Intermarket Trading Strategies

Chart Formations

Systems

Custom Indicators

NEURAL NET SYSTEMS

You can see in the table below the true out-of-sample performance of all 5 neural net systems on data that were not available when I wrote the book.
All systems outperformed the buy and hold method by a wide margin and with considerable less risk. In fact the combined system produced £73,140 of net profit by trading only one FTSE contract (£10 per point) which was 11 times the buy and hold profit (£6,460) with 4 times less risk (drawdown).
For the sake of consistence with the original tests in the book I used the same input parameters and optimization period (4/27/1993-1/1/2003). I had, however, to retrain all models as I upgraded to Neuroshell's latest version 6.1 beta Pro. I also increased the paper trading period until 8/1/2008 and of course the out-of-sample period from 8/1/2008-2/25/2011. I used the Gene Hunter optimization method and the objective that was used to optimize the network structure was to maximize the return on account equity curve correlation (recommended by Neuroshell).
The best performing systems during the most recent time period were the ROC (relative) and the Combined systems. The ROC system was the worst performer producing the lowest net profit with the highest drawdown. In addition I had to retrain the model a number of times in order to get these poor results, which was not the case with the other systems. To refresh your memory the ROC system was based on the % rate of change of the Intermarket securities whereas the ROC relative on the relative % rate of change between the FTSE and the Intermarket securities.
In the second case the inputs were more specific and hence the better results and less training time.
The Combined system used the outputs of the first three neural network tests. On optimization it rejected the Disparity and used only the first two for Long entries. The reverse was true for short entries (it used only the Disparity).
The Hybrid system used the outputs of three network systems plus two additional conventional indicators: The stochastic and the exponential moving average for long entries and only the exponential moving average for short orders. The model was configured to use a minimum of 3 out of the 5 in total rules for long orders, 2 out of 5 for sell orders,3 out of 4 for short and 2 out of 4 for buytocover orders. The stochastic and exponential averages were also optimized.
The last four rows of the table below indicate the contribution of each Intermarket security as optimized by the genetic algorithm.
It is worthwhile to note that all 3 neural net systems preferred to use the S&P Bank Index the most and only one system used either the XOI or the CAC. This indicates a change in correlations during the more recent paper trading period that was used to test the system.
Finally let's not forget my original objective in Chapter 14 of my book which was to compare conventional with neural net systems. In this case the conventional system produced only £18,000 of profits with only 4 trades during the 2.5 year test period compared with an average of £52,000 of the neural systems. In addition the two best neural net systems outperformed the conventional system on Reward/Risk basis. It is therefore worthwhile to add a good Neural Net program in your trading tools.

Performance of the Neural Intermarket strategies tested on one FTSE contract from 8/1/08 to 2/24/11 (US format) based on it's relationship with the French CAC40, the Amex Oil Index (XOI) and the S&P Bank Index (BIX). The COMBINED strategy used the outputs of the first three neural network tests and the last strategy was a hybrid neural and conventional rule-based system.
The last four rows indicate the contribution of each Intermarket security as optimized by the genetic algorithm.

©2011 Markos Katsanos - All rights reserved

aaaaaaaaaaaaiii