AI Studio Answer Guide: Direction (3-Class)

Status: Per-family deep reference (companion to the consolidated Fusion AI Studio & ONNX User Guide) Audience: first-time AI Studio user Answer family: Direction (3-Class) Prediction style: Classification

1. The Plain-English Question

Direction (3-Class) asks:

This is one of the clearest answer families when the user wants:

• explicit up target label
• explicit down target label
• explicit neutral target label

instead of one continuous score.

2. What Fusion Is Actually Building

Example:

• Target_Direction10

Meaning:

Fusion looks at the forward return over the chosen horizon and assigns one of three labels:

• +1 = up
• 0 = flat / neutral
• -1 = down

Plain-English rule:

if forward return > threshold      -> +1
if forward return < -threshold     -> -1
otherwise                          ->  0

That threshold is the neutral band you define in Step 1.

3. Why This Family Is Important

This family creates a three-class target-label answer:

That makes it easier to reason about than:

• Up / Down
• because 0 there means not up, not necessarily short
• Future Return
• because Future Return is continuous and still needs thresholds decided later

Direction (3-Class) is the cleanest built-in family when you want an explicit neutral class.

4. What The Step 1 Flat-Zone / Neutral Band Actually Does

This is the family where that setting matters most.

In Step 1, the field currently shown as:

• Direction Neutral Band (ignore tiny moves)

is the capture-time neutral band for direction-style labels.

It decides how large the future move must be before Fusion calls it:

• up
• down

instead of:

• flat

Example

If the threshold is:

0.001

that means roughly:

• future return greater than +0.001 -> +1
• future return less than -0.001 -> -1
• between those values -> 0

Very important

This setting affects how Direction (3-Class) labels are created during capture.

It is not:

• a live ONNX dead-zone
• a later trading threshold
• a universal setting for every answer family

5. What Look-ahead Windows Means For Direction (3-Class)

If Step 1 says:

1,5,10,20

Fusion can create:

• Target_Direction1
• Target_Direction5
• Target_Direction10
• Target_Direction20

Then in Step 2 you choose one of those for this model.

That means:

• Step 1 creates several possible direction labels
• Step 2 chooses one horizon for the actual model

6. What The Model Output Means

The target-label meaning of the classes is:

• +1 = up
• 0 = flat / neutral
• -1 = down

The ONNX block returns the decoded class output. User-defined conditions decide how that output is used in a workflow or strategy state.

Important current product note

Classification models may still need class decoding to recover target-label meaning cleanly from the saved model contract.

So the correct mental model is:

not:

This is a class decision, not a continuous signed score.

7. What 0 Means Here

For Direction (3-Class), 0 means:

• the future move stayed inside the neutral band
• or, in target-label language, “no clear directional edge according to this label design”

That is why this family is so useful:

• it gives you a real neutral class
• it is not forcing every bar into up or down

8. When Direction (3-Class) Is A Good Choice

Use it when you want:

• an explicit neutral target label
• a cleaner down / neutral / up target label than Up / Down
• a more rule-like target than a continuous regression score

It is a good first choice when:

• you want the model to return direction target buckets
• you want a neutral middle
• you want to reduce overtrading on tiny moves

9. When Direction (3-Class) Is Not The Best Choice

It is usually not the best first choice when you really mean:

• “How much move in ATR units do I expect?”
• use ATR-Normalized Return
• “How much raw percent move do I expect?”
• use Future Return
• “Which ATR barrier gets hit first?”
• use ATR Stop / Target Race
• “Is this bar later a confirmed swing high or low?”
• use Swing Reversal

10. Step-By-Step Direction (3-Class) Walkthrough

Step A: Prepare the raw dataset

In Export, make sure:

• OHLCV is on
• the features you want are included

Step B: Set the label behavior in Step 1

In AI Studio Step 1:

• choose Look-ahead Windows
• choose the Direction Neutral Band

This is the key design decision for this family.

Plain-English rule

You are deciding:

Step C: Create the prepared dataset

Click Create New Capture.

Now Fusion creates direction columns such as:

• Target_Direction5
• Target_Direction10
• Target_Direction20

based on the chosen neutral band.

Step D: Choose the actual training question

In Step 2:

• choose Classification
• choose Generated Answer Columns
• choose Direction (3-Class)
• choose one horizon

Example:

• 10 bars

Now the model is learning:

• Target_Direction10

Step E: Choose the features

Pick the inputs that may help answer the question.

Because this is a class problem, good features are often:

• momentum
• trend
• order flow
• volatility context
• mean-distance context

Step F: Train

Choose:

• model name
• algorithm
• tuning on/off

Good first algorithms:

• LightGBM Classifier if available
• fallback: Random Forest Classifier

Step G: Read the result

The right way to think about the output is:

+1 = up
 0 = neutral
-1 = down

not:

the slope of the plotted encoded class line

11. Good First Session

If you are new to AI Studio and want the cleanest first direction-class workflow:

Step 1

• Look-ahead Windows: 5,10,20
• Direction Neutral Band: start with the default

Step 2

• Prediction Style: Classification
• Answer Column Source: Generated Answer Columns
• Generated Question Family: Direction (3-Class)
• choose one horizon such as 5 or 10 bars

Step 3

• Algorithm: LightGBM Classifier
• fallback: Random Forest Classifier
• tuning off for the first pass

This gives one of the cleanest down / neutral / up target-label workflows in AI Studio.

12. How This Differs From Up / Down

This difference needs to be very clear:

Up / Down

• 1 = up
• 0 = not up

That means 0 mixes:

• flat bars
• down bars

Direction (3-Class)

• +1 = up
• 0 = flat
• -1 = down

That makes Direction (3-Class) the better family when you want symmetric down / neutral / up target-label logic.

13. What Not To Do

Do not assume the neutral band affects every answer family

This setting is mainly for direction-style label creation.

Do not assume 0 means “small regression value”

Here, 0 is a class label, not a continuous score.

Do not use the encoded class line like a normal oscillator

This is a classification family. The class meaning matters more than the raw numeric plot shape.

Do not assume wider neutral band is always better

A wider band creates more flat labels.

That may reduce noise, but it can also leave fewer decisive up/down target-label examples.

14. Current Product Notes From The Direction Guide Pass

These are the UI/product improvements this guide exposes.

A. The Step 1 neutral-band field should name this family directly

Suggested copy:

B. Step 2 should reinforce the resolved answer column

Example:

C. The product should clearly distinguish class outputs from analog scores

Suggested copy:

D. Classification decode still needs to stay visible as a product follow-up

The manual can explain trader meaning now, but the runtime/UI should continue moving toward explicit decoded trader-facing class meaning.

15. One-Sentence Rule To Remember

Direction (3-Class) means: