Brownian ReLU(Br-ReLU): A New Activation Function for a Long-Short Term Memory (LSTM) Network

By George Awiakye-Marfo, Elijah Agbosu, Victoria Mawuena Barns, Samuel Asante Gyamerah

Rating

981
Battle Count: 194

Relevance

6/10
The paper is moderately relevant to quantitative trading. It addresses financial time series forecasting (stock prices, S&P 500) which is directly applicable to trading strategy development. The improved MSE and R² metrics suggest better predictive accuracy for price movements. However, the paper does not address trading-specific concerns such as transaction costs, slippage, portfolio construction, or backtesting frameworks. The classification component (loan default) is more relevant to credit risk than trading. The stochastic nature of the activation function aligns well with modeling market uncertainty, but practical deployment considerations (latency, computational overhead of Monte Carlo simulation) are not discussed.

Implementation Complexity

5/10
Moderate complexity. The core idea is straightforward (replacing negative ReLU output with Monte Carlo Brownian path average), but implementation requires: (1) generating M independent N(0,|x|) samples per negative input per forward pass, (2) computing gradients with respect to the learnable α parameter through the stochastic path, (3) integrating this into existing LSTM gating mechanisms (cell state and hidden state), and (4) managing the computational overhead of Monte Carlo simulation during training. The algorithm is clearly specified in pseudocode, and standard deep learning frameworks (PyTorch, TensorFlow) can handle the stochastic sampling. The main challenge is the computational cost scaling with M and batch size.

Reproducibility

3/5
The paper provides detailed mathematical formulations, gradient derivations, and a complete training algorithm (Algorithm 1) with pseudocode. However, no GitHub repository or code link is provided. The Monte Carlo simulation procedure is well-described, and the datasets (Apple, GCB, S&P 500, LendingClub) are publicly available. The lack of hyperparameter details (learning rate, batch size, network architecture specifics like number of layers/units) and no code release limits full reproducibility.

About this paper

Methodology: Brownian ReLU (Br-ReLU). Problem types: Time Series Forecasting, Classification, Regression, Imbalanced Learning.

The interactive Everscope explorer (charts, battles, favorites) loads below.