TABLE 11.3 Performance comparison between models.

Performance comparison between the O&A class RTT, RTM trading models, and a (benchmark) 20-day moving average model. The displayed performance measures are the annualized total return R, the risk-sensitive performance measure Xeff< maximum drawdown D, the profit-loss ratio P/L and the dealing frequency F. These performance measures are explained in Sections 11.2.2 and 11.3.1.

TABLE 11.4 Performance comparison between markets.

The average risk-adjusted return Xeg for the different markets is shown as a percentage the average dealing frequency F is given in number of deals per week. The markets are listed in the order of their opening times in GMT.

USD-DEM, USD-JPY, GBP-USD, USD-CHF and DEM-JPY, and year period from December 1986 to March 1993 for USD-FRF, USD-NLG, and USD-ITL.

Table 11.3 shows the comparative performance of the two types of models (RTT and RTM) together with the performance of a simple 20-day moving average model tested with the same high frequency data and the same environment. All models produce a significant profit even when transaction costs are fully accounted for. However, they differ both in the size of the average profit and in the risk of temporary losses. This was formulated as the first conjecture in the introduction. These results are a good illustration of the possibility of having diversified strategies that are all profitable but correspond to different risk profiles.

Realistic trading models should be configured for traders located in particular geographical locations. Our high-frequency data give us the flexibility of configuring different opening hours for different markets. In Table 11.3, the models were computed within the market constraints of Zurich. Now we want to show how the effective return varies if the market constraints are changed. Six other markets are tested: Frankfurt, London, New York, Singapore, Tokyo, and Vienna. Table 11.4 shows the different parameters related to the active times of these markets.

The same eight FX rates used for the performance comparison in Table 11.3 were tested here. In Table 11.4, we present the average of Xeg over the eight FX rates for the seven markets and the corresponding mean dealing frequency. The bad results for Tokyo and Singapore are not surprising because these markets are the least liquid. Good results in these markets are only obtained for USD-JPY and DEM-JPY. For the other five markets, Xejr generally does not vary much (within 1 to 2%), but it clearly peaks on the most active market (London) although the models were optimized for the Zurich market. This presents the first empirical evidence for our third conjecture that within the active times, the performance is not very sensitive to certain changing conditions.

TABLE 11.5 The best Xeg as a function of opening hours.

The best Xeff, in percent, as a function of the number of daily business hours and the opening and closing times in MET. The sixth column shows the Xeg reached when the models are allowed to trade 24 hr. The last column shows the hour that produces the best Xeg when only I hour per day is allowed for trading.

At the beginning of this Section 11.6.1, we introduced two conjectures as subjects of research: it is not favorable to extend the dealing period to more than the normal business hours or even to 24-hr trading for our model types (conjecture 2); and the most profitable dealing periods should be the most active and liquid ones (conjecture 3). To test these conjectures, two main questions were asked: Is there an optimal daily business interval and do these optimal opening and closing hours differ for different rates? We present here the results of a study where we vary both the length and the starting point of the daily opening period. The two real-time model classes were tested with daily working intervals of 1, 8,9, 10,11, 12, 13, and 24 hr, shifting the opening time in 30-min steps from 0:00 to 24:00.

Table 11.5 shows the best Xeff values together with their corresponding working hours in Middle European Time (MET) for all rates and trading models used in this study. The models were optimized in-sample on 9 j hours from 8:00 to 17:30. Some first remarks can be made by looking at the results: shorter time intervals (8-10 hr) are generally preferred to longer ones (11-13 hr), thus confirming conjecture 2. There is not much profit in long working time intervals; these only tend to increase the number of bad deals because the indicators are more sensitive to