rckrdmrd 75c4d07690 feat: Initial commit - ML Engine codebase

Hierarchical ML Pipeline for trading predictions:
- Level 0: Attention Models (volatility/flow classification)
- Level 1: Base Models (XGBoost per symbol/timeframe)
- Level 2: Metamodels (XGBoost Stacking + Neural Gating)

Key components:
- src/pipelines/hierarchical_pipeline.py - Main prediction pipeline
- src/models/ - All ML model classes
- src/training/ - Training utilities
- src/api/ - FastAPI endpoints
- scripts/ - Training and evaluation scripts
- config/ - YAML configurations

Note: Trained models (*.joblib, *.pt) are gitignored.
      Regenerate with training scripts.

Co-Authored-By: Claude Opus 4.5 <noreply@anthropic.com>

2026-01-18 04:27:40 -06:00

3.8 KiB

Raw Blame History

Attention Score Model Training Report

Generated: 2026-01-07 03:39:38

Overview

The attention model learns to identify high-flow market moments using volume, volatility, and money flow indicators - WITHOUT hardcoding specific trading hours or sessions.

Configuration

Symbols: GBPUSD
Timeframes: 5m, 15m
Training Data Cutoff: 2024-12-31
Training Years: 5.0
Holdout Years: 1.0

Model Parameters

Parameter	Value
Factor Window	200
Horizon Bars	3
Low Flow Threshold	1.0
High Flow Threshold	2.0

Features Used (9 total)

Feature	Description
volume_ratio	volume / rolling_median(volume, 20)
volume_z	z-score of volume over 20 periods
ATR	Average True Range (14 periods)
ATR_ratio	ATR / rolling_median(ATR, 50)
CMF	Chaikin Money Flow (20 periods)
MFI	Money Flow Index (14 periods)
OBV_delta	diff(OBV) / rolling_std(OBV, 20)
BB_width	(BB_upper - BB_lower) / close
displacement	(close - open) / ATR

Training Results

Model	Symbol	TF	Reg MAE	Reg R2	Clf Acc	Clf F1	N Train	High Flow %
GBPUSD_5m_attention	GBPUSD	5m	0.6262	0.1596	59.08%	56.12%	310727	24.3%
GBPUSD_15m_attention	GBPUSD	15m	0.6953	0.2534	60.20%	56.62%	104434	35.7%

Class Distribution (Holdout Set)

Model	Low Flow	Medium Flow	High Flow
GBPUSD_5m_attention	6238 (8.4%)	49712 (67.3%)	17951 (24.3%)
GBPUSD_15m_attention	686 (2.8%)	15199 (61.5%)	8830 (35.7%)

Feature Importance

GBPUSD_5m_attention

Rank	Feature	Combined Importance
1	ATR	0.3542
2	ATR_ratio	0.1580
3	BB_width	0.1348
4	CMF	0.0814
5	MFI	0.0610
6	volume_ratio	0.0604
7	volume_z	0.0552
8	OBV_delta	0.0499
9	displacement	0.0450

GBPUSD_15m_attention

Rank	Feature	Combined Importance
1	ATR_ratio	0.3374
2	ATR	0.2368
3	volume_z	0.1040
4	volume_ratio	0.0950
5	BB_width	0.0617
6	MFI	0.0460
7	CMF	0.0437
8	displacement	0.0383
9	OBV_delta	0.0370

Interpretation

Attention Score (Regression)

< 1.0: Low flow period - below average market movement expected
1.0 - 2.0: Medium flow period - average market conditions
> 2.0: High flow period - above average movement expected (best trading opportunities)

Flow Class (Classification)

0 (low_flow): move_multiplier < 1.0
1 (medium_flow): 1.0 <= move_multiplier < 2.0
2 (high_flow): move_multiplier >= 2.0

Trading Recommendations

Filter by attention_score: Only trade when attention_score > 1.0
Adjust position sizing: Increase size when attention_score > 2.0
Combine with base models: Use attention_score as feature #51 in prediction models
Time-agnostic: The model identifies flow without hardcoded sessions

Usage Example

from training.attention_trainer import AttentionModelTrainer

# Load trained models
trainer = AttentionModelTrainer.load('models/attention/')

# Get attention score for new OHLCV data
attention = trainer.get_attention_score(df_ohlcv, 'XAUUSD', '5m')

# Filter trades
mask_trade = attention > 1.0  # Only trade in medium/high flow

# Or use as feature in base models
df['attention_score'] = attention

Files Generated

models/attention/{symbol}_{timeframe}_attention/ - Model directories
models/attention/trainer_metadata.joblib - Trainer configuration
models/attention/training_summary.csv - Summary metrics

Report generated by Attention Model Training Pipeline

3.8 KiB Raw Blame History