Machine learning in trading: theory, models, practice and algo-trading - page 1611

[mytarmailS]

I will use the word "predictors" in my article:
If by predictors you mean fiches, then I think in general it won't be equivalent, it depends on how to divide fiches. Most likely due to lack of data a model that could theoretically be trained by 1000 will not be trained by 100.
It is unclear why we should do so, the chips are chosen on the basis of giving the model a minimally sufficient set of data. Since it was originally conceived as minimal, how to divide it later?

Dividing to reduce dimensionality...

How do we know this sufficient minimum of features? Only by training, a hundred chips cannot be trained, a thousand chips is not enough for a computer, but what if we need 10 000 chips?

For example, to describe a two-candle feature, we need to create 45 variables, and 105 variables for three candlesticks...

[Maxim Dmitrievsky]

Alexander_K2:

Here's the first one I found:

There were also similar posts, but I'm too lazy to look...

Saber always writes in such a way that until you understand what it means, a week will pass, and then it turns out that he did not mean it. And he himself, in fact, uses the same strategy on the same symbol (the other has not seen).

Alexander_K2:

Max, do not scare me that you're completely gone from forex... That would be very sad... Everything is just beginning :))

Sometimes I`m trading arbitrage or hands... Neural networks are dead, they do not want to trade properly

[Aleksey Mavrin]

mytarmailS:

Dividing to reduce dimensionality...

How do we know this sufficient minimum of features? Only by training, a hundred chips cannot be trained, a thousand chips is not enough for a computer, but what if we need 10 000 chips?

For example, to describe a complete trait with two candlesticks, we need to create 45 variables, and with three candlesticks we need 105 variables...

Could you please explain how we get 45 variables from two candlesticks?

I think it's too much, or what I think.

[Evgeny Dyuka]

mytarmailS:

Dividing to reduce dimensionality...

How do we know this sufficient minimum of features? Only by training, a hundred chips cannot be trained, a thousand chips is not enough for a computer, but what if we need 10 000 chips?

For example, to describe a two-candle fixture, we need to create 45 variables, and 105 variables for three candlesticks...

Try https://colab.research.google.com/
In the settings choose TPU, works fast enough, the only disadvantage - do not let long use, switch to a regular CPU, but it too zhiviny, you can charge overnight

[mytarmailS]

Aleksey Mavrin:

Can you tell me more about how that makes 45 variables out of two candles?

It seems to me something too much, or what I think.

We have two vectors of variables, the current candle and the previous one ("-1")

a = "open", "high", "low", "close", "center"

b = "open-1", "high-1", "low-1", "close-1", "center-1"

the variable "center" is the middle of the candlestick (high+low)/2, without this variable it is impossible to describe a pattern like "eskimo" etc. I think the meaning of other variables is not necessary to explain, they are obvious.

So, let's create all kinds of logical combinations (which can also be non-logical).

[1] "1"               "open   >   high"
[1] "2"              "open   >   low"
[1] "3"                "open   >   close"
[1] "4"                 "open   >   center"
[1] "5"                 "open   >   open-1"
[1] "6"                 "open   >   high-1"
[1] "7"                "open   >   low-1"
[1] "8"                  "open   >   close-1"
[1] "9"                   "open   >   center-1"
[1] "10"             "high   >   low"
[1] "11"               "high   >   close"
[1] "12"                "high   >   center"
[1] "13"                "high   >   open-1"
[1] "14"                "high   >   high-1"
[1] "15"               "high   >   low-1"
[1] "16"                 "high   >   close-1"
[1] "17"                  "high   >   center-1"
[1] "18"              "low   >   close"
[1] "19"               "low   >   center"
[1] "20"               "low   >   open-1"
[1] "21"               "low   >   high-1"
[1] "22"              "low   >   low-1"
[1] "23"                "low   >   close-1"
[1] "24"                 "low   >   center-1"
[1] "25"                 "close   >   center"
[1] "26"                 "close   >   open-1"
[1] "27"                 "close   >   high-1"
[1] "28"                "close   >   low-1"
[1] "29"                  "close   >   close-1"
[1] "30"                   "close   >   center-1"
[1] "31"                  "center   >   open-1"
[1] "32"                  "center   >   high-1"
[1] "33"                 "center   >   low-1"
[1] "34"                   "center   >   close-1"
[1] "35"                    "center   >   center-1"
[1] "36"                  "open-1   >   high-1"
[1] "37"                 "open-1   >   low-1"
[1] "38"                   "open-1   >   close-1"
[1] "39"                    "open-1   >   center-1"
[1] "40"                 "high-1   >   low-1"
[1] "41"                   "high-1   >   close-1"
[1] "42"                    "high-1   >   center-1"
[1] "43"                  "low-1   >   close-1"
[1] "44"                   "low-1   >   center-1"
[1] "45"                     "close-1   >   center-1"

Just two candles, miserable two candles.....

[mytarmailS]

The clearest answer is why ensembles of models work better than individual weak models.

Explanation in terms of DSP (Digital Signal Processing) by the example of signal and noise, how noise suppresses itself when summed

https://www.youtube.com/watch?v=wqD892r-wfo&list=PLmu_y3-DV2_kpP8oX_Uug0IbgH2T4hRPL&index=13

I recommend to watch all this lecture series, you will get wiser and get new ideas... Great lectures! )

[Forester]

mytarmailS:

We have two vectors of variables, the current candle and the previous one ("-1")

a = "open", "high", "low", "close", "center"

b = "open-1", "high-1", "low-1", "close-1", "center-1"

the variable "center" is the middle of the candlestick (high+low)/2, without this variable it is impossible to describe a pattern like "eskimo" etc. I think the meaning of other variables is not necessary to explain, they are obvious.

So, let's create all possible logical combinations (not logical ones, either).

Just two candles, miserable two candles.....

All these combinations of NS/forest will analyze inside itself. Only OHLC and Center should be fed to the input. And Center will be determined internally if it is useful for the result.

[mytarmailS]

Wow))

[Evgeny Dyuka]

mytarmailS:

Wow)))

pure martingale, it's always like that with him ))

[Evgeny Dyuka]

mytarmailS:

Wow)))

Half a year of such strategies, the maximum result on backtests - x5 for the year, but once a year is sure to lose everything and it is impossible to solve this problem.