Attached Image
Machine learning is not a "one size fits all" problem.
Algorithmic Quant Trading (Machine Learning + Stat-Arb) 25 replies
Forex with Machine Learning Software Project 1 reply
Machine Learning + Retail Forex = Profitable? (Quant) 1 reply
Potential new machine learning style software. 79 replies
My most recent advancements into machine learning 16 replies
Disliked{quote} These are the results for the A=0, B=130, C=9, D=12 configuration on the EURJPY and USDJPY: {image} Machine learning is not a "one size fits all" problem.Ignored
Disliked{quote} Not so much of a difference actually. FX doesn't have that many gaps. The bars usually open very near of the close of the previous one. I use the close price simply because I cannot use the open of the bar the strategy receives. Per definition the bar it receives is closed (otherwise it is not a bar). Its open price is the price as of one hour ago in the simulated time. Therefore I need the open price of the NEXT bar. I made the exercise of shifting all of the indexes by one in order to use the open price instead. I take the current (simulated)...Ignored
Disliked{quote} Yes,It is a optimization problems about 4 parameters. and What do you think about trading in Asian sessions based on M15 or M5 TF,due to the long-term trend will be influenced by the fundamentals and our model did not consider it.Ignored
Disliked{quote} No cross validation. The machine learning method is already cross-validated since I retrain before each signal, the whole backtest is a cross-validation exercise. I see how the optimization of the machine learning variables can introduce curve fitting bias but I don't do anything to attenuate this fact besides using more than 20 years of data. This works for me, if I use less data (2007-2015 for example) I do get into curve fitting issues when going live. I use many different things for building of machine learning algorithms but this has...Ignored
Disliked{quote} I understand your view about using walk forward to avoid over optimization. However what do you think of these two papers on this subject: http://papers.ssrn.com/sol3/papers.c...act_id=2528780 http://papers.ssrn.com/sol3/papers.c...act_id=2308659 Data mining bias can come in even when you are not optimizing parameters directly. The fact that you tested many methods before finalizing one is itself a sort of optimization. I guess if you are doing a lot of experiments, then hurdle rate should be higher than usual...Ignored
Disliked{quote} I don't see why you cannot use the bar open. If you are doing an event-driven tick simulation you should have access to the current bar open and all ticks since open without any problems. You should also have access to all previous bar OHLC values for the training. Using opening prices is fundamental, if I use closes my results are completely different. While you don't use opening prices we simply cannot compare apples to apples. However I am completely sure that the F4 simulator does not allow data-snooping. That because of both my own...Ignored
QuoteDislikedMan, I feel tempted to buy you an F4 license just so that we can truly compare things up
Disliked{quote} Yes, I do an extensive data mining bias evaluation. I use a variation of White's reality check plus a method similar to the one suggested on the paper you posted. Still too early on the thread to deviate to this topic. I'll start talking about mining bias evaluation much later on.Ignored
Disliked{quote} I don't want to be mean but the GA you used to get A, B, C and D before backtesting over the whole set is already data snooping.Ignored
DislikedHello algoTraderJo, glad to see you back I am also following with great interest! As I subscribed to the Asirikuy framework I wondered sometimes if algoTraderJo was Daniel himself since, if I am not mistaken, you left us with similar questions and ML achievements/results as realised in the community There is a lot of valuable info in your thread as is in Daniels and the community forum - great! But the community seems to put more effort on the price action systems. I enjoy more the approach with ML and try to learn from you...Ignored
Disliked{quote} Can you reproduce the results posted by Jo with the Asirikuy framework?Ignored
DislikedYes, I have everything to reproduce the results (thanks for the files). I just have been so tied up in job and family the past few days that I had no chance to really follow up. With more practice it might be plug and play - for me the framework is still a huge world of dependencies to understand. This weekend I will have time and post my results.Ignored
DislikedI am very interested in how you go about choosing systems for ensembles. From what I understand, you 1. Have an idea about target (mse/mle regression in the last case), inputs (bars returns) and setup (hourly / daily with 4 Parameters to optimize) 2. Generate lots of systems with different parameters optimizing for Sharpe / R2 3. Pick individual systems that perform good 4. ... but how do you pick systems that perform good together as ensembles. As Asirikuy generates the system and does the Backtest at the same time, do you save generated systems...Ignored
As you say, these are questions I have asked before because they are important. I will not spoon feed the answers on the thread (as it took me years to get them!) but I would love to guide you folks through the process and help you with any guidance I can. For those who want to do the work I am more than willing to be helpful. Ask questions, do experiments.