File size: 2,025 Bytes
960eb1b
 
 
 
 
 
 
 
 
 
 
 
 
 
8f77c32
960eb1b
54d9eae
960eb1b
54d9eae
960eb1b
54d9eae
 
 
960eb1b
54d9eae
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
---
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.