README.md

---
language: en
license: mit
library_name: sklearn
tags:
- sklearn
- gold-price-prediction
- time-series
- classification
- financial-prediction
datasets:
- custom
metrics:
- accuracy
- f1-score
- roc-auc
model-index:
- name: Gold Price Direction Predictor
  results:
  - task:
      type: classification
      name: Binary Classification
    dataset:
      type: custom
      name: Antam Gold Prices
    metrics:
    - type: accuracy
      value: 0.55  # Approximate from training
      name: Accuracy
    - type: f1
      value: 0.56  # Approximate
      name: F1 Score
    - type: roc_auc
      value: 0.58  # Approximate
      name: ROC AUC
---

# Gold Price Direction Predictor

This model predicts the next-day direction of gold prices (up or down) based on historical Antam gold price data and technical indicators.

## Model Description

- **Model Type**: Binary Classification (Gradient Boosting / XGBoost / LightGBM)
- **Task**: Predict whether gold price will go up or down the next day
- **Input**: Feature vector with technical indicators (returns, lags, RSI, MACD, Bollinger Bands, etc.)
- **Output**: Probability of price going up (0-1), thresholded at optimized value for prediction

## Intended Uses & Limitations

### Intended Uses
- Financial analysis and decision support
- Educational purposes for machine learning in finance
- Research on gold price prediction

### Limitations
- Trained on historical Antam gold prices only
- May not generalize to other markets or time periods
- Prediction accuracy is around 55-60% (better than random but not perfect)
- Requires up-to-date feature computation for real-time use

## How to Use

### Loading the Model

```python
from huggingface_hub import hf_hub_download
from joblib import load

# Download model
model_path = hf_hub_download("theonegareth/GoldPricePredictor", "gold_direction_model.joblib")
model = load(model_path)
```

### Making Predictions

The model expects a pandas DataFrame with the same feature columns used in training.

```python
import pandas as pd

# Example feature vector (you need to compute these from your data)
features = pd.DataFrame({
    'ret': [0.01],
    'log_ret': [0.00995],
    'ret_lag_1': [0.005],
    # ... all required features
})

# Predict probability of going up
proba_up = model.predict_proba(features)[:, 1]
prediction = (proba_up >= 0.52).astype(int)  # Using optimized threshold
```

### Feature Engineering

To use this model, you need to compute the same features from your gold price data:

- Daily returns and log returns
- Lagged returns (1-5 days)
- Rolling means and stds (3,5,10,20 days)
- RSI (14-day)
- MACD and signal
- Bollinger Bands
- Day of week and month

See the training notebooks for the complete `add_features_adaptive` function.

## Training Data

- Source: Antam historical gold prices (Indonesian market)
- Period: [Insert date range from your data]
- Features: 25+ technical indicators
- Target: Next-day price direction (up=1, down=0)

## Performance

Based on holdout testing:
- Accuracy: ~55%
- F1 Score: ~56%
- ROC AUC: ~58%

See the confusion matrix, ROC curve, and feature importance plots in the repository.

## Training Procedure

1. Data preprocessing and feature engineering
2. Time-series split for cross-validation
3. Hyperparameter tuning with RandomizedSearchCV
4. Model selection based on F1 score
5. Threshold optimization for final predictions

Models compared: Gradient Boosting, XGBoost, LightGBM

## Contact

For questions or issues, please open an issue on this repository.

## License

MIT License