AI / ML
Stock Price Prediction — Stacked LSTM
A financial forecasting model built to explore how deep learning handles the noise and non-linearity of real stock market data — trained end-to-end with full evaluation metrics.
Project Overview
A time-series forecasting model built using a stacked LSTM architecture on OHLC (Open, High, Low, Close) stock data. The model was trained on historical price sequences normalized with MinMaxScaler and evaluated using RMSE and MAE metrics against a held-out test set.
The Challenge
Stock price data is inherently noisy and non-linear — standard regression approaches fail to capture temporal dependencies. The challenge was building a model that generalizes across market conditions without overfitting to historical trends.
The Approach
I built a stacked LSTM in PyTorch with two recurrent layers and a fully connected output. Data was preprocessed with sliding window sequences (60-day lookback), normalized, and split 80/20 for training and testing. Loss curves and prediction plots were generated to validate model behavior.
Model Performance
A validated forecasting model that captures price trends with quantified error metrics on real market data.
LSTM
Architecture
60d
Lookback Window
RMSE
Primary Metric
PyTorch
Framework
Technology Stack
Modeling
PyTorchStacked LSTMMinMaxScaler
Data
PandasNumPyOHLC Dataset
Evaluation
RMSEMAEMatplotlibSeaborn
Process Gallery
Predicted vs. Actual Price
Training & Validation Loss Curves
OHLC Data Preprocessing
Key Features
A rigorous ML pipeline from raw data to evaluated predictions.
Stacked LSTM Architecture
Two-layer LSTM with dropout regularization — designed to capture multi-scale temporal patterns in price sequences.
Full Evaluation Pipeline
RMSE and MAE computed on the held-out test set, with predicted vs. actual plots generated for visual validation.
Proper Data Preprocessing
Sliding window sequencing, MinMaxScaler normalization, and 80/20 train/test split applied before training.