Simulate common trading bots on real crypto data and prove that a neural networks is able to capture the pattern!
This project proves that a neural network can learn meaningful market patterns that stem from commonly used pre-programmed bots on the real market. It also permits its implementation in a real-time trading bot, and includes different tools for getting, analyzing and pre-processing data. It is meant to make this whole process more easy!
First install with:
pip install marketML
Example code:
cryptodata = CryptoDataGetter()
synth = SyntheticTrader(cryptodata)
synth_machine = LSTMachine().init(candle = "5min", layer1 = 40, layer2 = 15, lookb = 10, learn_rate = 0.03 , dropout = 0.1, reg = 1e-4)
""" ## Call historical data, simulate and apply an artificial trader ## """
_, _, synth_target, synth_features = cryptodata.get_historical_data_trim(
["1 August 2024 00:00:00", 15000], "BTCUSDT", Client.KLINE_INTERVAL_5MINUTE,
transform_func=synth.linear_RSI, transform_strength = 0.02, plot = False)
""" ############# Prepare Inputs, train the Neural Network ########### """
x_train, y_train, x_val, y_val, scaler = cryptodata.split_slice_normalize(lookb = 10, lookf = 5, target_total = synth_target, features_total = synth_features)
trainmean, train_std, valmean, val_std = synth_machine.fit(x_train, y_train, x_val, y_val, epochs = 50, batch = 16)
""" ############## Plot training and some examples #################### """
plot = MachinePlotter(synth_machine)
plot.plotmachine(trainmean, train_std, valmean, val_std)
plot.plot_tape_eval(x_val, y_val)From the original candlestick crypto data above,
We modify it by introducing a linear RSI trader that causes a price shift:
The input of the model, X, is always composed of whichever price over the last N timesteps and M features (we use 13). that can be computed directly.
In the histogram, we can wee the broadening of the price fluctuations due to the synthetic trader.
The model learns much more from this synthetic data than from the original pattern. This is due to the explicitly introduced correlation between the target price and a certain features: RSI.
After averaging of the price returns by lookf, we can compute the correlations C_i(tau) between returns and the i-th feature (RSI):
We can see theoretically the "predictiveness" of future prices based on their linear (there are higher order correlatios) correlation with past features/technical indicators. It remains the question of why our model could not capture this pattern, a safe assumption is that the signal is too weak compared to the noise.
- A question arises, to whether is correlation pattern is stable over time, this needs to be tested an quantified; that will give a notion of how often the model has to be retrained.
- Implementation of the trading bot strategy: Although this is working, there is no assurance of profit until the above pattern can be captured.
- A relation between the training and the trading must be established. Even if the synthetic trader could be modelled, would this give profits on future synthetic prices?
###################################################################