Overview:
This project delivers a full end-to-end time series analysis and predictive modeling pipeline to forecast USD/TRY exchange rates using real historical financial data from Yahoo Finance. It applies rigorous data diagnostics, statistical testing, and model evaluation techniques to build accurate forecasts using ARIMA and SARIMA models.
Key Features:
- Exploratory Data Analysis: Includes STL decomposition, volatility trends, moving averages, and daily fluctuation analysis.
- Stationarity & Seasonality Tests: ADF, KPSS, Hodrick-Prescott filter, and Kruskal-Wallis tests ensure valid model assumptions.
- Predictive Models: Implements ARIMA and SARIMA with hyperparameter tuning, seasonal order selection, and cross-validation.
- Model Performance Comparison: Evaluates MAE, RMSE, and MAPE. Performs residual diagnostics with the Ljung-Box test.
- Backtesting Framework: Rolling-window backtesting assesses long-term performance and forecasting stability.
- Tech Stack: Python · Pandas · Statsmodels · Matplotlib · TimeSeriesSplit · ARIMA · SARIMA
Dataset:
- Source: Yahoo Finance – USD/TRY
- Frequency: Daily
- Features: Open, Close, High, Low, Adjusted Close, Volume
Visual Highlights (suggested images for portfolio):
- STL decomposition of the exchange rate series
- ACF/PACF plots for ARIMA tuning
- Forecast comparison graphs (ARIMA vs SARIMA)
- Moving averages and volatility trends
- Ljung-Box residual diagnostics plots
Insights:
- ARIMA performed best for stable market conditions.
- SARIMA captured cyclical and seasonal behaviors but required careful tuning to avoid overfitting.
- The model setup serves as a blueprint for applying statistical forecasting to real-world currency data.
Use Cases:
- Currency exchange rate forecasting
- Financial planning tools
- Macro-economic scenario modeling
- Integration with trading automation systems