Skip to content
All library documents

Using GAN-Generated Price Paths to Test Trading Strategies

Article arXiv papers · Author: Ao Sun et al.
Read the original

Summary

The document proposes using generative adversarial networks with LSTMs to create synthetic historical price paths for evaluating trading strategies. Its premise is that a strategy selected for strong in-sample performance may be overfit, and should therefore also perform well on paths drawn from a model of the historical data distribution.

The proposed workflow fits the time-series distribution with the GAN-LSTM model, then backtests candidate strategies on generated paths as an additional check against overfitting. The text presents this as a method proposal, but gives no experimental results, implementation details, or comparison with other validation techniques. It is a translated summary of a thesis originally written in Chinese in 2018; the document itself cautions that some claims may be outdated.

Key ideas

  • Strong in-sample performance alone can select an overfit trading strategy.
  • A GAN combined with an LSTM is proposed to learn the historical time-series distribution.
  • Synthetic paths from the fitted model provide additional backtesting data for strategy evaluation.
  • The document provides no reported results or detailed validation of the approach.

Tags

Machine learningBacktestingStatistics
Full text
# Backtesting Trading Strategies with GAN To Avoid Overfitting


# Backtesting Trading Strategies with GAN To Avoid Overfitting









Many works have shown the overfitting hazard of selecting a trading strategy based only on good IS (in sample) performance. But most of them have merely shown such phenomena exist without offering ways to avoid them. We propose an approach to avoid overfitting: A good (meaning non-overfitting) trading strategy should still work well on paths generated in accordance with the distribution of the historical data. We use GAN with LSTM to learn or fit the distribution of the historical time series . Then trading strategies are backtested by the paths generated by GAN to avoid overfitting.(This paper is an tanslated English version of a thesis (10.6342/NTU201801645) which was originally written in Chinese in 2018, where some statements and claims are outdated in 2022)

Shown in full with attribution under the source's licence. Licence: abstract CC0

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