Apppublic

mirzafathir/usd-idr-forecasting

sourceHugging Facemitupdated 3mo agoView on Hugging Face

0likes

App README

📈 USD/IDR Exchange Rate Forecasting

XGBoost + BiLSTM Ensemble Pipeline — Prediksi kurs USD/IDR dengan akurasi tinggi

![Python](https://www.python.org/) ![XGBoost](https://xgboost.ai/) ![Streamlit](https://streamlit.io/) ![FastAPI](https://fastapi.tiangolo.com/) ![License](LICENSE)

Metric	Value
📉 MAE	37.9 IDR
📊 MAPE	0.224%
📈 R²	0.9842
🎯 Directional Accuracy	32.66%
⚡ Inference	< 1ms

</div>

🎮 Demo

Live Dashboard: Jalankan aplikasi Streamlit di atas untuk melihat forecast interaktif, analisis SHAP, dan prediksi 14-hari ke depan dengan confidence interval.

[image]

🔍 Overview

Proyek ini merupakan Final Project Big Data — Semester 6, TC ITS yang membangun pipeline forecasting kurs USD/IDR secara end-to-end:

—Data: Harga harian USD/IDR + makroekonomi (BI Rate, inflasi, emas, minyak) dari 2000–2026
—Target: Log-return USD/IDR (bukan level harga langsung) → menghindari extrapolation cap pada tree-based model
—Forecasting: Rekursif multi-step dengan back-transform: level[t+1] = level[t] × exp(predicted_return)
—Live Data: Harga USD/IDR real-time via Yahoo Finance (cache 5 menit)

🏗️ Arsitektur Pipeline

Raw Data (yfinance / CSV)
        │
        ▼
┌──────────────────┐
│  Feature Eng.    │  27 fitur: lag, rolling, RSI, BI-Fed spread,
│  (src/features)  │  commodity returns, kalender
└────────┬─────────┘
         │
    ┌────┴──────────────────────┐
    │                           │
    ▼                           ▼
┌─────────┐              ┌──────────┐
│ XGBoost │              │  BiLSTM  │  (opsional, butuh GPU/Linux)
│ + Q10   │              │  2-layer │
│ + Q90   │              │  Huber   │
└────┬────┘              └────┬─────┘
     │                        │
     └──────────┬─────────────┘
                ▼
        ┌──────────────┐
        │    Ridge     │  Meta-learner ensemble
        │  Ensemble    │  stacking XGBoost + LSTM
        └──────┬───────┘
               ▼
    Recursive Multi-Step Forecast
    + Quantile CI (Q10 / Q90)

Walk-Forward Validation:

Expanding Window:              Blocked (purge_gap=5):
[====Train====|Val]            [=Train=|gap|=Val=|gap|...]
[======Train======|Val]        [==Train==|gap|=Val=|gap|...]
[========Train========|Val]    [===Train===|gap|=Val=|gap|...]

📊 Features (27 Variabel)

Kategori	Fitur
Lag	`USD_IDR_Lag_1`, `_Lag_3`, `_Lag_5`, `_Lag_7`, `_Lag_14`, `_Lag_30`
Rolling Stats	`Rolling_Mean_7`, `Rolling_Mean_30`, `Rolling_STD_7`, `Rolling_STD_30`
Momentum	`Return_5d`, `Return_10d`, `Return_20d`, `USD_IDR_Diff_1`, `Volatility_10`
Technical	`RSI_14`
Makro	`BI_Rate`, `Inflation`, `BI_Fed_Spread`
Komoditas	`Gold_Price`, `Oil_Price`, `Gold_Return`, `Oil_Return`
Kalender	`DayOfWeek`, `Month`, `Quarter`, `IsMonthEnd`

📈 Model Performance

XGBoost (Test Set: Jan 2020 – Jun 2026)

[image]

Metric	XGBoost	Baseline (Naive)	Improvement
MAE (IDR)	37.9	~280	-86%
MAPE	0.224%	~1.6%	-86%
R²	0.984	~0.0	—
RMSE	51.4	~395	-87%

