I am sure this question will come up. The neural networks are trained on the data up until 6/2009. The last year 6-2009 till now 6-2010 is completely out of sample data.
The two methods I've seen for "proving" something isn't random are: (a) Throw random data at the trained neural net. Is the neural net able to distinguish random data from real data? (b) Use a random method like flipping a coin, tossing dice, or pseudorandom numbers to pick trades based on the training data. If the neural net outperforms random, it is "better" than random. In reality remember you need to hedge or manage risk in some way. And, it's easier to manage risk with a simple system vs one that's very complex.
Now this is actually getting very interesting. The statistics are incredible at this point... I now have 18 neural networks included. The strategy is still long only. I am starting to run out of ideas for new neural networks on the long side. Soon will be time to start working on the short side. PL: +$69,845 (Per Contract) Profit Factor: 2.81 Sharp: 1.05 Max DD: -$2,430 Trades: 494 Win %: 61 Largest Win: $2790 Largest Loss: -$815 At some point I plan to go back through these neural networks and see if any 1 in particular caused the big jump in DD and largest loss seen here..
Done for the night. I now have 29 neural networks. 18 long and 11 short. Results look good. Although historical results are only part of the puzzle. Will have to run for a while and see what happens. PL: +$120,780 (per contract) Profit Factor: 2.94 Sharp: .95 Trades: 826 Win %: 58% Expectancy: +$146 per trade Max DD; -$1590 The P&L is pretty impressive. Been designing strategies based off the YM for a long time. Don't believe I have had a strategy perform that well for a single contract. I will look to try and add a few more neural networks, however the easy money is now gone. To get a gain much higher than this, will be tough. So in 24 hours we have arrived at a very profitable strategy. The neural nets were trained until 6-2009. From then until 6-13-2010 (1 week ago) this is how the strategy performed: PL: +$18,775 Profit Factor: 2.19 Sharp: 1.22 Win %: 57% The strategy's key metrics have stayed within the expected range. Volatility is lower now than in the training period which accounts for the lower rate of return. 18k in a year for a single contract is not that bad however. Question is, what will happen in the coming weeks, months? Anyone care to make a guess?
good luck , I'd like to know the results , I suggest if you realize half this , your doing great , being a bit of an old cynic i'd put your chances at 50 ;50, which is pretty high for this old goat .
I have 4 distinctive groups of neural networks that make up this strategy. I wanted to see how "robust" the networks were so I ran each group of networks against the 30 stocks in the DOW. Two of the neural network groups were profitable on at least half of the DOW components. Two were only profitable on 3 of the DOW stocks. I do not expect these neural networks to be able to perform great on symbols outside of what it was trained on. However, it was nice to see 2 of the groups showing some degree of being robust. The 2 that performed very poorly, I will examine closer to see which neural networks in the groups hurt performance the most. Before getting to this step however, I need to train more networks. Would still like to get up to about 50 if possible. Once I have 50+ I can examine all of them individually and possibly eliminate ones that do not contribute much and are considered "not robust" when looking at other symbols. Profitability on even half of the symbols is robust enough for me considering they were not trained on that type of data and individual stocks normally react to the same pattern differently depending on who the market participants are.
Eliminate based on performance on the training data or eliminate based on performance on the test data?
Eliminate based on the training data. Right now I am finding that as I add new neural networks the profitability barely changes. A few hundred rarely a few thousand anymore... Yet the neural network by itself in many cases is +5k or more. This leads me to believe that many of the setups the new neural networks are seeing have already been represented by other neural networks. This means some of these other neural networks may no longer be needed or contributing much. If several new neural networks are able to represent the patterns an older one used to represent but with a much higher win%,profit factor, or lower draw down... Then I rather use the new neural networks and remove the older one which makes the overall system batter. The problem is, with so many neural networks now its hard to really tell what is being added.. what is overlapping etc... So it will be a lot of work and tedious to do this.. So my first pass through will be to run each neural network by itself and note its key metrics.. dd, pf, profit.. win % etc... Then i will add each of the neural networks back to the system 1 by 1 starting with the "BEST" first... Therefore by time I get to the ones with poorer metrics if I notice it only adds say a few hundred additional dollars to the system but brings with it large draw downs.. then I can leave that system out.. The other idea I will explore is testing each neural network interdependently against the dow 30 stocks to see how it performs. That will also play a role in the order its added back to the system in. At this point I believe the system makes more than enough per contract... I rather lose some of that profitability to lower overall risk. thoughts?
I was just curious what the earnings potential was like. So I assumed that 10k per contract was reasonable. So starting in 2007 with 10k in the account... And capping it at max 50 contracts to avoid liquidity issues and you arrive with 4.5 million today.. Maybe I should start one of those journals 10k to 4.5 million in 3 years or something.