Yes, all this thread are ex-post results, as well as the previous threads. Also the bot is now being used with real money. Tom
Here is the mini folio trading right now. Has closed (picture posted in your thread: http://www.elitetrader.com/vb/showthread.php?s=&postid=3123035#post3123035 ) and now is restarting trading (see the GOLD entry). <img src="http://www.elitetrader.com/vb/attachment.php?s=&postid=3123046" width="1200" /> Tom
Thanks Tom, That's very impressive. Do you have a long term chart of that, or just equity curve for a couple of months. Again, where there has been no change of parameter during the period. As if I had invested my money in it last summer...
Well, apart that my threads cover a span of several months, if you want to see the exact results on past data you just run the bot on the past tickdata file. It will provide exactly the same results as it were trading with real money. [ I also have a large collection of tickdata (the bot has an accessory to <a href="http://www.datatime.eu/public/gbot/TickDataBacktesting.htm">record and store tickdata</a> ). ] But lately i trust more random data, because <b>1 realization of past data </b> means really nothing (unless trading is negative! . No matter how good it is (actually if too good is not good sign at all ). To be sure to prevent overfit is useful to do massive resampling with random data. That's the only way (imho) to hope to create robust strategies. The rest is just <b>curve fitting</b>! Tom
It got pretty lucky with SI, pulling it out from a difficult situation: <img src="http://www.elitetrader.com/vb/attachment.php?s=&postid=3123098" width="1200" /> Tom
Very Nice trading Tom, You say "if you want to see the results over past data - just run the bot over it". What I am trying to see is what the bot(or parameter set) you had at the START of the data period would have done. i.e. one that had not seen the data in the period. I know you are an experienced professional Tom - so I am sure you can see what I am looking for - the unfitted results. Are you saying that NO market data from the past is included in the training(optimisation) process?
The strategy played has nothing to do with any period or optimization. I determine strategies based on (real time) random data. Then that strategies are applied to the real time data. If you run the bot on past data you can use any strategy you like and see what it would have done. My <b>strategy selection</b> process has really nothing to do with past data. It's only based on mechanisms which make intuitive sense (to me) and that are, then, tested and calibrated using realtime random data, to avoid the possibility of overfitting. There is no overfitted results. If I showed you an "overfitted result", that is the best one on the set of permutation of all rules on a given dataset, it would be just unbelievably good. But that would mean nothing. Probably you have not really got that what you see as simulation <b>is not 1 backtest</b>, but the <b>average performances over 1000 runs, complete with distributions of the most useful (for me) performance indicators</b> Check better the results: these histograms are just the <b>resampling distributions</b>. When you have time to read the past post of this thread everything will be even more clear. And yes! No past market data is included in the optimization process. It's just <b>random data, totally unpredictable, created on the fly and traded on the fly, tick by tick, and different for each simulations</b>. If you run the sim on your machine you will get completely different prices at each run. But averagely, on a sufficiently large number replications, of you will see about the same results for a given strategy. Tom
OK. Now I understand Tom - thank you for taking time to explain it. Can you tell me how I could run the "bot" over 1 years worth of data please to see the result of having trained on random data. Or if you have such a NAV chart handy... This is very interesting.
Hi Moon, I don't care much about producing that chart for a simple reason: that no matter how good it could be, I would not believe it ;-)) ...i mean it would be irrelevant. Because it would just be 1 run on 1 single price realization. Nothing that can really mean anything. I did run some test on particular occurrences like the flash crash, just to double check the hedging mechanisms. But if you try the random data, you will see plenty of worse "crashes" of all sort. Random data is not just to "train" (not a good term in this context, as the underlying philosophy is that nothing can be "learned" from price data ) but also to verify the <b>robustness</b> of strategies. Infact you can run it on any number of new random sets and explore the outcomes. If you wish to get the bot for any experiments and local paper trading, just follow the simple indications on my page (it's a free distribution). Tom
Tom, Think Moon wants to load historic data and run the bot to analyze past results. Much like the flash crash data we ran over the 40 - 50 instruments. Moon... The bot trades a folio of instruments as a tradeset hedging each others play. Individually they follow a set high probability algo making trades, adding to positions reversing, exiting and reentering. At the folio level the positions and results are aggregated with global rules to hedge and close all before any one instrument runs away. Each instrument is normalized in terms of trading increments so they all deal in the same dollar equalized trading increment ie. A standard scalp of $250 so ES may be 5 points, 6E 20 ticks etc. This way the relative position and incremental trade results can be uniformly compared and risk managed globally. Running a historic analysis on a single instrument will only identify the projected results of the base algo. You have to run the entire folio of instruments in your tradeset to get a more realistic picture. This is where the possibilities grow exponentially. Your results depend on the instruments in your tradeset, your scalp size and a few user defined constraints. Unfortunately results are not a simple fixed ratio or percentage.