usamaahmedsh/elliott-wave-scorer-training

---
license: mit
language:
- en
tags:
- elliott-wave
- synthetic
- finance
- time-series
- scorer-training
- binary-classification
- technical-analysis
size_categories:
- 100K<n<1M
---

# Elliott Wave Scorer — Training Dataset

Binary-labelled dataset for training a neural network scorer that distinguishes
valid Elliott Wave patterns (label=1) from invalid/borderline patterns (label=0).

## Dataset Summary

| Field | Value |
|---|---|
| Total rows | 30,000,000 |
| Good samples (label=1) | 15,000,000 |
| Bad samples (label=0) | 15,000,000 |
| Bad / Good ratio | 1.00 |
| Wave types | Corrective, bearish_impulse, impulse |
| Run | `run_20260318_135833_t5000000` |

## Label Semantics

| Label | Meaning | How generated |
|---|---|---|
| `1` | Valid Elliott Wave | Passed all 4 generation tiers + diversity filter |
| `0` (Type A) | Wrong Fibonacci ratios | Passed T1+T1b geometry shape, **failed** T2 Fibonacci gate |
| `0` (Type B) | Borderline pattern | Passed T1+T1b+T2 (correct Fib ratios), **failed** T3 MVN score threshold |

Type A = hard negatives (clearly wrong geometry).
Type B = soft negatives (plausible geometry, but low density vs real distribution).

## Schema

| Column | Description |
|---|---|
| `rule` | Wave type: `Corrective`, `impulse`, `bearish_impulse` |
| `wave_config` | Wave config string |
| `geo_0..geo_4` | Geometry ratios |
| `ensemble_score` | MVN-based quality score (0.0 for Type A bad, ~0.0–0.79 for Type B bad, ≥0.80 for good) |
| `fib_score` | Fibonacci proximity score |
| `is_synthetic` | Always `True` in this dataset |
| `label` | **Target variable**: `1` = good, `0` = bad |

## Recommended Input Features

```python
features = ['geo_0', 'geo_1', 'geo_2', 'geo_3', 'ensemble_score', 'fib_score']
target = 'label'
```

Optionally add derived Fibonacci-distance features:
```python
FIBS = [0.236, 0.382, 0.5, 0.618, 0.786, 1.0, 1.272, 1.618, 2.0, 2.618]
for i in range(4):
df[f'fib_dist_{i}'] = df[f'geo_{i}'].apply(
lambda v: min(abs(v - f) for f in FIBS))
```

## Training Notes

- Dataset is pre-shuffled (50/50 split, random seed)
- Stratify by `rule` when splitting train/val/test
- A 3–4 layer MLP or gradient-boosted tree works well given the clear feature separation
- `ensemble_score` alone achieves strong separation; geometry features add complementary signal

## Related Dataset

Good samples only (for synthesis evaluation): [`usamaahmedsh/synthetic-elliott-waves`](https://huggingface.co/datasets/usamaahmedsh/synthetic-elliott-waves)