The optimization framework is being upgraded.

User Rating: / 1
Written by Forex Automaton   
Tuesday, 26 May 2009 09:41

Our forex trading system optimization procedure is currently undergoing a change. This change is brought about by the recognition of the fact that a simple arithmetic average of monthly returns, a statistic used so far, is biased upwards and thus provides a falsely optimistic estimator of return. The popular Sharpe ratio statistic, if it incorporates such an estimator, is to be avoided or redefined. Here are some outlines of the upgraded approach.

The essential core aspects of the "old" approach, applied so far to AUD/JPY, EUR/USD, USD/JPY, GBP/USD, and AUD/USD are believed to be sound. These are:

  • The algorithm may not trade the data used to train its decision making. The algorithmic learning is continuous and lasts as long as the trading lasts, but is limited to the past data.

  • Stability of profit is of higher priority than the profit figure itself. Some kind of a second-order statistic, such as variance or RMS, is needed to monitor risk. The maximum draw-down proved to be a useful figure of merit; the worst annualized monthly return was used for the purpose.

  • The key concepts of conditional projection distributions and profile histograms, and their software implementation, remain as before.

The distinctive new features of the upgraded approach are:

  • Annualized return for the full period of trading is always used. Previously, due to over-reliance on the arithmetic average of the annualized monthly returns, these were believed to be redundant and not even monitored for some of the data sets.

  • Linear returns are replaced with logarithmic returns. The advantage is that the arithmetic average of such returns, taken on a month-by-month basis, provides an unbiased estimate of the return obtained by comparing just the beginning and the end of the trading period (and in fact, equals that in the case when the start of trading and its end coincide with the month-end performance-monitoring points). Obviously, this property owes to the fact that the logarithm of a product is the sum of the logarithms of the factors.

Bookmark with:    Digg    reddit    Facebook    StumbleUpon    Newsvine
Last Updated ( Monday, 04 January 2010 12:35 )