AI Studio Interactive Teaching Lab

AI Studio Mock Controls

Answer mode

Trading Goal preset, not algorithm

Preset selects a continuous ATR-normalized forward-move answer. It usually behaves like a signed score around zero.

Step 1 Answer Families next capture

Fixed-horizon returnsready ATR barrier raceH/L/C Swing reversalH/L Mean reversionSMA Volatility regimestate

These settings build future answer columns only when a new capture is created. A saved capture stays frozen.

Step 2 Answer Mapping Regression

Prediction Style

Regression predicts a number. Only numeric answer columns are available.

Generated Question Family

Selected Training Horizon 10 bars

ATR Barrier Distance 2.0 x ATR

The answer is a generated historical label. Live ONNX input never receives future bars.

Feature Candidates inputs

Selected features are current-row evidence. Their order becomes the ONNX input contract.

Step 3 Model Configuration fit method

Algorithm Enable hyperparameter tuning slower

LightGBM is a fast boosted-tree first choice for tabular regression when available. Leave tuning off for a first pass.

What The Model Is Being Asked To Predict

Target_ATRReturn10

ATR-normalized forward move

At each historical row, AI Studio records current features and creates a future answer by looking 10 bars ahead.

Current Row

Bar 78 – close 104.12 – ATR 1.01

Bar 78

Training answer

+0.84

Known later in historical data. This is what the model tries to learn.

Model output

+0.61

Estimated from selected current-row inputs only.

Common first use

> +0.50

Plot first, then test thresholds or classes with Builder logic.

Model Card And ONNX Block

Ready to deploy

What This Model Outputs

A signed numeric score in ATR units. Positive is upward expected movement, negative is downward expected movement.

How To Read It

Zero is the centerline. Values near zero are weak. Larger positive or negative values need your own thresholds.

Training Notes illustrative

Regression with 7 selected inputs. Time split: 80 percent train, 20 percent held out. Tuning is off.

AI Studio Log illustrative

Canvas-Linked Top Inputs illustrative

ONNX Block Mapping illustrative

ONNX Model Block7 inputs

Condition / Gateoutput > 0.50

The model output is still just one value. Builder logic decides what threshold or class matters.

Demo threshold 0.50

Result Plot Firstmoving

Review scale, flatlines, warmup gaps, and mapping warnings before using strategy blocks.

The Whole Workflow In One Picture

0 Export

Choose OHLCV and block outputs. This is the evidence AI Studio can capture.

1 Capture

Freeze rows and generate answer columns. Future labels are created here.

2 Answer

Pick one target and a clean feature set. One model learns one answer column.

3 Train

Fit an algorithm to the feature-to-answer relationship and review held-out behavior.

4 Deploy

Copy `.onnx` plus `.fusion.json`, then load it into an ONNX block and map ports.

Plain-English Target Decoder

Regression Outputs

ATR Return: signed move scaled by ATR.
Future Return: signed raw percent move.
Signed Reversion: directional move toward the mean.
Future Volatility: expected movement size, not direction.

Classification Outputs

Direction: down, neutral, or up.
Barrier Race: upper barrier, lower barrier, or no winner.
Swing Reversal: swing low, neutral, or swing high.
Regime: low, normal, or high state bucket.

Recommended First Session

Start with 5 to 12 features you can explain.
Use ATR-Normalized Return at 10 bars for a first regression test.
Use Direction (3-Class) for a first clean classification test.
Leave tuning off until capture, feature mapping, deploy, and ONNX plot behavior all make sense.
Plot the ONNX output before connecting it to Strategy logic.