Training Machine Learning Models to Choose Grid and Martingale Trades

Summary

This article proposes training a machine learning model to choose between buy and sell grids. Instead of labeling examples from the direction of a single position over a future horizon, it simulates a grid across that price window and labels the side using aggregate profit, including triggered orders and trading costs. It also discusses alternative labels based on profit per order or the number of activated orders. Grid spacing and lot multipliers can vary, allowing regular, martingale, or other weighting schemes.

The author describes an exploratory implementation and reports a historical test in which the system worked from late 2016 onward but failed earlier. A more aggressive version is reported as returning 1600% over three years with 40% drawdown, while carrying a hypothetical risk of losing the full deposit. The author offers no explanation for the regime dependence. These results are presented as backtest evidence, not proof of future performance, and aggressive multipliers make the risk substantial.

Key ideas

• Label grid examples using the total outcome of all triggered positions over a chosen future window.
• Grid spacing and lot multipliers define whether the simulated system is regular or martingale-like.
• The article tests profit-based labels and notes alternatives based on order count or profit per order.
• The reported backtest is highly period-dependent and includes a stated possibility of losing the full deposit.
• Results from the historical test do not establish that the pattern will persist.

Tags

This summary was written by Stratmill's research agent from the original; it is not a copy of the source.