Walk-Forward Validation

[image]

14-Day Forecast dengan Confidence Interval

[image]

Feature Importance (SHAP-based)

[image]

🚀 Cara Penggunaan

1. Clone & Install

bash

git clone https://huggingface.co/spaces/YOUR_USERNAME/usd-idr-forecasting
cd usd-idr-forecasting
python -m venv venv && source venv/bin/activate
pip install -r requirements.txt

2. Jalankan Dashboard

bash

streamlit run app.py
# → http://localhost:8501

3. Jalankan API

bash

uvicorn api:app --reload --port 8000
# → Swagger: http://localhost:8000/docs

4. Generate Artifacts (tanpa training ulang)

bash

python generate_artifacts.py
# ~10 detik, menghasilkan semua CSV/PNG di output/ dan figure/

5. Training Ulang (opsional)

bash

# Fast mode (~15-20 menit)
python train_fp.py --fast

# Full GridSearchCV (~60-90 menit)
python train_fp.py

🔌 API Endpoint

bash

# Health check
curl http://localhost:8000/

# 14-day forecast
curl -X POST http://localhost:8000/forecast/multistep \
  -H "Content-Type: application/json" \
  -d '{"n_steps": 14}'

# Metrics
curl http://localhost:8000/metrics

Swagger docs: http://localhost:8000/docs

🐳 Deployment

Docker Compose (Streamlit + FastAPI)

bash

docker-compose up --build -d
# Dashboard  → http://localhost:8501
# API        → http://localhost:8000

Streamlit Cloud / HF Spaces

Push repo ke GitHub/HF → auto-deploy via app.py + requirements.txt.

GitHub Actions CI

yaml

# .github/workflows/ci.yml sudah tersedia:
# - Syntax check semua file Python
# - Import check (pastikan semua dep tersedia)
# - Dataset integrity check
# - Docker build

📁 Struktur Proyek

.
├── app.py                    # Streamlit dashboard v3.0
├── api.py                    # FastAPI inference server v3.0
├── train_fp.py               # Pipeline training XGBoost + BiLSTM
├── generate_artifacts.py     # Generate semua output/ tanpa training ulang
├── requirements.txt
├── Makefile
├── docker-compose.yml
├── .streamlit/config.toml
├── .github/workflows/ci.yml
│
├── src/
│   ├── config.py             # Konfigurasi global (dataclass)
│   ├── features.py           # Feature engineering pipeline
│   ├── evaluation.py         # Metrics + walk-forward CV
│   └── inference.py          # Multi-step forecast functions
│
├── output/                   # Model artifacts + CSV results
│   ├── xgboost_model.pkl
│   ├── xgboost_q10.pkl       # Quantile 10% (lower CI)
│   ├── xgboost_q90.pkl       # Quantile 90% (upper CI)
│   ├── model_config.json
│   ├── evaluation_results.csv
│   └── multistep_forecast_14d.csv
│
└── figure/                   # 16 visualisasi PNG
    ├── summary_dashboard.png
    ├── actual_vs_pred_xgboost.png
    ├── feature_importance_xgboost.png
    ├── walk_forward_validation.png
    └── ...

📚 Tech Stack

Layer	Tools
Data	pandas, numpy, yfinance
Features	scikit-learn, statsmodels
Model	XGBoost 3.2, TensorFlow/Keras (BiLSTM)
Explainability	SHAP TreeExplainer
Visualization	Plotly, Matplotlib, Seaborn
API	FastAPI + Pydantic
Dashboard	Streamlit
DevOps	Docker, GitHub Actions

👤 Author

Mirza Fathir R. — Teknik Komputer, Institut Teknologi Sepuluh Nopember (ITS) Final Project Big Data, Semester 6, 2025/2026

<div align="center"> <sub>Built with ❤️ for Big Data Final Project — TC ITS 2025/2026</sub> </div>