Stock Market Forecasting with Statistical & Deep Learning Models
Stock prices are notoriously difficult to forecast, but improving prediction accuracy even slightly can help businesses and analysts make more consistent, data-driven decisions.
I built this project to compare traditional statistical models with deep learning architectures on real financial time series data, focusing on how different approaches capture time-dependent patterns and reduce forecasting error.
Using Google’s historical stock dataset, I built a forecasting pipeline that included preprocessing, stationarity checks , model design and evaluation, allowing me to assess where each method performed best and how such techniques can extend to domains like demand prediction, credit risk analysis etc..
-
Designed time-series forecasting pipelines with preprocessing steps including ADF stationarity testing, differencing, and parameter selection via ACF/PACF with AIC/BIC criteria.
-
Benchmarked statistical vs. deep learning models: ARIMA, SARIMA, Prophet, Transformer vs. RNN, LSTM, CNN-LSTM encoder–decoder.
-
Implemented a sliding-window approach for sequence generation, enabling models to learn from past windows to predict the next 30 days of closing prices.
-
Achieved strong predictive performance: LSTM reduced RMSE to <5% of the average closing price, significantly outperforming baseline regression and ARIMA models.
-
Validated stability with R² and error variance analysis, ensuring consistent performance across different forecast horizons.
-
Showed how the forecasting pipeline generalizes beyond finance, applicable to areas like demand forecasting, credit risk modeling, and energy consumption prediction.
Skills
- Time Series Forecasting: ARIMA, SARIMA, Prophet, Transformer
- Deep Learning: RNN, LSTM, CNN-LSTM encoder–decoder
- Techniques: Stationarity testing (ADF), sliding-window sequence generation, RMSE, R²
- Languages & Libraries: Python, Pandas, NumPy, Scikit-learn, TensorFlow/Keras
