Jun 2024
As part of my Deep Learning final project, I built an LSTM-based time-series model to predict the next-day closing price of Amazon (AMZN) and Cisco (CSCO) stocks. I prepared the historical closing-price data using a 5-day window to predict the following day, then split the data into 80% training, 10% validation, and 10% test sets. The baseline model used a 50-unit LSTM with ReLU activation followed by a single dense output layer, and was evaluated using RMSE, MAE, and MAPE.
I also experimented with model architecture and training strategies by applying Early Stopping and, for CSCO, adding a second LSTM layer. Early Stopping significantly improved the AMZN model, reducing its test RMSE from 72.2 to 40.01, MAE from 63.58 to 29.18, and MAPE from 3.65% to 1.7%. For CSCO, the baseline model achieved a test RMSE of 0.79, MAE of 0.54, and MAPE of 1.25%, while the two-layer architecture resulted in slightly higher errors.
Here are the visualizations of the predicted price vs the actual price
See the code here: Google Colab
More on my thought process here: Youtube