Machine learning in trading: theory, models, practice and algo-trading - page 242

 

It's all bullshit - the MO. It's like trying to know the weather for tomorrow from historical weather data. It's like applying facial recognition to a person who constantly has cosmetic surgeries and uses new makeup every time. The person is the same, but you can't recognize him or her.

In general, why try to catch and constantly adjust to changes in the market? Wouldn't it be better to look for something that doesn't change?

 
Andrey Dik:

It's all bullshit - the MO. It's like trying to know the weather for tomorrow from historical weather data. It's like applying facial recognition to a person who constantly has cosmetic surgeries and uses new makeup every time. The person is the same, but you can't recognize him or her.

In general, why try to catch and constantly adjust to changes in the market? Maybe it is more correct to look for something that does not change?

You got it all in one pile: what to look for and what to look for.

In this thread we're looking forwhat doesn't change?

And to do this, we use a huge set of proven tools, which are grouped under the name of "machine learning". Although this is not quite correct, because there are also tools that belong to datamining

 
SanSanych Fomenko:

You've lumped together what to look for and what to look for.

In this thread we are looking for exactlywhat doesn't change?

And to do this we use a huge set of proven tools, which are grouped under the name of "machine learning". Although this is not entirely accurate, since we also use tools that are classified as datamining

No, they're not. You learn from the weather in April, validate in May, and then wonder why the forecasts don't work in June.

And I guess those who checked the "huge set of proven tools" are all already billionaires? Or is the point - check it, make sure it doesn't work and then turn it in at R? (I mean the developers of "a huge set of proven tools")

 
SanSanych Fomenko:

That's not what we're talking about.

Now I see what you mean. The "divide a table of 4 predictors into 50 groups, and pick a profitable action for each group" method does have a lot of tweaking and shamanism, I agree.

But it's better to spend an hour and try it, test the strategy with a walk-forward test, and see if it works/(doesn't work). Sometimes such a simple strategy gives some positive result, or at least does not drain the balance, which is difficult enough to create just by choice.

 
Andrey Dik:

No, it doesn't. You learn the weather in April, validate in May, and then wonder why the forecasts don't work in June.

You can make twelve models. One for January, the second for February, the third for March, etc. Validate for different years. That would be good.
 
Dr.Trader:
You can do twelve models. One for January, one for February, one for March, etc. Validate with different years. That would be fine.
Of course you can. Only the next year it will be frost in May, constant rain in July and +5 instead of dry and hot weather, hot in September +35, and in December it is raining instead of snowing, and the Saudis are riding camels on snow-covered dunes. And things are much more fun in the market than the weather.
 
SanSanych Fomenko:

If we're talking about candlesticks, there are many books that list numerous combinations of candlesticks. Moreover, there are people..........................

And not one of these "fool's pamphlets" has statistics on the triggering of these candlestick patterns... What's that about? Let's use our brains...

SanSanych Fomenko:

My claim to all these candlesticks in particular, and to technical analysis in general, is that the profitability in the past has NOTHING to do with the future.

And MO has to do with profitability in the future????? Oh, well...

what you put in there is what you get out.

SanSanych Fomenko:

And if we take on models that are much more complex than the Three Soldiers, we should get something in return. And in my opinion this is proof that the future will be like the past.

MO is cool, MO is a lot different than some "three soldier" soldiers))) But when I put the last two candles into MO, it turned out that it didn't just not recognize the shapes, it confused the color of the candle, the fucking color of the candle. Think about it, Sanych!

It's not a problem of MO's stupidity, it's our problem, because you have to pre-process the data before you send it to MO, you have to make sure that MO can really understand what you want it to understand.

And you're always going on and on about picking out noise from non noise, but you don't understand the fact that you're working with noise from the beginning and you want to pick out the non noise in the noise ))))

Andrey Dik:

It's all bullshit - MO. It's like trying to know tomorrow's weather from historical weather data. It's like applying facial recognition to a person who constantly has cosmetic surgeries and uses new makeup every time. The person is the same, but you can't recognize him or her.

In general, why try to catch and constantly adjust to changes in the market? Wouldn't it be more correct to look for something that doesn't change?

For example?

 
Andrey Dik:
Of course you can. Only the next year in May it is freezing, in July it rains constantly and +5 instead of dry and hot weather, in September it's +35, and in December it rains instead of snow and the Saudis ride camels on snow-covered dunes. And things are much more fun in the market than the weather.

So what's the solution?

I understand that everything is not stationary, but is there a solution?

 
mytarmailS:

such as?

If the question is about a person with makeup - Take a pulse, fingerprints, retinal scans - a person's biophysiological portrait stays the same throughout life.
 
mytarmailS:

So what's the solution?

It is clear that everything is not stationary, but is there a solution?

I told you to look for market properties that do not change. What are the "parameters" of a price series? Volatility, the average value of the return of candles, what else? Which parameters change all the time and which are unchanging? Look for unchanging parameters of the price series, and then build your TS on this knowledge, then the TS will not depend on the changes of those parameters that are changeable.

As I said before - create a number of increments of RNG with parameters (distributions, probabilities, etc.). Twist the knobs of the parameters of this series. Try to build a TS which shows results independent of the twirling parameters of the series. If you succeed then you have found a solution. If no - then nothing will help, neither MoD and certainly not R with its "huge number of proven tools", nothing at all.

And the technical analysis, that has been frequently kicked in recent years, is not responsible for bad trading results, neither for the MO, but for the approach to the market when trying to track changes in the market and the more so, when it is unknown the next market "change" when previously created models stop working.