AI in trading

ketansp
Anyone here exploring AI/ML based strategies for active trading?

I have been working on this for couple of years on the side and last few months full time. I have been creating simple ML models of different combinations. Though I am fairly early in my journey, what seems to be working out for me is 0 to 5 min timeframe, where the impact of macro factors in minimal. The best performing model that I have today takes 10 candles of 1 min timeframe and tries to predict the 11th candle. Have been able to achieve profitability without brokerage and transaction costs. Need to fine-tine this model more or apply it on the right stocks. I will continue on this path for a while and see what happens; before calling it quits.

Would love to learn what others have been doing on the AI/ML side. Always open to exchanging notes in this field.
Tagged:
  • ANL
    I have a couple of ML logics that I have developed for long term purposes. I am specializing in LFT, so it really does not work, so not using it right now.
    Have been able to achieve profitability without brokerage and transaction costs
    Whatever you do, you must check the real cost. We can't say profitability without checking the cost. As a trader, brokerage and transaction costs are a nightmare as we are facing tax terrorism. So always check your strategy with cost, then confirm whether your logic is effective or not.
  • ketansp
    @ANL agreed. The litmus test for any AI or any algorithm for that matter is if it is able to make you any significant money, after all the costs, after taxation, after all the hassle.
  • ketansp
    An update here.

    I changed my AI model to predict 11th and 12th candle and used its combination to generate a signal. While the number of signals have dropped drastically, the model seems to have achieved profitability after transaction costs. Will keep fine tuning this further.
  • naveenarya
    import numpy as np
    import pandas as pd
    from sklearn.preprocessing import MinMaxScaler
    from tensorflow.keras.models import Sequential
    from tensorflow.keras.layers import LSTM, Dense

    # Load data
    data = pd.read_csv('one_minute_data.csv')

    # Feature selection
    features = ['Open', 'High', 'Low', 'Close', 'Volume']
    data = data[features]

    # Scaling
    scaler = MinMaxScaler()
    scaled_data = scaler.fit_transform(data)

    # Create sequences
    def create_sequences(data, seq_length=10):
    X, y = [], []
    for i in range(len(data) - seq_length):
    X.append(data[i:i + seq_length])
    y.append(data[i + seq_length, data.columns.get_loc('Close')]) # Predicting 'Close' price
    return np.array(X), np.array(y)

    X, y = create_sequences(scaled_data)

    # Split data
    split = int(0.8 * len(X))
    X_train, X_test = X[:split], X[split:]
    y_train, y_test = y[:split], y[split:]

    # Build model
    model = Sequential()
    model.add(LSTM(50, input_shape=(X_train.shape[1], X_train.shape[2])))
    model.add(Dense(1)) # Predicting one value (e.g., next 'Close' price)
    model.compile(optimizer='adam', loss='mean_squared_error')

    # Train model
    model.fit(X_train, y_train, epochs=10, batch_size=32, validation_split=0.1)
  • ketansp
    ketansp edited October 1
    @naveenarya applying LSTM on ohlcv data is the first thing folks tend to do. I did that as well years back. Unfortunately either it did not work out for me or I gave up too soon on it. Simpler deep neural network kind of architecture works for me.

    Nonetheless, I would love to hear how does this model perform for you.
  • naveenarya
    Hi Ketan ,

    Yeah it didn't work as expected, too many negative signals.

    How is your approach for this.?

    Would you mind sharing notes.?
  • ketansp
    LTSM, GRU etc work on dynamic memory. My hypothesis is not dynamic memory, but rather for fixed set of candles, last 10 to be precise. A simile deep neural network that replaces a deterministic approach of technical analysis should work. I have also simplified the input data instead of 5 different inputs, I just input a normalised value of (closePrice - openPrice) over the range.

    Also, coupling your neural network with genetic algorithms with save you lot of hassle of trying different combinations of the structure.
Sign In or Register to comment.