Multivariate LSTM forecasting

DeepStock

A stacked LSTM trained on five years of daily Google stock data to forecast prices over a two-month horizon.

The problem

A study in multivariate time-series forecasting: how well a deep recurrent model tracks a real price series when trained on open, high, low, close and volume together.

Approach

• Exploratory analysis and cleaning of daily GOOG data from Yahoo Finance.
• Train on Jan 2019 – Jun 2023, validate on Jul – Dec 2023, test on the first two months of 2024.
• Min-max scaling and sliding input sequences.
• Four stacked LSTM layers of 100 units with dropout, trained with Adam on mean squared error.

Stack

• TensorFlow
• LSTM
• Time series
• Scikit-learn