Interesting article, the findings are similar to what I found myself when I walk-forward validated a ML based trading system relying on a combination of technicals / price data as features. What the article does not consider is the combination of returns and any sort of alternative data (e.g. news sentiment) as features. This on the other hand should paint a different picture altogether.

Andreas - I strongly believe agree with your comment.

May I add - when this technology becomes highly used it will increase market volatility much as we presently see 'Robotic Trading' has. That will change the market forever in a somewhat negative way.

Check the STOX project / cryptocurrency that might be related to this topic

Agree. Tried every approach, from big data set (years) to shortest period like 30 days with continuous model retraining but nothing really have been giving me an edge

Probably you have to include macroeconomic data and global situation data and spot long term trend

How highly amusing.

Python + scikit-learn ( for Support vector machines (SVMs) ) = Trading Model

Of course machine learning is curve fitting.

Only a script can.

I bet that the observations and findings likely still hold true. But this paper was published 1991. Is there new techniques with the new advancement in computing power?

"Thirdly, the choice of sample period may be criticized: it could be that the models were over-conditioned on a too- lengthy in-sample data set, which in one case ran from 1993 to 2008, with just two years (2009-2010) of out-of-sample data. "

The paper was published 2011 I presume so about 6 years ago.

I don't know if there were any other major advances in the technology since then.

I do understand from 3rd party not directly( I might be wrong) that the AI and ML are not working in the area of finding the trading pattern/s in the chaos by it self , In the area of execution it self we all agree that the machine is much better than the Human Trader . I am referring to the

"Thirdly, the choice of sample period may be criticized: it could be that the models were over-conditioned on a too- lengthy in-sample data set, which in one case ran from 1993 to 2008, with just two years (2009-2010) of out-of-sample data. "

The paper was published 2011 I presume so about 6 years ago.

I don't know if there were any other major advances in the technology since then.

I do understand from 3rd party not directly( I might be wrong) that the AI and ML are not working in the area of finding the trading pattern/s in the chaos by it self , In the area of execution it self we all agree that the machine is much better than the Human Trader .

" "We tested over 1,000,000 models using the latest machine learning techniques" sounds suspicious: just imagine, how much working time computational power one neeeds to [thoroughly] test "over 1,000,000 models"?!

But I share an opinion that ML and AI is hype, at least, mostly hype. Here is my non-technical essay on subject: https://letyourmoneygrow.com/2016/10/23/big-data-and-deep-learning-a-technology-revolution-in-trading-or-yet-another-hype/

Not until they become self aware. Until then machines will only know what we tell them.

@Tom Kadala "What we have found is that our algo is excellent at predicting a direction for a specific amount of time (about 3 times the period) but may or may not follow through with our intended exits. "

Can you elaborate more on that issue please.

4 example what is "(about 3 times the period)" what do you mean by that ?

and "not follow through with our intended exits."

How come a software would not follow the intended exit ?

Thank you for the inspiring comment.

Eli Weiss - Thank you for your interest. Our algo at RagingFX.com does a hefty mixture of evaluating pattern recognition, market sentiment (as it applies to our algo specifically), optimal technical parameter alignments, and more... prior to issuing a viable signal. Our signals include risk managed Stop Loss, high probable/optimal Entries, and Take Profit levels that match an aggregate score from a group of pairs. Each signal includes the Period (M5, M15...D1) that triggered the signal, so a typical signal may look as follows:

SECURITY TRADE ENTRY PRD STOP TP PROB LAST

EURCHF BUY 1.14396 H1 1.14228 1.146490 .75 1.14571

Here our algo is telling us that there's a 75% probability that if you place a market order to BUY EURCHF that the price will rise over three times the period or for the next three hours. The probability of hitting our intended exit or Take Profit level from a market entry is about the same, if not higher.

For signals that hit our suggested Entry level, the probability of hitting our Take Profit within our maximum 31 hour window, drops by about half. We are working diligently on improving this number.

If you like you can follow our progress at RagingFX.com. We update our results weekly on our site. You can also request to receive our Daily Dashboard Report at no charge for closer monitoring. Just send us an email from our site.

We expect this month with Germany/Italy's elections plus Draghi's initial comments on a QE reversal to offer an excellent testing period for our algo.

@Eli Weiss - Your advice makes sense if you're looking at an outcome over a few days but would, more than likely, fail over a sustained period. It's important to ride profits out to their fullest potential to offset many small losses. That's why algo's that seek small, high frequency trades tend to fail over time, ...which may also explain the reason you came to your conclusion in this blog when you stated that algos and machine learning don't work. All it takes is one black swan trade or trades to eat up profits.

That's why we've taken a different tact with our algo at RagingFX.com. Our approach is to let profits run using calculated risks within a fixed trading period, while focusing on avoiding losses altogether, (as opposed to minimizing them). It's easier said than done, but we are making progress. Ultimately, our goal is to develop machine learning enhancements that will sustain this behavior over a long period of time.

Tom Kadala -"the price will rise over three times the period or for the next three hours."

Does That mean that if the signal was taken from the hourly data the prediction that the price of EURCHF will rise with 75% probability is good for the next 3 hours?

Exit target mean the level that the EURCHF will trade at is only 50% probability ?

Maybe you can drop the Time target from say 3 hours to 30 min. and that will rise the probability that the EURCHF will rise to 90%.

Add to that instead of 50/50 price target you can move to 90% probability by lowering the price target.

All this valid only if the issue with the algo/ system is time horizon.

Means if the time horizon goes down the probability goes up.

If this assumption is right than go for the smallest possible time frame prediction with best price prediction probability and do it so many time a day. Execute with an ML , AI to get the best execution wise. and you good to go.

Your suggestions would be great if the math that governs probabilities and general uncertain behavior were that straightforward. Unfortunately, it's not. Cutting down the time from 3 hours to 30 minutes, actually works in reverse to your description. Rather than increase the probability, it decreases it. Compare it to driving a car at 60 mph for three hours. If you drove for only 30 minutes, what is the likelihood that you will have arrived to your original destination? Zero!

The handling of probabilities with algo's and machine learning code is not an exact science, so you need to use other parameters to further define likelihood of success. I wish it were as easy as you described. We would have been finished with this project much sooner.

Your approach to narrow the risk window does have some merit though but can probably be best achieved by reducing trade size and averaging. We are looking at other ways using some clever math and will be testing our theories later this month.

I was just looking to help by pointing to a different direction that might help maybe in generating new line of thinking.

Good luck with building the system.

Indeed! Machines are not creative but better executors then humans...