What is it about?
We test whether a small design choice in AI—the “activation function” inside a neural network—changes how well it predicts cryptocurrency prices. Using data from Jan 2016–Jun 2022, including COVID-19 and the start of the Russia–Ukraine war, we compare several models and show which setups track price moves more reliably.
Featured Image
Why is it important?
Crypto markets are noisy, and forecasting errors can be costly. This study is timely because it tests model robustness during “stress” periods, not just calm times. Uniquely, it isolates the impact of activation functions (ReLU, sigmoid, tanh) across RNN/GRU/LSTM and hybrid models, and shows that tuning this choice especially tanh can materially improve accuracy, with GRU performing best overall.
Perspectives
What stood out to me is how much performance can hinge on something many papers treat as a default setting. Seeing tanh lift weaker models and GRU remain consistently strong across different market conditions reinforced a practical lesson: before chasing ever-more complex architectures, it’s worth stress-testing and carefully tuning the basics especially when the goal is better risk management in turbulent markets.
Tibor Bareith
ELTE Centre for Economic and Regional Studies Hungary
Read the Original
This page is a summary of: Investigating the Role of Activation Functions in Predicting the Price of Cryptocurrencies during Critical Economic Periods, Virtual Economics, December 2024, The London Academy of Science and Business Limited,
DOI: 10.34021/ve.2024.07.04(4).
You can read the full text:
Contributors
The following have contributed to this page
