LSTM–CNN GAN Stock Forecasting

Article BigQuant

Summary

This learning note outlines a stock-price forecasting pipeline that uses an LSTM as a sequence generator and a CNN as a discriminator in a generative adversarial network. It describes a broad feature set: related assets and rates, technical indicators, news sentiment, Fourier-derived trends, and ARIMA forecasts. It also discusses VAE-generated features, XGBoost feature ranking, PCA reduction, and Bayesian optimization plus deep reinforcement learning for tuning model parameters. The LSTM uses rolling windows to predict the next price, while the discriminator evaluates generated patterns against observed data.

The reported example is a Goldman Sachs forecasting exercise using historical data split into training and test periods. The note says forecasts appeared stronger after repeated training and reinforcement learning, but supplies no numerical evaluation or reproducible code, so the claimed quality cannot be independently assessed from this account. Its central assumption is that historical price patterns can recur; regime shifts, seasonality, feature quality, overfitting, and performance on other stocks remain important limitations.

Key ideas

The proposed GAN pairs an LSTM sequence generator with a CNN discriminator for stock-price forecasting.
Features combine market data, technical measures, news sentiment, trend extraction, and statistical forecasts.
VAE features, XGBoost ranking, and PCA are used to construct and reduce the input set.
Bayesian optimization and reinforcement learning are presented as tools for adapting model hyperparameters.
The account reports qualitative improvement on one stock example but provides no metrics for independent evaluation.

Forecasting Goldman Sachs Prices with an LSTM–CNN GAN

Summary

Key ideas

Tags