The following posts will outline detailed statistics on the 3 markets I will initially be trading. ES, CL, and 6E. The statistics are from 2009 until present as this was the "out of sample" period. --CL Results-- Per Contract Profit: +$229,220 Max DD: -$655 Trades: 1510 Percent Profitable: 85.23% PF 7.51 Sharp 1.87 Average Trade: 151.80 Attached is the equity curve for the period
--6E Results-- Per Contract Profit: +$130,160 Max DD: -$782 Trades: 1513 Percent Profitable: 88.63% PF 6.00 Sharp 3.18 Average Trade: $86.03
--ES Results-- Per Contract Profit: +$66,352 Max DD: -$415 Trades: 1342 Percent Profitable: 86.96 PF 5.36 Sharp 1.47 Average Trade: $49.44
profit factor over 5 with that many trades? looks too good to be true. are you sure you haven't made backtesting error? like look in the future. i prefer to trade systems with PF 1.6 to 3.0
I am using NinjaTrader. As far as I am aware there is no way to "look" into the future with it. In the past with custom trading engines I have built I had that issue. Part of the reason I use a system like NinjaTrader now is to shield me from those mistakes. If someone knows any common pitfalls as far as looking into the future with NinjaTrader that I can look out for let me know.
Those are some beautiful equity curves. One generally needs lots O' trades to realize such smoothness. Some more stupid questions for you: 1) What does your expectation/trade look like on ES? (Using your stats, I'm getting about $50/trade or about 4 ticks net). 2) Any execution assumptions that might give you issues in sim vs live environments? Do you enter at market at the open of a 1min bar or do you use limit orders that execute within the range a 1min bar? I'm curious about your NN topology; I've played with quite a few in my research. Are your 20 NNs essentially 20 different strategies or are they more of a consensus (ie: are they wired in an OR configuration or an AND configuration?). Based on your equity curves, I'm assuming they are wired as multiple strategies. Thanks again for the journal; Excellent thread this.
This is just an aside on the subject of neural nets to trade: There's a fairly new research area in NN known as the HyperNEAT project. As an example, it was trained to become an expert checker player in part because it could realize the spacial relationships between the pieces on the board without having to learn this information for itself over repeated trial and error (because there was previously no way of describing such spacial information from the get-go). It seems as though it should be possible to train such a network with price action alone given that it has a leg up with solving these kinds of positional problem domains. That is, price action is price action, nothing else needed like X no. of indicators thrown into the mix. Anyway food for thought, given that this thread is more like "let's see if the smartie pants who has a Grail NN in test can make $$$ in reality" rather than, "let's build a NN that can trade pure price action so everyone can get in on the fun."
I agree.... I really don't want to be a detractor here, but, those backtests are a huge red flag... intraday systems with proper fill and slippage assumptions AND with a profit factor > 2 are very rare. Anything over 3 is generally subject to very poor fill ratios. Frosty, I hate to say this, but, your testing process is flawed. You either assumed frictionless entry/exit and/or perfect fill ratios. How are you testing fills? Are your entries/exits limits? Did you simulate full-tick trade through for every transaction?
nothing beats good old tradestation and good old moving averages. i'm happy to see 99% of developers think otherwise and love to overcomplicate things i've tested thousands of different systems. 6,000 trades in 10 years? such systems always, always produce negative equity curves, going straight down. if properly tested. frost, your systems however may generate huge interest from the brokers who will probably get super excited to offer them to clients Interactive brokers should sent you a bottle of dom perignon if you'll trade these systems for more than 3 months.