Machine learning in trading: theory, models, practice and algo-trading - page 2043

[[Deleted]]

Alexander_K:

Hmm... I'll have a look. I haven't worked with TF above M15 yet...

Where's your signal? I put it in my favorites to watch it, and now it's gone. What happened?

[Evgeniy Chumakov]

Oleg avtomat:

Where's your signal? I put it in my favorites for easy viewing, and now it's gone. What happened?

He must have found the Grail and gone to the Quiet House. )

[Aleksey Vyazmikin]

Interesting video, Maxim Is this the approach you are taking?

[Aleksey Vyazmikin]

And here is about what I do. Interestingly, the two authors from the video above and below say the opposite - one says the nets don't work, the other says the trees do :)

[[Deleted]]

Aleksey Vyazmikin:

Interesting video, Maxim Are you engaged in this approach?

Sort of, yes, but I haven't done it yet) as I have little faith in the success of this activity.

Networks work better on homogeneous data such as images or signals. Trees are better on heterogeneous like a lot of different non-normalized features.

how long has yandexair been around? i didn't know there was a youtube counterpart before

HOWEVER, here are some examples of the coolest thing you can find for time series at the moment - transformers

https://timeseriestransformer.readthedocs.io/en/latest/notebooks/trainings/training_2020_04_27__093505.html

but it all looks like a lagged forecast, just like with LSTM. Like the current value of the series is the best predictor of the next one, like in SB

[mytarmailS]

There is a request !!!

It is necessary to write a simple script for mt4 !

The essence is as follows

1) I click on a specific candlestick

2) The script writes in a notebook the date and time and closing price of this candle

That's it!

[Valeriy Yastremskiy]

Maxim Dmitrievsky:

sort of, yes, but I haven't done it yet) as I have little faith in the success of this endeavor.

Networks work better on homogeneous data, such as images or signals. Trees are better on heterogeneous like a bunch of different non-normalized features.

how long has yandexair been around? i didn't know there was a youtube counterpart before

HOWEVER, here are some examples of the coolest thing you can get for time series - transformers

https://timeseriestransformer.readthedocs.io/en/latest/notebooks/trainings/training_2020_04_27__093505.html

This is another one about transformers.

[[Deleted]]

Valeriy Yastremskiy:

I'm not sure how to translate it, but it's more or less understandable.

I have it in my favorites) like, threw before

[[Deleted]]

GRU on random marks on small dataset

Epoch 20 train err: 0.3469601273536682 tst err: 0.40891700983047485

What could be the pros and cons of this random label sampling?

def add_labels(dataset, min, max, markup):          #min, max - минимальная\максимальная продолжительность сделки, в барах
    labels = []                                     #сюда сохраняем метки
    for i in range(dataset.shape[0]-max):
        rand = random.randint(min,max)              #случайно выбираем продолжительность следующей сделки
        if i == 0:                                  #если это первый элемент массива, заполняем значениями 0.5, rand-1 штук\
            for a in range(rand-1):                 #поскольку нет более ранних цен для определения метки       
                labels.append(0.5)
        if dataset['close'][i] > (dataset['close'][i + rand] + markup):    #если текущая цена больше чем цена + rand баров вперед\
                labels.append(1.0)                                         #то метка 1.0 (продажа)
        elif dataset['close'][i] < (dataset['close'][i + rand] - markup):  #если меньше, то покупка
                labels.append(0.0)              
        else:
                labels.append(0.5)
    dataset = dataset.iloc[:len(labels)].copy()
    dataset['labels'] = labels
    return dataset

now i'll put the cuda pack for my videocard and use the videocard to do the math on the big data

the chips are just sequences of increments, 15 pieces per input. you can scale up

[Aleksey Vyazmikin]

Maxim Dmitrievsky:

sort of, yes, but I haven't done it yet) as I have little faith in the success of this activity.

If I understand correctly from the video, there is a function / library, which is looking for signs in the convolutional network, ie, ready-made templates that should be found patterns / predictors - I wonder what is expected to find there, how this mask was made - what is the logic, do not know by any chance?

Maxim Dmitrievsky:

how long has yandexair existed? i didn't know there was a youtube analog before

For a long time, about two years, I think there is.

Maxim Dmitrievsky:

Z.Y. Here are some examples of the coolest thing available for time series at the moment - Transformers

https://timeseriestransformer.readthedocs.io/en/latest/notebooks/trainings/training_2020_04_27__093505.html

but it all looks like a lagged forecast, just like with LSTM. Like the current value of the series is the best predictor of the next one, like in SB

I'm not sure if our input data is suitable for this network, it looks very cool in the pictures.