Machine learning in trading: theory, models, practice and algo-trading - page 2131

 

I'm thinking of doing a "full search" of regularities, without a target like "what will happen on the next candle" and the like...

Search consists in the fact that I will look for just regularities, the target is to find a pattern, and not "what will be on the next candle", also regularities can be stretched in time, for example if today was "event 1" and then "event 2" For example, if today there was "event 1" and then "event 2", and then " event 3", then tomorrow at 14:05 there will be a rising candle, something like that))

I have a better idea of how it should look like, and what algorithm to apply, but I would probably need some computing power, which I don't have (


By the way a question, how many repetitions of an event to consider it a pattern?

 
elibrarius:

Theoretically, it should be the same.
The number of different options in days, hours and minutes equals the number of options in sines and cosines. Both there and there in 7 days are 10080 different values, changing once a minute.
If there is any randomization in training, it may be because of this difference.

What did you train with, catbust?

Experiments are always more important.

Look at the first predictor T1 (Den_Nedeli_S), or more exactly its grid

T2


And days of week without time conversion.

As you see, the grids are different and there are different deltas between the digits, although the partitioning settings are the same:

catboost-0.24.1.exe fit  --learn-set train.csv   --test-set test.csv     --column-description %%a        --has-header    --delimiter ;   --model-format CatboostBinary,CPP       --train-dir ..\Rezultat\RS_208\result_4_%%a     --depth 6       --iterations 1000       --nan-mode Forbidden    --learning-rate 0.03    --rsm 1         --fold-permutation-block 1      --boosting-type Plain   --l2-leaf-reg 6         --loss-function Logloss         --use-best-model        --eval-metric Logloss   --custom-metric Logloss         --od-type Iter  --od-wait 100   --random-seed 0         --random-strength 1     --auto-class-weights SqrtBalanced       --sampling-frequency PerTreeLevel       --border-count 208      --feature-border-type MinEntropy        --output-borders-file quant_4_00208.csv         --bootstrap-type Bayesian       --bagging-temperature 1         --leaf-estimation-method Newton         --leaf-estimation-iterations 10        

Which means that you can more accurately match the split, which can either result in a fit or an improved result...

 
mytarmailS:

I'm thinking of doing a "full search" of regularities, without a target like "what will happen on the next candle" and the like...

Search consists in the fact that I will look for just regularities, the target is to find a pattern, and not "what will be on the next candle", also regularities can be stretched in time, for example if today was "event 1" and then "event 2" For example, if today there was "event 1" and then "event 2", and then " event 3", then tomorrow at 14:05 there will be a rising candle, something like that))

I have a better idea of how it should look like, and which algorithm to apply, but it will require computing power, which I probably do not have (

Oh, I also will do something similar :))))


mytarmailS:

By the way a question, how many repetitions of an event to consider it a pattern?

I use the criterion - not less than 1% of the entire sample and the "frequency" of recurrence of an event with the same outcome is important. I don't know how to measure the "frequency".

 
Aleksey Vyazmikin:

Experiments are always more important.

Look at the first predictor T1 (Den_Nedeli_S), or rather its grid

T2


And days of week without time conversion.

You see, the grids are different and there are different deltas between the digits, although the partitioning settings are the same:

Which means that you can be more precise with the split, which can either result in a fit or an improved result...

Okay. Sine + cosine is better not only for NS, but also for trees.

 
elibrarius:

Okay. Sine+cosine is better not only for NS, but also for trees.

I wouldn't jump to that conclusion - so far we can say that the result is not identical.

 

wheel of time


.

 
mytarmailS:

I'm thinking of doing a "full search" .....

I just threw in the time and day of the week and the color of the candle...

The data as a single week, forty weeks in all, and within them searched for patterns


Friday_18:20_dw means Friday - 18:20 - falling candlestick


confidence - percentage of rule 1 working is 100%

count - how many of such rules were found

 lhs                          rhs           support confidence coverage lift     count
[1]  {Пт_18:20_dw}             => {Чт_1:0_up}   0.500   1          0.500    1.290323 20   
[2]  {Пт_16:15_up}             => {Пт_5:0_dw}   0.500   1          0.500    1.290323 20   
[3]  {Пн_21:0_dw}              => {Пт_5:0_dw}   0.500   1          0.500    1.290323 20   
[4]  {Ср_12:50_dw}             => {Чт_22:55_up} 0.525   1          0.525    1.538462 21   
[5]  {Пт_18:40_dw,Ср_22:15_dw} => {Пн_14:50_up} 0.500   1          0.500    1.290323 20   
[6]  {Пн_0:0_dw,Пн_9:20_dw}    => {Пн_23:55_dw} 0.500   1          0.500    1.428571 20   
[7]  {Вт_20:40_up,Пн_0:0_dw}   => {Вт_21:5_up}  0.500   1          0.500    1.481481 20   
[8]  {Вт_9:40_dw,Пн_14:50_up}  => {Чт_1:0_up}   0.500   1          0.500    1.290323 20   
[9]  {Пн_0:0_dw,Чт_1:10_dw}    => {Чт_2:55_dw}  0.500   1          0.500    1.379310 20   
[10] {Пт_9:25_up,Ср_2:5_dw}    => {Пн_14:50_up} 0.500   1          0.500    1.290323 20   
[11] {Пн_14:50_up,Пт_9:25_up}  => {Ср_2:5_dw}   0.500   1          0.500    1.538462 20   
[12] {Вт_13:0_dw,Ср_2:5_dw}    => {Пн_14:50_up} 0.500   1          0.500    1.290323 20   
[13] {Чт_23:55_dw,Чт_4:20_dw}  => {Пн_0:0_dw}   0.500   1          0.500    1.250000 20   
[14] {Вт_18:55_dw,Пт_1:0_up}   => {Пн_14:50_up} 0.500   1          0.500    1.290323 20   
[15] {Вт_18:55_dw,Чт_1:0_up}   => {Пн_14:50_up} 0.525   1          0.525    1.290323 21   
[16] {Вт_2:45_up,Пн_9:50_dw}   => {Пн_0:0_dw}   0.500   1          0.500    1.250000 20   
[17] {Вт_2:45_up,Ср_20:40_up}  => {Пн_0:0_dw}   0.500   1          0.500    1.250000 20   


this rule

[1]  {Пт_18:20_dw}             => {Чт_1:0_up}   0.500   1          0.500    1.290323 20   

it means, if on Thursday at 1AM there was a rising candle, on Friday at 6:20 PM it will be falling. 20 rules were found, and the rule has worked 20 times out of 20 found

IDD...

 
Aleksey Vyazmikin:


I use the criterion of at least 1% of the entire sample and the "frequency" of the event with the same outcome is important. I don't know how to measure "frequency.

Same events are events with the same outcome.

100% of sample / % repetition. 1% is frequency, but without regularity. It's harder to measure regularity. Break it down into periods and look at the regularity of events. You can just minimize the number in a period and maximize it and divide min by max will get relative regularity, and you can use root mean square).

 
Oleg avtomat:

the wheel of time

you can add more harmonic minutes, and sum up the sinusoids, you get one curve to describe the three signs

But what to do and weekends and holidays, you need to take it all into account, what the hell is it all for?
 
mytarmailS: By the way, such a question, how many repetitions of an event to consider it a regularity?

I tried my metric , but it only works for SL=TP, for other ratios you have to count Hearst.