Hugging Face's logo

kyLELEng
/
mag7-tabnet-model

PyTorch
Joblib
finance
tabnet
tabular
mag7
private-dataset-trained
Model card Files
xet
 Community
mag7-tabnet-model / README.md
kyLELEng's picture
kyLELEng
Upload README.md with huggingface_hub
54d9eae verified
preview code
|
raw
history blame contribute delete
2.03 kB
metadata 
license: other
tags:
  - finance
  - tabnet
  - tabular
  - pytorch
  - mag7
  - private-dataset-trained
library_name: pytorch

MAG7 TabNet Model

Public optimized TabNet classifier trained on a private MAG7 daily ML dataset.

Data Access

The model is public. The dataset remains private and is not distributed in this repository.

  • Private training dataset: kyLELEng/mag7-ml-daily-dataset-5y
  • Dataset file used internally: mag7_ml_daily_dataset_5y.csv
  • No raw dataset rows, target labels, or future-return columns are included here.

Training Setup

  • Task: binary classification
  • Target: target_next_10d_outperform_qqq
  • Architecture: TabNet classifier
  • Feature count: 247 selected numeric features plus ticker dummies
  • Split: chronological train / validation / test
  • Train end: 2024-03-06
  • Validation end: 2025-02-19
  • Test start: 2025-02-20
  • Hardware: Hugging Face Jobs, a10g-large
  • Device used: CUDA

Results

Validation was used for hyperparameter selection. Test metrics are out-of-sample on the final chronological holdout.

Metric Value
Best validation AUC 0.5404
Test AUC 0.5037
Test accuracy 0.5009
Test precision 0.4801
Test recall 0.1614

The holdout edge is weak. Treat this as a research artifact, not as a production trading signal.

Best Trial

  • n_d: 32
  • n_a: 32
  • n_steps: 5
  • gamma: 1.3473
  • lambda_sparse: 0.0000595
  • Learning rate: 0.000251
  • Mask type: entmax
  • Best epoch: 2

Files

  • tabnet_model.zip
  • scaler.joblib
  • feature_columns.json
  • metrics.json
  • trial_results.csv
  • test_scores_public.csv

test_scores_public.csv contains dates, tickers, and model scores only. It intentionally excludes labels and future returns.

Intended Use

This model is for personal research and model comparison on MAG7 daily technical features. It is not investment advice and should not be used for live trading without separate validation, walk-forward testing, slippage assumptions, and risk controls.