Forex Automaton

During the month of February, the second month of real-time documented performance, the system kept running on complete autopilot, with no code upgrades or parameter changes. This document consists of a summary section followed by 14 subsections, dedicated to the individual exchange rates tracked by the system. Those contain color-coded charts of the performance and details pertinent to the specific currency pairs. Usage strategies and effects of various approaches to selecting the forecasts to trade are discussed. For comparison with the previous month, you may want to take a look at January review.

No changes in the algorithm

No code upgrades or parameter changes took place during the month.

Figures of merit

The figure of merit used, Pearson correlation coefficient between the forecast and real logarithmic returns on a day scale, is a measure of forecasting quality.

By construction, the Pearson correlation coefficient is a quantity bounded between -1 (the forecast move and reality are total opposites) and 1 (the forecast is perfect). A success or lack thereof on every trading day makes a contribution to this quantity.

In order to make a large positive contribution, one needs a coincidence of a large move in a currency pair with a large forecast move in the same direction. Since a hypothetic rational operator of the system will not pursue small forecast moves, understanding this to be a noisy system, a forecast with large magnitude is more likely to result in a successful trade.

By the same property of a product, an impact on the Pearson correlation coefficient of even a large forecast move in a wrong direction can be moderated if the actual price move turned out to be negligible -- again what we want, since in this case the effect of the system's imperfection on the operator's portfolio is likely to be similarly insignificant.

Likewise, an impact of a wrong forecast move of negligible magnitude will be negligible both for the Pearson correlation and for the operator's portfolio since in this case, the operator is likely to ignore the trading idea.

Finally, the worst case is the one when the system predicts a large movement in the currency pair and a large movement does materialize -- but in the opposite direction. In this case, a large negative contribution to Pearson is recorded.

Why use Pearson correlation coefficient instead of running a model portfolio? This independent figure of merit, characterizing the forecasting component regardless of the money management strategy used, allows one to split the complex problem of trading system development into independent tractable pieces, and solve each piece of the problem separately, eventually combining a winning forecasting system with a winning money management one, having an independent quality assurance process for each.

Performance tables

As started in the previous report, I take last day of the month's forecast as the basis for the month's performance. In the tables, data for the "month" column are taken from the "month" column of the last day's forecast, which reports 24-trading-day running average of the Pearson correlation figure of merit. End of February 2010 data are taken from the February 27 update. End of January 2010 data are taken from the January 30th update.

The month's effect on the performance since inception was, as Table 1 indicates, positive for GBP/USD, USD/CAD, USD/JPY, AUD/JPY, CHF/JPY, EUR/AUD, EUR/CHF, EUR/GBP, EUR/JPY, and GBP/CHF (10 markets in total) -- their performance since inception improved, and in some cases quite significantly so. Outstanding results in forecasting quality for day close, as measured by the Pearson correlation coefficient between logarithmic returns and their forecasts, have been obtained for AUD/JPY and GBP/CHF. Overall, the mean over markets for the Pearson correlation coefficient was 0.107 +- 0.037 where the latter number gives a (Gaussian) estimate of the precision of the mean.

For day low and high, the result of the v0.5 upgrade performed in January, seems to be felt -- the Pearson correlation coefficients went up for both low and high.

Usage strategies

January performance report discussed two hypothetic strategies. One strategy, called Day Range, would consist in submitting market orders as soon as possible after the release of the forecast, buying and selling forex pairs in the direction of the forecast day trend for close, provided that the predicted change for the daily high and low would be in the direction of that trend. That is, selling not unless lower low, lower high, lower close are expected; buying not unless higher low, higher high, higher close are expected. Each such market order would be accompanied with a stop loss and and a take profit target. These would be set on the previous day's low and high, depending on the direction of the trade. That is, if the order is a sell, previous day's high would serve as a stop-loss and previous day's low would be the profit target; the roles would change for a buy order.

A subsequent quantitative study revealed that the Day Range strategy would not be successful unless more strict and specific conditions are imposed when selecting potential trades for execution.

Another strategy, referred to as Close-to-Close, would be obtained by removing profit target from the Day Range strategy. It was analyzed quantitatively in pure form and with a stricter requirement of a large forecast magnitude. In both cases, it was found to be superior to the Day Range strategy. As a result, there will be no red-green-magenta-cyan Day Range strategy plots in this report.

These two strategies are the easiest to analyze and lend themselves to full automation. Users of the system eager to get more involved may want to go for a Close-to-High or Close-to-Low strategy which would involve monitoring of the market on time scales finer than once a day, having Close-to-Close as a fall-back safety net if no attractive opportunities to close the deals with profit better than expected from Close-to-Close are found during the day. However, such an algorithmically-aided, as opposed to fully algorithmic, trading strategy, in theory has the potential of delivering results worse than the pure system would have delivered, especially if the operator's intervention tends to take quick limited profits and let the losses run -- or if it tends to take losses while letting profits turn into losses.

Effects of trade idea selection

The following 14 subsections are dedicated to the specific currency pairs, useing a particular method of charting to illustrade performance. Three charts will be presented for each. The first chart will show all trade ideas. This corresponds to trading in every pair every day.

The second chart will highlight the days when forecasts for all three components of a day's candlestick were pointing in the same direction. I call this level zero (L0) requirement or L0 trigger.

Likewise, L1 trigger combines L0 with a stricter requirement of a sufficiently large forecast magnitude.

Table 0.4 shows that while L0 reduces the total number of trades, and L1 reduces it even more, the quality of trade selection, at least as judged by the ratio of wins to losses, grows quite impressively as the selection becomes more and more strict. Note that such a simple counting approach leaves the question of the relative impact of these wins and losses open. Therefore it can only serve as a supplement to more quantitative studies.