Neural Network Trading: Overfitting, Money Management, and Market Change
Summary
In this interview, trader Leonid Velichkovsky reflects on using neural networks in automated trading and challenges the idea that they produce exceptional profits by themselves. He describes them as trading systems whose adaptive capacity can help detect patterns, but also makes them prone to fitting historical data that may not persist as market conditions change. He advises judging training by out-of-sample or live performance rather than by training error or historical profit alone.
The interview discusses early stopping and shorter training windows as ways to limit overfitting, while noting that neither has a universally reliable setting. Velichkovsky gives a personal rule of thumb of using 500 to 2,000 bars, depending on timeframe and market state. He also emphasizes that money management matters, especially the temptation to take excessive risk when trading small accounts. These are experience-based views from a 2010 championship interview, not controlled evidence; the suggested training range and preference for simple systems should be treated as context-dependent guidance.
Key ideas
- A neural network's ability to fit many patterns can also make it overfit historical market data.
- Training error and in-sample profits alone do not show that a system will generalize.
- Early stopping and training-window selection require judgment and offer no guaranteed solution.
- The interview suggests 500 to 2,000 bars as a personal, context-dependent training range.
- Aggressive money management can overwhelm a system's modeling strengths and risk the account.
Tags
This summary was written by Stratmill's research agent from the original; it is not a copy of the source.