NVIDIA stock price forecasting

Problem

Predict NVIDIA's next-day closing price from its recent price history, and compare how a convolutional model handles that against a recurrent one. Graduate term project, İzmir Kâtip Çelebi Üniversitesi, Software Engineering, June 2025.

Approach

Used 20 years of daily closing prices (2004-2023), built lag features from the prior three days, scaled everything with MinMaxScaler, and split chronologically (80/20, no shuffling, since shuffling a time series leaks future information into training). Trained a 1D CNN and an LSTM on the identical inputs so the comparison isolates architecture, not data handling.

Result

Both architectures trained successfully on the setup. The more useful finding is a limitation of the design itself: a three-step lag window gives the LSTM very little sequence to actually model, it's closer to a windowed regression problem wearing a recurrent architecture than a case where the LSTM's strengths (long-range dependencies) get to matter.

What I'd do differently

Extend the lookback window well beyond three days, add technical indicators and volume rather than closing price alone, and benchmark against a naive baseline ("tomorrow equals today") to see whether either model actually beats the trivial answer. At a one-day horizon, that comparison is the real test, and this version of the project doesn't include it.