you scan with it like any other filter.

    Some setups are easier to recognize than to describe. A pattern model suits those: you point at the charts, and it works out what they have in common.

    How a model learns

    • An example is a window of a chart. A window is the 128 daily bars ending on a date you pick, called its anchor. The model only sees the bars inside the window.
    • Positive examples are windows you would call a match. A model needs at least 3. Negative examples are near misses you reject. They're optional and teach the model what to leave out.
    • Training fits the model to your examples, then scores today's chart of every stock in the market. It takes a few seconds.
    • The score is a rank. Each stock gets a similarity to your examples, ranked against the whole market. A scan keeps the top share you choose. A newly trained model starts at the top 1%.

    Models train on daily bars. Scale stretches the window: scale 1 covers 128 sessions (about six months), scale 2 covers 256 and scale 4 covers 512. Every example in one model uses the same scale.

    Tour

    Open Model Trainer in the sidebar on the web, or the Models tab on the mobile app. Train is the workspace for one model. My Models lists your models and your drafts. On mobile it lists the presets too.

    Slide the window along the chart, then add it as an example.

    • Examples and Candidates: the two modes. On Examples you pick windows. On Candidates you judge what the model found.
    • Ticker box: the stock on the chart. From a saved scan lets you step through a saved scan's results instead.
    • Timeframe and Scale: the window's shape. Timeframe is daily on every new model. Scale locks after the first example.
    • Window strip (under the chart): the shaded band is the window. Drag it to move the anchor. Older bars load as you drag back.
    • Add as example and Reject as example: file the window as a positive or a negative.
    • Examples list: your examples with their anchor dates, and how many more you need.
    • Train model: fits the model. It turns into Retrain once the model has trained.
    • Model menu: Edit, New model, Save model, Load model, Model presets, Export examples (.csv) and Delete.

    Train a model

    1. Open Train on the web, or tap New model on My Models on mobile. A new model starts as Untitled model. Use the pencil next to the name to name it and add notes.
    2. Pick a ticker whose chart shows the setup.
    1. Drag the shaded band until its last bar sits where the setup completes.
    2. Click Add as example. For a near miss, click Reject as example.
    1. Repeat on other stocks until the list shows at least 3 examples. Use different tickers: a model trained on one stock learns that stock, not the setup.
    2. Click Train model. The first training saves the model, and it takes one of your plan's model slots.

    You can also start from a chart elsewhere in the app:

    • Right-click a ticker anywhere in the app and choose Train a model on $AAPL.
    • On a ticker's chart, turn on Training mode, drag across the chart to frame a window, then click Train. The window snaps to about 6 months, 1 year or 2 years of bars. The first time you turn the mode on, a short tour shows how.

    A model you haven't saved yet is a draft. Drafts stay in this browser or on this device, under Drafts on My Models, and take no model slot.

    Review what it found

    After training, open Candidates. The model shows the stocks it scored highest, above and below your threshold.

    Each candidate shows its rank and similarity. The threshold sets where a scan cuts the list.

    1. For each candidate, click Accept, Reject or Skip. The keys A, R and S do the same, and ← and → move between candidates.
    2. If the setup ends a few bars before or after the window, slide Window, then click Accept adjusted.
    1. Accepted candidates become positive examples and rejected ones negatives. The button then reads, for example, Retrain with 4 new examples. Click it to fit them in.
    2. Set Threshold. The slider shows how many stocks match today at that cut. Raise it until what's left looks like your setup. A filter you add after this starts at this cut, and a scan that already has the model keeps its own.

    By default, candidates come from today's charts. To see what the model finds in the past, click Sweep a range and pick Last 12 months or a calendar year. A sweep scores every window in that range and usually takes under a minute. You can keep reviewing while it runs.

    On mobile, after a sweep the threshold stops at the lowest score the sweep found, marked by an amber dot under the slider. Tap the dot to see why: below that score the list can't grow.

    Start from a preset

    Presets are models scanr trained, with the examples they learned from.

