A One-Day Return Neural Network for Directional Trading
Summary
This document presents a directional trading strategy that applies a fixed artificial neural network to a price-change input. It uses the prior day's percentage change as its single input, passes that value through two tanh hidden layers with 5 and 33 nodes, and produces a prediction from a linear output node. The description says predictions above a threshold trigger long signals and predictions below its negative trigger short signals. The supplied implementation instead carries forward a binary long-or-short state based on the output's sign, so the threshold parameter's stated role is unclear.
The notes characterize the approach as trend following and suggest that a nonlinear model might capture patterns missed by simpler rules. The provided backtest settings specify Bitcoin futures and a test window, but give no performance statistics or evidence of out-of-sample validation. The document cautions that the model may overfit, can behave poorly in ranging markets, and is difficult to interpret. It suggests adding inputs such as volume, testing different architectures, using multiple horizons or ensembles, and incorporating risk controls; these remain proposals rather than demonstrated improvements.
Key ideas
- The network uses the previous day's percentage price change to forecast a subsequent change.
- Its described architecture has two tanh hidden layers and a linear output.
- The notes define threshold-based long and short signals, while the provided implementation appears to switch direction based on output sign.
- The listed Bitcoin futures backtest settings provide no reported performance or validation results.
- Overfitting, limited features, ranging conditions, and model opacity are identified as risks.
Tags
This summary was written by Stratmill's research agent from the original; it is not a copy of the source.