Another undesirable feature of NN is that they truly are "black boxes": it's difficult for the programmer to understand just why a buy/sell decision has been generated. Exactly as it may be difficult for a truly discretionary trader to explain precisely why he acted in that particular way. GS
One of the most interesting facets of using NN's or true statistical analysis is that when one looks at the least significant element, (the tick) and the range, (or actually lack of adequate range between one tick and the next), there is truly not enough range variation to make even a pittance of a reliable projection in terms of germane statistical analysis long term or short term for that matter. That doesn't mean you can't obtain range from subgrouping the data into various bars, either minute, tick, volume type bars, etc. The problem is that the numerical functions that feed the remaining parameters still need to come from random variation found in the least significant element, initially sampled at 100% of a lot size, which individual ticks cannot provide adequately. From a programming standpoint, creating an input engine and NN that a trader can feed multiple data into, we only need to know the mathematical NN model one cares to use, or even various models to opt calling on. The assumption would be that a trader would care to take the result and plan to enter / exit at the two ends of the extremes, less however much unreliability the NN is proving to have in the recent history. I simply have not as yet, seen a true NN that is both recursive and "moving" which will stay in tune with the market and remain forecasting reliably, because temperament changes both short and long term, affecting the reliability of how long a period the NN will need to be fed, versus how far into the future it's reliable. Without being able to verify the percent of error that's occurring at any given time, the end result is still a discretionary exit based upon emotion or a fixed set of P/L/R analysis and you don't need an NN if you can do Profit/Loss/Risk. Accordingly, one's NN will always under-perform the true market at some unknown value and the trader's decision based on the NN, will under-perform the NN itself, so in the end, I'd personally rather stick to conventional methods that have both a simpler and more historically reliable "sense" of security about them. While I've seen a number of NN's hit the market dead on well enough to create a great marketing demonstration and video-bite, I have yet to see a well documented instance of long term performance guaranteed from historical significance. ;-)
As a final "ya-but" to NN approaches, I know several guys spending to have redundant 6000 Mbit feeds into their homes, thinking they can pluck off peaks and valleys more reliably. Yet, they fail to understand that the timestamps on their data are being generated at their own system as the data arrives, not the true market time stamp from the exchange. (Even true of cloud served VPN feeds from the big boys). Reliably, on high volume swings in the futures market, the consolidated data vendors at the exchanges are seeing bottlenecks that can delay data up to several seconds in worst case, while it appears to the trader his data latency was dead on. How do you teach an NN to deal with that, if one gets hung with 20% of the contracts rounding the turn or huge slippage to get out on the peak or valley? I struggle to even teach other traders that, let alone have an algorithm account for it. LOL I'm pleased to find 70% of a trend came home
Thank you for all the inputs. The poll basically indicates 1/3 of the people is profitable with NN, 1/3 is not and 1/3 is working on it. The numbers do not look all that bad but the real troubling part is that some very good people like slacker, Spike500 etc had not seen the type of results expected from a lot of hard works and the bitch changes her predictions each day. Maybe there is a luck factor here. I think I am gonna wait for the next generation protein biochips with some sour cream on them in case they don't work. From what I heard from friends, many hedge funds make short term predictions... I just want to see if I can come up with ways to put in a couple of contracts along with them.
This thread has been a fascinating read. Considering the fact that the markets trend about 33% of the time and trade back and forth 66% of the time, I would think that it would not be that hard to create a NN that could start with those stats ... and build from there. Here's a thread where the trader seems to be trading a NN System, and he's doing pretty well: http://www.elitetrader.com/vb/showthread.php?s=&threadid=70418 Best, Jimmy
Confucius says: The superior man makes demands on himself, the inferior man makes demands on others. Now let's translate this into this discussion. The markets change all the time. Tell me how can something that is static and "trained" on historical data detect changes and adapt to these changes real time? NN can at the most "look forward" a timeframe of approx 10% before it has to be "retrained". The other thing to remember is (and which is readibly observable when the markets are quiet, like during Globex hours) is that buying or selling will affect the perception of the other participants (ehhh: flippers and market matching at the exchange immediately flipping the market with their large shown orders). This is never taken in account while doing any mechanical backtesting. Give me a discretionary trader who can "read the market conditions" real time and (s)he will outperform a machine. After all trading is a psychological exercise of fear, greed and manipulation. vital statistics
If Tradingpro agrees i will post my tradingsignals for a number of days that were published by them. Just to show what i mean that NN is inferior to human beings. A human being has a huge flexibility to react to an ever changing environment. Confucius was right: The superior man makes demands on himself, the inferior man makes demands on others.
http://www.research.ibm.com/deepblue/home/html/b.html I got a feeling Spike is gonna win because his opponent is running Bill Gates' software on a crappy PC.
It depends on what kind of NN you are talking about. Some of them give prediction without any additional information but other can provide user with information that can help user to take right decision. There are types of NN where neurons excitation takes place when excitation signals come on all input connections of the neuron. If these signals go from receptors that represent indicators then we can interpret NN conclusion as concatenation of signals that were received from these excited receptors. NN can also calculate frequency of such receipt combination met during train session when TS was profitable and when it was not. This information can help trader to decide if he want to enter into market or not.