Summary of the trading system optimization results. Step One.

User Rating: / 2
PoorBest 
Written by Forex Automaton   
Tuesday, 21 July 2009 14:28
Article Index
Summary of the trading system optimization results. Step One.
Stop-loss
How often to place new trades?
Forecasting
The multivariate optimum
Winners. Future plans.
All Pages

I've gone through three rounds of historical data optimization (back-testing) for the major forex pairs, performing simulated trading in each pair independently. This mode of the market analysis, with the various exchange rates being treated by the algorithm in isolation from one another, is not the way the production trading system will operate. However, understanding the markets in isolation and optimizing the trading system in this simpler problem setting is seen as the first step towards optimization of the more complex algorithm, where the amount of information at every point in time will be radically increased by combining analyses of multiple exchange rates within the same algorithm. This study is a cross-check to see how the indvidual optima hang together, trying to define a coherent dry residue from the combined experience of the 150,423 simulated trading histories.

1. The "consensus" optimum.

Why is consistency of optima found for the different forex markets interesting? Ideally, if the problem is treated correctly, the algorithmic optimization must deal with bare essentials of the market predictability, with all the transient and non-essential specifics and particulars of the individual exchange rates being left behind and "abstracted" away. In the opposite case of self-delusion, the trading system, optimized using a certain data set, will not function successfully in real life. Wild variation of solutions (the parametric optima), if found for the different markets, would look suspicious. Lack of applicability of XXX/YYY solution for UUU/VVV would cast doubt on the applicability of today's XXX/YYY solution tomorrow. A necessary (but not sufficient) condition of success therefore seems to be some sort of universality in the solutions found. Defining such universality may not be easy.

The data collected so far (see the latest optimization studies for AUD/USD, EUR/USD, GBP/USD, USD/CAD, USD/CHF, and USD/JPY ) allow me at leat two ways of approaching this question:

  1. Compare the "lanscape" of the optimization problem. Similarility of the landscape entails similarity of the solutions. To a considerable degree, this has been already done as successive optimization reports typically included comparisons with similar reports for other markets.

  2. The "consensus" approach: work out a "consensus" optimum, the one-size-fits-all compromise solution to fit all the different markets. See if it fits, and how good the resulting performance is.

I am not saying that such an optimum is necessarily the way to go -- there may be legitimate differences among the optima that work for the different markets. I am just saying that defining such a common optimum and trying to test how well it fits is a good excercise.

The basic framework remains the same: a run of the program included simulations of trading histories of over 20 thousand independent "virtual traders" (forex robots), each of them being an incarnation of the same algorithm differing by the setting of the adjustable knobs. The algorithm "learns" continually using the past (but not future) data. In total for the six markets researched, there are 150,423 such robot trading histories under study. As in all other published studies so far, the trading is performed on one-decision-a-day time scale, with 1:100 leverage, risking no more than 10% of the trading capital at any point in time. The data used cover the time interval from August 20, 2002 to March 23, 2009, with the actual trading starting in March 2006 (when the initial "training" of the system was completed). The key concepts of conditional projection distributions and profile histograms have been explained before. The trading system control parameters remain as previously defined.

Construcing the consensus optimum in stop-loss parameter Construcing the consensus optimum in entry parameter Construcing the consensus optimum in exit parameter Construcing the consensus optimum in forecasting parameter

Fig.1.1. Comparing the optima, from top to bottom: stop-loss, entry, exit and the forecasting parameter. The edge values are included. The gray shadow band shows the "consensus" optima.

In Fig.1.1, the chosen optima are highlightened by gray bands. The choice of the optimum for the stop-loss is the easiest one to challenge, but the rationale for that choice will be explained below.



Last Updated ( Monday, 04 January 2010 12:30 )