OPTIMISATION as the basis of a trading system. - page 2

New comment
 
alsu:

Proof of this is in the studio.

So far, I have not seen a graph of ideal parameters that does not exhibit stationarity breakdowns. If there is a (new) sensible idea, it's no problem to program it.

Stationarity breaks do happen.... I don't want to prove anything to anyone.... I made a parameter graph in excel six months ago... if I can find it, I'll post it.
 
jelizavettka:
Stationarity breakdowns happen.... I don't want to prove anything to anyone.... I made a parameter graph in excel six months ago... if I can find it, I'll post it.

How do I explain this... Probably it is, but we should take into consideration an important fact: the final result of the algorithm operation is a certain deal that is executed at certain prices. This means that variance of any parameter forecast should be recalculated in price variance, and only then the real value of the forecast will become clear.

We can use the simplest example that many people bang their heads against: a wave built on N counts is roughly predicted N times better than the price. But when you make a trade by this prediction its variance is recalculated to the variance of entry/exit price, i.e. back to N times. Plus the increased by N times dispersion of sampling noise that occurs in calculations - and as a result the forecast is even worse than the forecast at net price.

I'm not saying that all algorithms will have such a negative result, just the opposite - I suggest to think of such variants in which positive effect of the forecast exceeds the effects of increased noise variance. This is a necessary condition for a forecasting system to be profitable.

 
Any segment of history is essentially stationary, because it is always possible to find a pattern on which you can make as much money as possible. Non-stationarity (or market change) always happens in the future, which we do not know and do not assume....))))
 
LeoV:
Any segment of history is essentially stationary, because you can always find patterns on which you can make the maximum amount of money. Non-stationarity (or market change) always happens in the future, which we don't know or assume....))))


)))) What does "stationarity" and "you can find patterns" have to do with it? There are also regularities in non-stationary series.

"Any segment of history is essentially stationary" - what is "essentially"? I only knew before - stauionary and non-stationary series, but "essentially" - I don't know.

 
alsu:

How to explain it... Very often parameter/indicator charts look much smoother than the price chart, which gives the impression that it is easier to make forecasts based on them... Maybe it's true, but we should take into account an important fact: the final result of the algorithm operation is a certain deal that is made at certain prices. This means that variance of any parameter forecast should be recalculated in price variance, and only then the real value of the forecast will become clear.

We can use the simplest example that many people bang their heads against: a wave built on N counts is roughly predicted N times better than the price. But when you make a trade by this prediction its variance is recalculated to the variance of entry/exit price, i.e. back to N times. Plus the increased by N times dispersion of sampling noise that occurs in calculations - and as a result the forecast is even worse than the forecast at net price.

I'm not saying that all the algorithms will give such a negative result, just the opposite - I suggest thinking of such variants in which the positive effect of the forecast exceeds the effects of increased noise variance. This is a necessary condition for a forecasting system to be profitable.

Yes,... I've been thinking about that too... A small deviation in the forecasting of the period of the wagons - and a loss.... But i think classic syss tem will lag far behind the system with parameter forecasting because in the second case optimisation is used after each crossing.
 
jelizavettka: In the second case we use optimization after each intersection.

This is where OpenCL and a powerful graphics card come in handy.

But first you have to come up with an algorithm. And unfortunately the author is not very good at it.

 
Mathemat:
That's where OpenCL and a powerful graphics card come in handy. But first you have to come up with an algorithm.
And figure out what to optimize at all)) And since we don't have a powerful visual programmer in our brain, we don't want to go through all the variants - if only we could think up what we need))
 
Mathemat:
That's where OpenCL and a powerful graphics card come in handy. But first you have to come up with an algorithm.
Ooh. That must be something cool. Can't a regular CPU do it? How about just.... autosampling in expert code.......
 
Demi:


)))) What does "stationarity" and "you can find patterns" have to do with it? There are patterns on non-stationary series as well.

"Any segment of history is essentially stationary" - what is "essentially"? I only knew before - stauionary and non-stationary series, but "essentially" - don't know.

An intuitive definition of "stationarity in essence" ((c) LeoV), for example, might be this - it's easy to see where trading decisions had to be made...and which ones.

;)

 
Mathemat:

That's where OpenCL and a powerful graphics card come in handy.

But first you have to come up with an algorithm. And unfortunately the author is not very good at it.

You are reading my mind. (about the graphics card). about nalitu - yes, but will it work. there's also fuzzy logic in an implicit form and everything else.
1234567
New comment
Reason: