Many of you will laugh at this idea and with good reason. Common sense dictates that this should be a bad idea, and everything you've ever read probably said do not do this. However, I am bored, and there could be some "merit" to this. The basic idea is you develop x number of neural networks. All of them independent of each other and trained on different types of inputs. I am sure many of us have already done something similar. Grab the ones that appear the most curve fit. Such as very few trades (However at least 10 trades), extremely high profit factors etc... Now you run each of the "cherry picked" neural networks on outside data (not from your training set). All of them that performs at least 1 trade and maintained exceptional metrics keep. Now you build a large strategy using all of these neural networks, so that if any of the neural networks fire off you take the trade. The idea behind this is YES you are curve fitting. But think about what curve fitting is. Put together 50... 100+ separate neural networks and you now have an overall strategy that trades often, but each individual component is only trained to trade when it sees a very specific pattern. The main down side is those specific patterns each neural network has seen are not statiscally significant. Meaning the results from each pattern could be the result of random chance.. How do you weed out those patterns that were just pure luck? Anyone try something like this before? How were the results?
I did about 10 neural networks around 2001 just before the market went to decimals (while it was still trading in 8ths) to predict the next uptick or downtick for stocks. It was phenominally accurate for accurate (over 90% accurate with its predictions). Then the market went to decimals and a lot of level II shenanigans started with the hidden orders and large orders appearing and then disappearing to deliberatly mislead and it stopped working. stocks started moving .04 up, then .09 down and the only one getting rich was the broker. I think the notion of creating NN for a lot of stocks would take an enormous amount of computing power and too much time and effort to make it worth while. The bigger issues will come down to "what happens when you are wrong"? do you sell at a loss or do you add to your losers? Even a great strategy falls apart if you don't have equally good money management. I venture to say that good money management is the most important thing in trading. If you do it right, you can have more losers than winners and still come out on top. I think all of the time spent on the NN can be better spent elsewhere. Ed
Ed, Thanks for the reply. You are correct, exits makes a strategy work. For these tests, I am having a set time exit. For example exit after 5 bars. The reason is, I am trying to find an edge with these neural networks. Becomes very simple. Is the neural network finding a pattern that 5 bars from now will be up more often than its down.. or up more than down.. ie does it have a positive expectancy. Like this I do not have to worry about exits. I can grade all the neural networks evenly. Does the pattern represented by that neural network have an edge over randomness. To truly gauge these neural networks a baseline test must be performed prior. Such as run random iterations using the same exit strategy to determine the bias in the testing period. Then each neural network can be graded based on outperforming that baseline significantly. Once the neural networks have been selected, a final step can be to try different exit strategies/money management techniques to make the overall strategy more efficient.
I understand what you are doing. In theory I agree. However, from my experience, NN training is long and combersome and involves too much. Whereas, you can accomplish that task sooooooo much easier with simpler conventional methods, like backtesting. I do 90,000 backtests on each symbol, each week, and find a series of optimal numbers for each symbol. For instance, it tells me what kind of return to go after for each symbol. For example going after 8% return on 1 symbol may make sense, while 1.5% may make more sense for another, etc. My strategy is too involved to go into now, but I will say that just building a good backtest model is a lot easier, and in my opinion, a lot more valuable than the effort it will take to design, create, train, and test a bunch of NN's. All of that said, if you are up for the daunting task of creating the NN's, I believe in your theory that you can skew the odds in your favor with the use of a well built and managed series of neural networks. Yes, buy all means, I agree that it would work. Ed
I am not an expert in neural networks. So, take whatever I have to say with a grain of salt: 1) Training a neural network is very much like calibrating a polynomial approximation. For some things it works. For others it doesn't. 2) Before you start, it helps to understand limitations of a particular neural network. For example, if it is linear, there are certain patterns it will never be able to reproduce... I understand the idea Edbar is trying to convey is human brain can figure out repeatable patterns in price action much better then a run-out-off-the-meal neural network.
The idea used neural networks as that is what more people are familiar with. The idea I am getting at can be applied to any method capable of learning patterns. The thought being you take something and curve fit it beyond what the general "masses" say is acceptable. Then combine many of those pieces together into one large strategy...
OK. I'm game. Let's do it. Creating robotic traders happens to be what I do. I see your NN as a single indicator that will return a percent. The closer to 100%, the better. We can combine your indicator with a whole host of other indicators/rules already present in my system, so theoretically the strategy should be that much more accurate. How capable are you at cranking out the neural networks? Do you have the computing power, and NN software? My suggestion is to start with just one symbol, say U.S. Steel, stock symbol "X", as it moves $1 to $4 a day. Where will you get your input data to create and train the NN? Where will you get your input data to make the real-time output? I can create something to get your indicator results into my robotic traders in real-time, and the robotic traders will follow the strategy rules to the letter and take over from there. If you are up to it, send me email to edbar@exmsft.com. Like I said before, I have worked with NN in the past and know how powerful they can be! The NN's combined with the robotic traders can be a good combination. Ed
I am done for the night, but wanted to share some of what I am looking at. So far I only have 4 neural networks combined. All long only. The market I am testing is dow futures (YM). Trading only a single contract. This may not work out.. but should be a fun experiment if nothing else. Gain: +$14,025 (per contract) Trades: 181 Win%: 59 Profit Factor: 2.28 Sharp: .58 Max Draw down: -$1100 Largest Gain: $1215 Largest Loss: -$670 Nothing special yet... curious to see how it performs after say 50 neural networks.. Attached is a monthly breakdown of profitability..
Now up to 8 combined neural networks. Getting better, but still a long ways to go to get to the goal of approx 50. PL: +$24,460 (per contract) Profit factor: 2.37 Sharp: .72 Trades: 292 Win %: 60 Max DD: -$1270 Largest win: +$1445 Largest Loss: -$670 Only a few more Neural networks can be constructed from the current input set. After this the inputs fed to the neural network will have to be altered to start producing new ones. One thing with neural networks is that proper input is the key...