thanks for the reply, but who actually asked for your credentials? I was just curious whether all your previous posts were just some random stuff or whether I missed the coherent red thread somewhere. You answered my question, thanks. By the way, GSAM does recently poorly, Citadel as well, LEH is dead...not sure how much that speaks for itself. But no matter how successful you are, and I wish you all the success you work so hard for, pride usally does not get you very far in life, just my 2 cents being pretty much of the same age as you (or is this all an issue of self-identification?). Cheers.
There, this is the last time (3rd time trying to post my BATS MMM blotter... it gets deleted for reasonable reasons... I understand) I'm going to do this: "Name" = name of the model. "Code" = XXXYYY. XXX is the AI that made the model. YYY is the # in the model set it developed. "Condition" = How much the model is in sync with the market. "xxxx W" = W is the weight in the portfolio. "Custom Selection #1" = The P&L for that day. "Custom Selection #2" = The performance relative to how it should be acting on "normal" conditions for that day. My intentions were: 1. Before I hibernate away from posting in ET, I thought I'll post my blotter in ET. 2. I just wanted to tell people how great I am and my advice is very much worth it in ET. 3. Market making (Stat Arb) is viable for individuals (using BATS). 4. I AM A TRADING GAWD!!!! (It's a custom app and it doesn't prove how great I am in trading and I can have a negative equity on non-BATS trades, prolly only says, I can program in WinApp...)
TS - Interesting setup. I like the way you're monitoring how in sync with the market your systems are. One question - I presume that the 'code' used to describe the AI learning tool that created the model is some kind of identifier to describe the type of model & unique ID or something. Am I correct? If so, would you describe the process you go through to determine which type of AI model to use for the situation/market condition you're looking to exploit? Thanks, Eric
I am being forced to finish up where I left off and some people are single-mindedly concluding useless shit so I'm going to go back to the 95% of the trader fails talk... OK... As I mentioned previously, "MOST" (NOT ALL) Tech Analysis states are determined by the following: 1. Higher High / Lower low. 2. Close relative to High or Low. 3. Current Bar relative to Swing(# of Bars) So I made a simple system with a bunch of indicators that considers the 3 components above. I'd like to mention that this eliminates the suvivorship bias by providing all instances of the 3 parameter and 10 steps. (hence 1000 results). But to make this test easier for me to program and analyze... I've fitted the dataset and the models so that it's easier to see the difference with the test (hence the conclusions under a strong selection bias). In another words, the datasets (real historical data) are fitted based on end product of the so that conclusion is easier to see... The data consists of 2 columns: 1. Name of the model instance. + Instance #. 2. P/L of the model instance. --------- Models with a bunch of TA combination... With Commish and 2 tick slippage... 4.6% of the models are most likely to survive the test period. Others will lose money.
So... From the previous, I add the general TA trading system concepts. Things to note are: 1. Determining the nature of the signals. In this case, it's a mixture of Swing and Trend-following (Generally speaking) so I added the stops, sizing and money management to fit it from general terms. 2. Logic of the signals are untouched. Meaning each test instance will provide the same signal. So the number of Buy Signals that NBM0001 from the previous test and this test will be the same. 3. All added rules are generalized. The new rules are not optimized based on each instances. Trade Management helps a dumb system. ----------------------------------------------- About 34.7% of the models were profitable.
Now... I add a bunch of IT and Quant. perks to it where it manages the models without altering the signals. (Same signals different ways of handling it.) So now... I have a crappy system but by applying top-notch risk management logics and routines, you can get it to a point where it won't lose... I have to mention that there is NO computer learning involved. And there are no other set of data added to the signals. Just one little proof that you can increase your risk of ruin by having good risk (not money) management. ----------------------------------------------------- 51.4% chance of success. Statistically significant, no edge. (Well... the edge is risk management added but like I mentioned previously, the tests are under a major selection bias to make a few points)
Finally... I filter the signals with a definite edge that exposes a market inefficiency. I don't think there's much to add here...
So I posted 4 samples of data based on the same set of Tech. Analysis signals. 1. The first post (pureTA_PL.txt) represents a typical newbie who trades without any understanding of the market. 2. The second post (pureTA_MM_PL.txt) represents a typical ET trader who trades starts realizing that "money" management is important. 3. The third post (pureTA_Stat_Max.txt) represents a profitable trader who knows how to TRADE and has the ability to survive in markets. 4. The final post (pureTA_EdgeFiltered.txt) represents a trader with an edge. ------------------------------------------------------------------------ Again, this test is trivial. But it's one instance of an opinion to consider. 1. The trade management forum is kewl but not enough. 2. There's no psychological / niche / instinctive factors involved. Everything is coded. 3. Signals do not matter. High probability trading is BS. Trading as career is about the ability to handle a trade. (Signals come and go with the market changes) Chuck LeBeau's (Van stole his quote) study about "Entries being least significant" still stands. The study is somewhat of that proof, only that I do it in a whole different level as what people talk about in ET (Hence, the pureTA_MM_PL.txt) Finally... Yeah... I can take a retarded signal to a profitable system... Imagine how I will trade when the signal itself are good to great... Seriously, I'm a genius. Treat me like a GAWD!!!! PS. (Added): Everyone talks about thoeries about what Goldman Sachs does, High Frequency trading, character/personality categories and all the other useless shits. ET Monkeys can never seem to understand that there are 2 sides to a coin...
Column1 = Test Instance Column2 = Simple Probabilitistic MM (pureta_mm_pl.txt) Column3 = Deterministic MM Few simple stats (I get this on the Win. but I don't want post a pic) System Avg PL: -16500 to -2076 System Std Dev: 29195 to 30091 System Profitability: 34.7% to 48.5% The Std Dev. remained the very close because of we're dealing with the same set of signals. The distribution of each trade's P/L didn't change much. The key was that the deterministic MM rules stopped the models out early or filtered out the signals. ---------------------------------------------------------- 1. Google: "probabilistic deterministic". 2. Statistics are important, no doubt about it.