I'd have to agree. The software sounded interesting until I read that. Now I would definitely not consider using it.
With regards the server-based security aspect: 1. All communication between the client PC and the server is encrypted (standard digital certificate based as used by the banking sector). 2. The user space on the server is encrypted also. This is encrypted in such a way as to be unreadable by anyone (including the staff and programmers internally) and is linked to the client pc (it uses public/private key encryption, with the private key held by the client PC). It is not physically possible for anyone, including our company, to view your data or strategies. Only you. 3. A comprehensive legal agreement will be electronically signed by the company that sets out the above in a legal fashion (i.e. that we cannot (both literally and legally) view any information that is on our own servers that has been created by you the user). Also, that any information provided by you, the user, to us voluntarily (e.g. for the purposes of support/training/customised indicators/whatever) is strictly private and confidential and the sole ownership of the user in question. I totally understand the concerns you express and this is incidentally one of the reasons why the product is not yet ready is because all of thes technological and legal issues have to be in place prior to launch. Thanks for the query, hope that cleared up a few things Andy
gk1998, As regards comparison to Neuroshell: Neuroshell is a good product and would be are main competitor. There are a number of technological reasons why we feel our product is better: 1. The proprietary Neural Network building algorithms are extremely advanced (there is nothing like this on the marketplace at the moment- this is the reason that our software being server-based is a requirement: the processing speed required is far beyond the average PC even at todays standards). 2. The resulting Neural networks created are far more complex than those created by Neuroshell and provide vast scope in terms of the operations that can be performed. 3. The interface is much easier to use. This is a system tailored towards the average user who knows something of the stock market but is not a dedicated multimillion dollar day trading specialist (although they are catered for as well if you want to turn on the "advanced interface"). You can point and click your way through the wizard screens and end up with a fully back-tested strategy (the power is in the learning cycles). 4. On the non-technical side, the pricing model is much more flexible and caters for people who want to "try out" a system like this for a few months without a massive upfront cost. There is also an annual subscription model but having read your post I will put forward for discussion the possibiliity of having an "outright purchase" option or maybe a long-term contractually fixed price option also. Thank-you for your post/questions, please feel free to ask me any further questions here or pm me if you prefer. Andy
This doesn't make much sense. If the data and strategies are actually "run" or processed on the server side, obviously the server side has access the unencrypted strategy and data - otherwise it wouldn't be able to process it.
I believe that your business model is a non-starter mainly because you require that the system be run from your servers. Not only is there a security concern of a proprietary system being compromised, there are a number of integration issues you havenât addressed: How does one integrate a complex proprietary algorithm into your system? Normally this would be done by calling a user developed DLL from within the strategy as is done in NeuroShell Trader? The DLL has to be in binary executable form in order to run, what are the guarantees that an employee couldnât disassemble the DLL and then reverse engineer it? How would one integrate your system with a broker? Since the strategy runs on your servers, the user is at the mercy of how your administrator has set processing priorities as well as the current processing requirements of competing strategies running on the servers. Then there is the issue of delays in the connectivity between your servers and the user. The resulting latency in receiving trading signals could be harmful to oneâs equity especially during fast market conditions. ========= Getting back to your business model, the only reason that I can think of for centralized server based system is that you what to prevent your NN from being reversed engineered, otherwise the added expense doesnât make any economic sense at the lease prices youâre quoting. Itâs hard to imagine a NN that is so unique and proprietary to require that kind of protection, unless you have licensed some technology that would prevent its resale or incorporation into a commercial product.
To Answer your questions... Quah: The result of the training algorithm is a neural network. This resulting network is downloaded to the client machine upon completion and is not stored on the server. All the scanning takes place on the users PC using their own market data connection (IQfeed, tradestation, etc). For the non-professional version (EOD only), the scanning does take place at the server-side (if requested), and in this case the internal structure of the Neural Network is encrypted with the private key residing with the user. As far as disassembling or memory dumps of the servers being a technical issue, the supporting legal documentation will be very clear that the ownership of the data, strategy and network are the sole private property of the user (this is similar to the software development companies that develop strategies for traders explicitly- they must be told the strategy in order to program it but the legal documentation covers the rights of the user). WolfVector: >>How does one integrate a complex proprietary algorithm into your system? We don't support that at all currently. If you already have a complex propietry algorithm that works effectively then you probably don't need to build another strategy to add to it. >>How would one integrate your system with a broker? As I said above, the actual market scanning happens at the user's PC so it integrates in the same way as other packages. This covers the processing priorities question too. Only the non-pro version runs on our servers and that is an EOD package and timing isn't as critical. RiskTaker: >>So...if you guys went out of business, what would the trader be left with? Nothing? You would still have full functionality of all created strategies as they reside on your local system. Thank-you for you questions... I hadn't realised that security was such a thorny topic! Any other questions, please let me know, Thanks, Andy
Wolfvector, Also, to answer the question of the reason for the centralised server: Currently in testing, the average learning time for the algorithm is about 4 hours (thats running off 2 Powerful servers linked). On an average, single-cpu desktop machine this would be more like 16 hours dedicated. It would also require over 1Gb of memory. Upon launch there will be 16 servers in a cluster handling the learning (with additional servers being added as necessary). This will mean that even at peak load, the learning requests will take no longer than 8 hours for any particular user (with email notification upon completion). As I said, this is an extremely powerful algorithm unlike anything else in the market place but its not something that the average PC will handle. I suppose we could offer a stand-alone version to professional users who want to have total control over the algorithm with no communication with our servers and who have the hardware required - its something I will suggest. Thanks for the comments, Andy
That sounds good. But even if I was interested, which I am not since I develop my own neural networks, I would rather have it run on my own computer. I think running it in my computer for 16 hours is better than running it in somebody's else computer for 4 hours. I can dedicate a computer just for that, or do during the week end when I am golfing, or even in the background while I am using my computer (intel HT...). I think the concept is good, but you guys need to offer a standalone solution that people can install in their own PCs. P.S. For the records: Bigger/more complex Neural Network are not necessarily better. Taking so much time for training is definitely not better either, because all you end up doing is overtraining (overfitting) your NN. They will respond very good for past performance, but very bad for unseen scenarios. Whoever is interested in the subject should take a class from Stanford University Masters Program called: "Modern Applied Statistics: Learning". You'll know all the limits of NN and their training.