I'm starting a journal to post some trades I'm making using BioComp Systems' Profit 8 product, which is a fairly advanced modeling product incorporating neural networks and other technologies. I thought I would post 30 trades or so, using a system for trading the ES that I developed a few months ago. This is a fairly highly-tuned system, so while it's done well for me so far, I'm kind of expecting performance to taper off in the future. In fact, I'm currently evaluating another Profit 8 system that should be much more robust, so I'll have something to switch to when this one fades away. This is a swing trading system, with the average trade being held for 2 days, more or less. While I'll be trading multiple contracts, I'll just post the results for 1 contract here, to keep things simple. I'll be using a disaster stop that is volatility-based, but set fairly wide. I'll start with 150 points, given current volatility levels. Eventually I'll probably move that to about 2.5 times the 20-day ATR. (right now that would be 205 points) I'll try to remember to post the signals and trades at the time I make them, which is usually about 8pm or so EST on weekday nights, earlier on Sundays. So, let's get started. I have a SHORT signal for tonight, which I will fill at the open this evening, and post the details later. Good trading!
Given that you will be using disaster stops of 150 to 205 ES points, what profits per trade or ATR does this system generate?
It has generated about 10 points average per trade in the past. In terms of reward to risk, that's obviously not very good, but the disaster stop gets hit very rarely. (once in the last 16 months, which of course was one of the biggest drops in decades) The short was filled at tonight's open at 965.75, so the current status is as follows: Open Position: S&P mini (ESZ8) - SHORT @ $965.75 Closed equity: 0
Got a LONG signal tonight, so exited the short position at exactly $950.00, for a gain of 15.75 points, or $787.50. I forgot to mention this earlier, but I'm not going to fool with commissions in this thread. Everyone's commissions are different, and it just creates more work for me. But keep in mind that you have to account for them. So our status is as follows: Open Position: S&P mini (ESZ8) - LONG @ 950.00 Closed equity: $787.50 1 down, 29 to go.
The trade went 42.75 points against you before you netted your 15.75 points. Is this typical for this trading system? From just the first trade, I know I don't have the stomach or risk capital for this system's style of swing trading.
Would you care to comment on the average number of inputs per NN and the total number of NNs used (with voting I gather)? Just for the curious. FWIW, I evaluated BP years ago and decided Matlab was the way to go as: - Nets were essentially a black box - Far too many bugs - Felt like it was developed in someone's garage as a weekend project (no offence). Note - I don't use Matlab NNs to generate Buy Sell signals.
That was about a 4.4% adverse excursion. About 10% of the trades exceed that level of MAE. Of course, most of the data for this analysis comes from backtesting, as I only have a couple of months out-of-sample on this system. In general, my experience is that systems tend to trade worse in real life than they do in backtesting. However, so far this one has been an exception to that rule, but it may not continue.
Equalizer, I can't tell you the number of NN's used, because that info is not available. But I can say that Profit is more than just a NN, although it does incorporate them. As far as the average number of inputs goes, I used an average of about three, although you can use as many as you want. You're right about the nets being a black box. That's almost always the case with NN systems. I think you either have to be able to live with that or pass on using them. It doesn't bother me, and I consider it an advantage that some people stay away from them for that reason, although I do understand their thinking. As far as bugs, P8 is a new version, the 8th generation. It does have some bugs, to be sure, although they are rapidly being dispatched. My biggest criticism of it is that the user interface is not as good or intuitive as P7 had. I understand that P8 is pretty much a complete rewrite, and the BioComp crew were focused on working under the hood, adding robustness and speed, which I think they did well. The user interface work still needs work in my opinion, but they are aware of it and working on it. In the meantime, I think it is definitely a usable product.
Got a 'short' signal from the system after the close on Friday, so I executed it at tonight's open, at 943.75. That resulted in a 6.25 point loss, or -$312.50. So the status is: Open Position: S&P mini (ESZ8) - SHORT @ 943.75 Closed equity: $475.00 2 down, 28 to go.
Last time I looked, the number of Nets used was given, in fact you could determine the voting of the different Nets that survived the curve-fitt.... err... I mean ... searching process. Anyway, re your second point, I think these jokers have taken far too long to address basic usability issues and bugs. Their software is not cheap, and there is no excuse for what they put out if these types of bugs still exist. Really, no offence to Carl, but his team should concentrate on quality not quantity. Just MHO. These reasons, and more were why I decide to roll my own Nets using Matlab. And like I stated earlier, I do not use nets to generate Buy/Sell signals. Your trades might vary. All the best of luck.