    1. On the web, open the Model menu and click Model presets. On mobile, presets are listed on My Models.
    2. Open one to see its examples and its candidates. A preset is read-only.
    3. Scan with it as it is: Scan with this preset on the web, Scan on its row on mobile. Scanning with a preset takes no model slot.
    4. Save a copy to make it yours. The copy scans straight away, takes one model slot, and you can add examples and retrain it.

    A preset marked Unavailable hasn't finished training on our side. It can't scan or be copied until it has.

    Use it in a scan

    1. Add a Pattern model filter: click + on a filter group (or Add filter in an empty group), then Pattern model. On mobile, tap Add filter in a group on the Filters screen.
    2. Click Choose a model and pick one of yours under Ready to scan, or a preset under Presets.
    3. Set the cut-off. The filter starts at the model's own threshold, and shows how many stocks match today. On mobile, the cut-off reads Keep the top, with shortcuts from 0.5% to 10%. The scan stores the cut-off as a share of the market, so retraining the model keeps the scan about the same size, and moving the model's threshold later leaves the scan's cut-off alone.
    4. Run the scan.

    To rank the results, set Sort to Pattern score in the universe settings and scan again. On the web, add the Pattern score column to see each stock's rank within the model's universe. Scan on a row of My Models opens the scanner with that model as a filter.

    A scan with more than one model sorts each stock on its highest similarity across them, and the column shows the rank from that model. Ranks from different models don't compare, so they may not line up with the sort order.

    You can also name a model in the AI terminal: "Scan for stocks that match my Tight Flag model above $10". Add "(preset)" to the name to use a preset.

    A pattern model works with Time Machine on past dates, on restated prices only.

    Reference

    Model status

    StatusWhat it means
    DraftNot trained yet. Until you save it, it stays in this browser or on this device.
    TrainingThe fit is running.
    ReadyTrained and ready to scan with.
    Needs retrainingThe price history behind an example changed, usually after a stock split, or we improved the pattern encoder. Retrain it. Your examples are kept. After a price history change it still scans on its last fit. After an encoder change, scans refuse it until you retrain.
    FailedThe last fit didn't finish. Your examples are saved.

    Limits

    LimitSuperSuper Pro
    Saved models520
    Pattern-model filters per scan33
    Positive examples to trainat least 3at least 3

    Presets never count toward your saved models. Pattern models are part of Super and Super Pro. On Free, the Model Trainer shows the upgrade, and training, saving and scanning stay off, presets included. A lapsed paid plan counts as Free: your models stay listed, and you can delete them.

    Tips

    • Pick examples from different stocks and dates. Variety teaches the setup; five windows of one ticker teach the ticker.
    • Put the anchor on the bar where you'd act. The model learns what the chart looks like up to that bar.
    • Reject near misses during review. Negatives sharpen a model faster than more positives.
    • Raise the threshold until the matches look right, then add the model to a scan. A loose cut returns stocks that only share the trend.
    • Combine a model with plain filters such as Market cap≥2B or Previous Day Volume>500K to drop thin stocks.

    Troubleshooting

    IssueCauseSolution
    Train model stays offThe model has fewer than 3 positive examples.Add more windows you'd call a match.
    "Not enough usable examples survived"Some examples had too little history behind them.Add a few more examples, then train again.
    "Your examples are too much alike"Every example is the same stock.Add examples from other tickers.
    "We could not rebuild the price history behind one of your examples"A ticker's history changed or is missing.Remove the examples marked in the list and train again.
    "The pattern encoder was unavailable" or "The fit took too long"A temporary problem on our side.Your examples are saved. Train again in a minute.
    I can't change ScaleScale locks after the first example.Remove every example to change it, or start a new model.
    My scan says a pattern model "needs retraining"We improved the pattern encoder, and the model hasn't moved onto it.Open it in the Model Trainer and click Retrain.
    It says I have used all my model slots, or "Model limit reached"Every model slot on your plan is used.Delete a model. On Super, you can also move up to Super Pro.
    A past-date scan refuses a pattern modelPast dates run on restated prices only, and the stored history has limits: the newest week arrives a few days late, and older dates may predate it.Switch the scan's price basis to Restated, or pick another date.

    Was this page helpful?

    Still stuck?

    Ask scanr AI in the AI Terminal. It answers from this page. Included with Super and Super Pro.