Forex Automaton

This USD/CAD visual back-testing analysis continues the series which began with EUR/USD. Simulated track records of six best Step Two algorithmic traders are studied graphically. For a more numbers-oriented approach to performance, see the article explaining the trading system optimization process which led to the selection of these six robots.

Table 1. Various modes of the trading system operation. The present report deals with performance of selected Step Two algorithmic traders.

A run of the back-testing program created simulated track-records for just the six independent "virtual traders" (forex robots), each of them being an incarnation of the same algorithm, differing by the setting of the adjustable knobs. The robots had been selected on the basis of Step One vs Step Two performance comparison, and are believed to balance risk with return well while taking advantage of the multi-market capital allocation available in the Step Two algorithm. (The article linked above gives details of the traders' input parameters and the performance figures of merit). The algorithm learns continuously, using only data from the past of the time series. Therefore, every trading step, even during the simulated trading, tests applicability of the past experience to the present moment. In Step Two mode, co-existing trading opportunities in markets other than the market under study may influence which trades are placed. There is a competition between the markets.

Fig.1. USD/CAD bar charts for the period under study. Day scale. The chart includes training time when no trades could be placed.

As in all other studies posted here so far, the trading is performed on the one-decision-a-day time scale, with 1:100 leverage, risking no more than 10% of the trading capital at any point in time. Historical EUR/USD, USD/JPY, GBP/USD, USD/CAD (Fig.1), USD/CHF and AUD/USD day scale data are used, covering the time interval from August 20, 2002 to August 21, 2009, with the actual trading starting in April 2006 when the initial "training" of the system was completed. This report covers the trading in USD/CAD, for all six traders.

2.1 2.2 2.3 2.4 2.5 2.6

Fig.2. Performance track record charts. Day scale. Bars in the charts are color-coded according to the type of trading position held, as explained in the table. Black corresponds to no trading position. The time period, same for all traders, is chosen to best show all trades of all traders.

Fig.3. Performance track record chart for T588, arguably the one of the six traders with the best USD/CAD performance. Day scale. Bars in the charts are color-coded according to the type of trading position held, as explained in the table. Black corresponds to no trading position. The time span is chosen to highlight the time of the most active trading.

Discussion of USD/CAD performance

As was already noticed for EUR/USD, USD/JPY, GBP/USD, and USD/CHF, the selected robots trade only during the period of highest volatility. No trades in USD/CAD have been placed in 2006 and 2007. Selected on the basis of portfolio success, the six robots under study show no particular success trading USD/CAD. As with other robots, it's seen very clearly that this is not a trend following system. While it is seen to place trades on break-outs and break-downs, it does not shy away from changing its opinion about the market direction instantly if it considers such a change warranted -- just look at the October 2008 break-down in Fig.3 and what happens the next week.

I dislike the fact that the system does not trade unless the volatility is extreme, since now there seems to be a risk of launching a system which will never place a trade if the volatility storm brought about by the financial panic of 2007-2008 finally abates. This is a topic of special discussion, but the core problem is now easy to identify: unlike the stop-loss parameter, the entry and exit parameters are formulated in a way which is not adjusted for volatility. In other words, the threshold is applied to a quantity (relative expected price change) which is large when volatility is large and small when the volatility is small. That this quanity is a relative price change does not change the fact that the threshold effectively favors high volatility. The reason I end up with strategies trading only high volatility is because I try to minimize the risk, and the risk is minimized by minimizing the number of trades, by being selective about them.

This problem is quite different from the problem of obtaining the market direction forecasts on which the trades are based: it has to do with what fraction of the trade ideas that get accepted, not with the composition of the pool of trade ideas subject to this filter. This problem is not too difficult to fix by changing the way selectivity threshold is calculated, e.g. by including a volatility estimate into the threshold.

Bookmark with: