thirty days out leaves one open to events of randomness where a prediction with that percentage would be fantastic for the next day. so my argument is if you can predict thirty days out "why then" would your results not be as good or more accurate closer to the target? say the next day, see my point.

if you arrived at these results thru observation rather than optimization also this would have significant meaning. of course i could easily produce one hundred percent winning systems ignoring draw-down and time. so i would suggest not to get too excited until you have actually traded your results with real money which is the truth serum for all dreams.

These results were through optimization, and it could work for day-day trades, though the data gets far more expensive to run. Would you be available for a call to discuss further?

exactly - when you observe an anomaly or edge and then test for it typically this is experience proving itself. when you find and edge thru optimization the program will always aim to please by giving the optimal results.

if you test the same parameter setting over any other number of days parameter and it fails to have within proximity results, that is a sure sign of over or curve fitting. this was my point about the mine field of optimization.

there is a balance when doing optimization you should see good results within a large field of settings and parameters then you have something that is robust and could endure.

one other thing just to look at since i think you are using eod is that the prices you are running on maybe false because there is a close price and there is also a settlement price. the difference can cause completely different result deviation from real time trading. Especially if you're using any sort of "on close" argument.

First of all, of course 74% is really impressive but at the same time not the only statistic. It is quite possible that the rest 26% could be quite devastating for the pnl, so if you want meaningful feedback and not have ppl guessing, you need to provide better description.

Second, the idea or collaboration really does not work out in practice very well. I mean if you found edge and you are truly convinced it is edge why would you share it with anybody. That certainly crosses my mind and as Mark said, it is quite possible that results you get might be due to over fitting. You would get ppl to take you seriously only after a good time period of live trading but if they look good, you will find easily investors and them not need anybody to collaborate with...

Best of luck.

My program has reached 92%

Alex, unfortunately your post is a typical example of using terms on "common sense" basis: no one knows what that "efficiency" is in your terms, but it seems obvious. Could you specify what you mean? If only the percent of winning trades, then I should discourage your that it's very easy to create a strategy featuring 99% of winning trades, and which will lose money. However "efficiency" most frequently is used to talk about the precision of entries, MAE, MFE, average sunup and so on. So, a clarification from you would help making this discussion more meaningful.

Apart from that I'd also like to note that 30 days sample is nothing to discuss seriously.

can you say what's the risk reward per trade? or the profit factor (average profit in 74% trades/ average loss in the 26% trades).. once this is clear, one can come to conclusion is your strategy can make money.

Alex, please do elaborate on "efficiency". Are you predicting the stock price 30 days into the future or simply "higher/lower" ? I must agree with previous replies that you are not giving enough data for us to understand what your are claiming.

I am with Alex Krshtop, Nikolay Stoykov and Mark Brown, it is easy to find EA´s through optimization, the difficult is find systems with good rates (PF, SQN, Ram, RaM, RF, %Winner.....) out of sample.

I am working in a project building EA for MT4 with, almost 8 year of optimitation and after that i run all EA more than 1 year out of same and after that 1 year in Demo account and after Real accounts... and belive me that is not easy to find EAs that are stable, but is possible ;).

Now i have 78 EA in DEMO and 12 in REAL account with SQN > 2.

I guess everyone who has survived this business for more than 5 years knows that the key things is the ability to adapt to changes in the market regime. Something that cannot be found only in price time series alone.

@ Alex Krishtop

My dear friend, I beg to differ. As long as your model/system is not (over)optimized at the start, you can make good money with the unaltered parameterset. And, yes, I have been around for more than five years. Of course, there will be drawdowns, but a solid framework can cope with changing market conditions.

Yes (short term at least) and no :)

What is your "out of sample" success rate, and actual return per unit invested? Can you actually trade the entitities in any meaningful size?

Alex, I am in the same mindset as everyone else here in desiring to know what and how you define "efficiency". Our algorithm, MarketDNA, produces an 84% accuracy historically over the last 9 years in determining equity price movement, 65/35 long short with an average 8% equity gain and on avergae of 17 business days. We don't describe accuracy as "efficient" in our realm.

Do keep up the good work but as evident in the comments so far; you need a bit more dimensions to even begin to discribe your achievements, especially to fellow systemic traders. Otherwise you risk sounding like a cheap beginners guide book. "Make a million pounds trading forex" they'll often claim but utterly baseless. Do keep at it though. No discouragement intended.

Would love to see "picks" for first few weeks of 2016.

Ryan Fitts and all others - The post we just published would give you an idea of our algorithm, MarketDNA, and the "picks" we produce.

https://www.linkedin.com/pulse/marketdna-research-infy-returns-76-while-overall-markets-rossi?trk=hp-feed-article-title-like

Alex, you have to remember about one thing. I made algorithm, tested it, everything was ok. On the real account apperead trades, which didn't appear during tests (parallel to real trades I test it again). I don't know reason, but it's true. I think the problem is with delay of data on the platform. If you trade stat arb (example min. two instruments) and quotation of one of them is delay a few ms on the account can appear trades which there aren't during tests.

The first prediction model i developped a few days ago using machine-learning algorithms obtains quite similar results: 75% accuracy for DAX forecasting at 40days, backtested over 10years. This is a first draft and i truely think i can do better.

@Mark Brown : The stochastic nature of price movements makes it way harder to predict on a day to day basis than to detect longer, short/mid-term trends...If you really want to do that just stick to TA , which for some obscure reasons ( self-realisation ? ) also works sometimes.

@ clement

As for the original poster, please elaborate on your forecasting. WHAT do you forecast ? A pre-specified level of the DAX 40 days out ? A trend (higher/lower) 40 days out ? and how do you determine if a prediction is correct ?

@Marc I built a simple classifier which predicts the trend (up/down) of the DAX 40 days out. At First I tried to quantify the variation by trying to fit a polynomial regression, but ( probably because of the non-linearity of the problem ) it gave poor, inconsistent results. The solution i see is to this would be to extend my classfifier to multi-class, with variation intervals as new labels.

How do i test: Well when you make a prediction model based on Machine-learning you will split your data in 2 subsets : the training set and the test set (sometimes 3 for cross-validation).

Once you fit your model with the training set, you will confront it with the "reality" of the test set: 75% accuracy means my model is right at predicting the test set for 75% of its samples (it's the variance of the model).

I hope this answers your questions.

C.

@ Clement

Thanks but now you raise another question. What are the results of taking positions based on your predictions ? If you predict "higher" and 40 days out, it is 1 % higher, you make 1 %. However, if it is 10 % lower, you lose 10 %. So, having a 75 % prediction accuracy (higher/lower) does not say anything about the profitability of the model.

Can you elaborate on that ?

thanks

Alex, I am getting outperformance from a strategy put together for a european index: 69% of the picks outperforming the index by 16% and 31% of the picks underperforming by 10% for a holding period of 6 months. This is for a portfolio of 12 holdings and the returns (above) don't include compounding or dividend payments (or trading costs, although these are more than covered by yields). The cumulative average index adjusted return has stabilised at just over 6% for a 6 month holding period based upon 3 1/2 years of data from a real money portfolio (with the portfolio fully invested after the first 6 months). This performance also matched pretty well for about a couple of years back-tested data (I didn't have an easy means of historic data download when I put this together, so backtesting was limited). The index has also done well over the last 3 years as well, so I have close to doubled my money over this period. Not quite the 74% efficiency that you mention, but how do the index adjusted returns look?

>So, having a 75 % prediction accuracy (higher/lower) does not say anything about the profitability of >the model.

No that's right,

What you must keep in mind is that this classifier is just an element beyond others of an higher level trading algorithm. Ultimately this is it who will decide what to do with the predictions, and also handle the trades life cycle and money management.. So if you want a real insight of the returns, you must finally backtest your algo as a whole and not just the part which makes the prediction..but to be honest, i haven't integrated it to a full strategy yet so i can't give the figures.

@Clement

okay, thanks

Ive gotten 64% accuracy and 86% YOY return in the FX markets. Live trading, no back-testing, over 19,000 trades placed per year for the past 3 years.. still using it to this day.

Nobody really has reached over 92% efficiency in a period of 30 years?