Forex Automaton

For the first time I am able to discuss performance of a ForexAutomaton system without the "benefit of hindsight" caveat: the results for January have been obtained in real time, hence no hindsight.

Executive summary

In the absence of a capital allocation and trade-idea discrimination system, the main figure of merit is the Pearson correlation coefficient between real and forecast logarithmic returns in day high, low and close. In January, these figures of merit appear to remain in line with longer range historical performance. A type of a trade strategy specifically designed to take advantage of the superior forecasting quality for daily high and low, while minimizing exposure to the forecast for close, is discussed, with an attempt to evaluate performance using the recorded output of the system in January.

The document consists of a summary section followed by 14 subsections, dedicated to the individual exchanged rates tracked by the system. Those contain color-coded charts of the performance and more detailed discussion focused on the specific currency combinations.

Changes in the algorithm

On January 14, the system was upgraded to version 0.5. The upgrade is expected to improve prediction quality for daily low and high, even though the effect may not be observable immediately.

Performance tables

Day close strategies

These are trade-at-the-market strategies pursuing trades from day close to day close, which in case of prolonged moves may last for several business days. Not realizing the potential of day's high and low forecasting, this is the type of strategy I was having in mind while working on the system optimization just a few months ago. A particular way of colored bar charts was developed specifically for that (see examples for AUD/USD, USD/CAD, USD/CHF, GBP/USD, USD/JPY, and EUR/USD), and I am going to keep using that here.

Day range strategies

There is one thing the data are quite unambiguous about: the forecasting quality for day's extremes, low and high, is greatly superior to that of day close. Instead of trying to inject that information back into the algorithm to improve the quality of close, here is a better way: one can trade with a limit take-profit order and stop-loss. The issue of correctness of the forecast for close is, in effect, bypassed since a large fraction of trades will be closed by either of these orders even before the next close is reached. The time scale of a trade becomes effectively intra-day. Naturally, the high quality knowledge on the direction of next low and high, with respect to the previous day, enables betting on the coming day's dynamics with respect to the previous day's low and high. Specifically, having a bullish forecast for the coming day's high, low and close (we predict returns, that is changes in these quantities), one can bet with relative confidence that the previous day's high will be exceeded while the previous day's low will never be touched. The way to use this knowledge is to open a long position with the previous day's low as a stop-loss and a the previous day's high as an automated take-profit target. Situation for the short trade is a mirror-reflection of this.

In the following section, an implementation of this strategy is analyzed in a very simple-minded way by counting would-be profitable and would-be lossy trades. Such a simple tally is indeed relevant to the system profitability, assuming that the odds for such a trade to be a win or a loss are uncorrelated with how many pips it contains. Whether this is a good assumption remains to be seen. One may argue that trades with a small distance to a profit target and a long distance to the protective stop have a better chance of turning into profits than into a loss; on the other hand, the size of the possible profit is smaller than the size of the possible loss. What is the interplay of the two effects? How does the bias introduced by the fact that the trades are forecast to reach profit and not to reach loss by the system under study, influence the answer to the previous question? These are open questions at the moment.

A simple tally of profitable and lossy day-range trades, following the strategy just described, goes as follows: 141 winning trade, 94 losing trades.

Charts in the following sections give the reader, among other things, an opportunity to see how the variations in the seemingly abstract quantity such as the Pearson correlation coefficient correspond to variations in the potential trading performance.