Rogendo/forex-lstm-models

Model Card for Forex-LSTM-Predictor-Collection

This repository hosts a collection of 20 specialized Long Short-Term Memory (LSTM) neural network models designed to forecast future price movements and directional trends for major currency pairs. Each model is trained on a specific pair and timeframe tuple.

Model Details

Model Description

This project explores the application of Deep Learning to financial time-series forecasting. Unlike generic models, this collection treats every currency pair and timeframe combination (e.g., EURUSD on 15m) as a unique environment requiring a dedicated model.

The models utilize historical OHLCV (Open, High, Low, Close, Volume) data to predict two simultaneous outputs:

1. Price Vector: The projected closing price for the next time step.
2. Directional Probability: A confidence score indicating the likelihood of the price moving Up or Down.

Note: These models are part of a larger full-stack research project ("FX-Predict") involving a FastAPI backend and a real-time dashboard.

• Developed by: Rogendo
• Model type: Dual-Output LSTM (Regression + Classification)
• Language(s): Python 3.12 (TensorFlow/Keras)
• License: MIT
• Finetuned from model: N/A (Trained from scratch)

Model Sources

• Repository: https://huggingface.co/rogendo/forex-lstm-models

Uses

Direct Use

These models are intended for:

• Algorithmic Trading Research: Backtesting ML strategies on forex markets.
• Signal Generation: Providing baseline buy/sell signals based on price action history.
• Market Analysis: Identifying potential trend reversals via directional confidence scores.

Downstream Use

The models are designed to be consumed by an inference engine (like the model_service.py in the FX-Predict app) which applies post-processing logic:

• Risk Management: Calculating dynamic Stop Loss/Take Profit levels using ATR (Average True Range).
• Confluence checking: Comparing predictions across timeframes (e.g., 15m vs 4h).

Out-of-Scope Use

• Unsupervised Live Trading: These models are Proof-of-Concept tools. They should not control real capital without a strict risk management wrapper (see Limitations).
• Long-term Forecasting: The models are optimized for the next candle ($t+1$) only.

Bias, Risks, and Limitations

The "Accuracy vs. Profitability" Paradox

During testing, a critical discrepancy was observed:

• High Directional Accuracy: The models frequently achieved >55% accuracy in predicting the color of the next candle.
• Low Win Rate (~40%): Without external logic, the raw model predictions often led to losses.
  • Cause: The model predicts the vector correctly but not the path. On volatile timeframes (5m/15m), market noise often triggered standard Stop Losses before the prediction materialized.
  • Mitigation: We strongly recommend using ATR-based dynamic stops rather than fixed percentage stops when using these models.

Data Limitations

• Data Source: Trained on data from yfinance (Yahoo Finance).
• Lag & Limits: Yahoo Finance data can be delayed. Crucially, intraday data (5m/15m) is restricted to the last 60 days. This introduces Recency Bias, as the model has not seen historical market regimes (e.g., the 2008 crash or 2020 volatility).

Feature Poverty

The inputs are currently limited to raw OHLCV. The models lack "Macro-Awareness" (Economic Calendar events) and Order Flow data, limiting their ability to react to news events.

Recommendations

Users should treat these models as Directional Compasses, not Crystal Balls. They work best when validated by "Market Breadcrumbs" such as:

• Session High/Low breaks (e.g., 9:30 AM NY Open).
• Key psychological levels (0.000 / 0.500).
• Multi-timeframe confluence.

How to Get Started with the Model

You need the .h5 model file and the corresponding .pkl feature scaler file to ensure data is normalized exactly as the model expects.

import numpy as np
import pickle
import os
from tensorflow.keras.models import load_model
from huggingface_hub import hf_hub_download

# 1. Select Pair and Interval
PAIR = "EURUSD_X"  # Options: GBPUSD_X, USDJPY_X, etc.
INTERVAL = "15m"   # Options: 5m, 15m, 30m, 1h, 4h

# 2. Download Files
model_path = hf_hub_download(repo_id="rogendo/forex-lstm-models", filename=f"{PAIR}_{INTERVAL}_model.h5")
scaler_path = hf_hub_download(repo_id="rogendo/forex-lstm-models", filename=f"{PAIR}_{INTERVAL}_features.pkl")

# 3. Load Model
model = load_model(model_path)
with open(scaler_path, 'rb') as f:
    feature_info = pickle.load(f)

print(f"Loaded {PAIR} {INTERVAL} | Lookback: {feature_info['lookback']}")

# 4. Prepare Dummy Data (Replace with real OHLCV data scaled via RobustScaler)
# Shape: (1, Lookback_Steps, 5_Features)
dummy_input = np.random.rand(1, feature_info['lookback'], 5)

# 5. Predict
prediction = model.predict(dummy_input)
price_change = prediction[0][0][0]
direction_conf = prediction[1][0][0]

print(f"Predicted Change: {price_change:.5f}")
print(f"Direction Confidence: {direction_conf:.2f}")

Training Details

Training Data

The models were trained on historical Forex data fetched via yfinance.

Pairs: EURUSD, GBPUSD, USDJPY, AUDUSD, USDCAD, USDCHF, NZDUSD.

Timeframes: 5m, 15m (approx 60 days history); 30m, 1h, 4h (approx 2 years history).

Training Procedure

Preprocessing

Scaling: RobustScaler (sklearn) was used to handle outliers in financial data.

Windowing: Data was transformed into sequences of 15 lookback steps.

Training Hyperparameters

Optimizer: Adam (learning_rate=0.001)

Loss Functions:

Price: mean_squared_error

Direction: binary_crossentropy

Batch Size: 32

Epochs: 50 (with Early Stopping patience=15)

Evaluation

Testing Data, Factors & Metrics

Testing Data

Data was split 80/20 for Training/Validation. Due to the rolling window nature of time series, the validation set represents the most recent market data available at the time of training.

Metrics

RMSE (Root Mean Squared Error): Measures price prediction magnitude error.

MAE (Mean Absolute Error): Average error in pips.

Directional Accuracy: % of time the model correctly predicted Positive vs Negative close.

Results

4H, 14min Models: Showed the highest stability and profitability because price trends are cleaner.

5M Models: Showed high noise. While directional accuracy remained >50%, the realizable profit was often eaten by spreads.

Technical Specifications

Model Architecture

The architecture is designed to capture temporal dependencies:

Input Layer: Shape (15, 5)

LSTM Layer 1: 50 units, Return Sequences=True, Dropout=0.2

LSTM Layer 2: 25 units, Return Sequences=False, Dropout=0.2

Dense Layer: 20 units, ReLU

Output 1 (Regression): 1 Unit (Price Change)

Output 2 (Classification): 1 Unit (Sigmoid - Direction)

Future Roadmap

To improve the "Realizable Profitability" of these models, the following upgrades are planned:

Data Pipeline Overhaul: Move away from yfinance to a professional provider (OANDA/Alpha Vantage) to access 2+ years of 5m data and faster api.

Feature Expansion: Triple the input features to include:

• Rolling Technical Indicators (RSI, MACD, Bollinger Bands) as inputs (not just validators).

• Time-of-day embeddings (to learn session volatility).

Architecture: Experiment with CNN-LSTM hybrids (to catch chart patterns) and Transformer models (TimeGPT).

For questions regarding the implementation or the "FX-Predict" dashboard integration, please open a discussion in the Community tab.