The content discusses the comparison of three popular neural architecture search strategies - Bayesian optimization, hyperband method, and reinforcement learning - in the context of financial time series forecasting. It explores the challenges, data preparation, architecture types (feedforward networks, CNNs, RNNs), search spaces, and performance estimation strategies. Results show LSTM and 1D CNN outperforming FFNN with hyperband and Bayesian optimization yielding better results than reinforcement learning. The study highlights the difficulties in predicting financial markets and the impact of random seed variance on model performance.
Til et andet sprog
fra kildeindhold
arxiv.org
Vigtigste indsigter udtrukket fra
by Denis Levche... kl. arxiv.org 03-25-2024
https://arxiv.org/pdf/2403.14695.pdfDybere Forespørgsler