|
|
|
|
SGT_FW
by Aku_Aku. 06/02/25 17:54
|
|
|
|
|
|
|
|
|
|
|
|
|
0 registered members (),
357
guests, and 3
spiders. |
|
Key:
Admin,
Global Mod,
Mod
|
|
|
Multi-Factor Gaussian FX Strategy
[Re: TipmyPip]
#488292
07/06/24 15:56
07/06/24 15:56
|
Joined: Sep 2017
Posts: 147
TipmyPip
OP
Member
|
OP
Member
Joined: Sep 2017
Posts: 147
|
Following more factors of correlation, and Market sentiments, we can also include factor exposure, market capitalization, and industry data from various data sources on the web : #include <default.c>
var alpha, beta;
var filt, filttr, hband, lband;
vars PriceClose, GaussianFiltered, GaussianFilteredTR;
// Define additional factor variables
vars FactorExposure, MarketCap, IndustryData;
var beta0, betaFactor, betaMktCap, betaIndustry;
// Initialize factors for each time period
void initFactors() {
FactorExposure = series(0);
MarketCap = series(0);
IndustryData = series(0);
// Set coefficients based on your model or optimization
beta0 = 0.01;
betaFactor = 0.02;
betaMktCap = 0.03;
betaIndustry = 0.04;
}
void init() {
BarPeriod = 60; // 1-hour bars
LookBack = 144; // Lookback period for the Gaussian filter
PriceClose = series(priceClose());
GaussianFiltered = series(0);
GaussianFilteredTR = series(0);
alpha = calculateAlpha(144, 4); // Sampling Period and Poles
beta = (1 - cos(2 * PI / 144)) / (pow(1.414, 2 / 4) - 1);
initFactors();
}
var calculateAlpha(int period, int poles) {
var beta = (1 - cos(2 * PI / period)) / (pow(1.414, 2 / poles) - 1);
return -beta + sqrt(beta * beta + 2 * beta);
}
void calculateGaussianFilter(vars src, int poles, int period) {
vars filtSeries = series(0);
vars filtSeriesTR = series(0);
var lag = (period - 1) / (2 * poles);
vars srcData = series(0);
vars trData = series(0);
for(int i = 0; i < Bar; i++) {
if(i > lag) {
srcData[i] = src[i] + (src[i] - src[i - lag]);
trData[i] = TR() + (TR() - TR(lag));
} else {
srcData[i] = src[i];
trData[i] = TR();
}
}
for(int i = 0; i < Bar; i++) {
filtSeries[i] = filt9x(alpha, srcData[i], 1, poles);
filtSeriesTR[i] = filt9x(alpha, trData[i], 1, poles);
}
GaussianFiltered[0] = filtSeries[0];
GaussianFilteredTR[0] = filtSeriesTR[0];
hband = GaussianFiltered[0] + GaussianFilteredTR[0] * 1.414;
lband = GaussianFiltered[0] - GaussianFilteredTR[0] * 1.414;
}
var filt9x(var a, var s, int i, int poles) {
static vars f[10];
var x = 1 - a;
var filt = pow(a, i) * s + i * x * f[1] - (i >= 2 ? 36 * pow(x, 2) * f[2] : 0) + (i >= 3 ? 84 * pow(x, 3) * f[3] : 0)
- (i >= 4 ? 126 * pow(x, 4) * f[4] : 0) + (i >= 5 ? 126 * pow(x, 5) * f[5] : 0) - (i >= 6 ? 84 * pow(x, 6) * f[6] : 0)
+ (i >= 7 ? 36 * pow(x, 7) * f[7] : 0) - (i >= 8 ? 9 * pow(x, 8) * f[8] : 0) + (i == 9 ? 1 * pow(x, 9) * f[9] : 0);
for(int j = 9; j > 0; j--) {
f[j] = f[j - 1];
}
f[0] = filt;
return filt;
}
bool crossOver(vars data, var level) {
return (data[1] < level && data[0] >= level);
}
bool crossUnder(vars data, var level) {
return (data[1] > level && data[0] <= level);
}
// Predict returns based on factors
var predictReturn() {
var predictedReturn = beta0 + betaFactor * FactorExposure[0] + betaMktCap * MarketCap[0] + betaIndustry * IndustryData[0];
return predictedReturn;
}
void run() {
if(is(INITRUN)) {
init();
}
calculateGaussianFilter(PriceClose, 4, 144);
bool longCondition = crossOver(priceClose(), hband) && predictReturn() > 0;
bool closeAllCondition = crossUnder(priceClose(), hband) || predictReturn() < 0;
if (longCondition) {
enterLong();
}
if (closeAllCondition) {
exitLong();
}
}
Last edited by TipmyPip; 07/06/24 15:58.
|
|
|
Gaussian Bands Strategy
[Re: TipmyPip]
#488293
07/06/24 17:16
07/06/24 17:16
|
Joined: Sep 2017
Posts: 147
TipmyPip
OP
Member
|
OP
Member
Joined: Sep 2017
Posts: 147
|
Here is another version for the Gaussian Filter Strategy, while it can be considered non-linear under multi-correlation factors of overlapping data : #include <default.c>
// Variables
var alpha, beta;
var filt, filttr, hband, lband;
vars PriceClose, GaussianFiltered, GaussianFilteredTR;
vars UpperBand, LowerBand;
// Initialization
void init() {
BarPeriod = 60; // 1-hour bars
LookBack = 144; // Lookback period for the Gaussian filter
PriceClose = series(priceClose());
GaussianFiltered = series(0);
GaussianFilteredTR = series(0);
UpperBand = series(0);
LowerBand = series(0);
alpha = calculateAlpha(144, 4); // Sampling Period and Poles
beta = (1 - cos(2 * PI / 144)) / (pow(1.414, 2 / 4) - 1);
}
// Calculate Alpha
var calculateAlpha(int period, int poles) {
var beta = (1 - cos(2 * PI / period)) / (pow(1.414, 2 / poles) - 1);
return -beta + sqrt(beta * beta + 2 * beta);
}
// Modified Gaussian Filter with Bands
var filt9x_with_bands(var a, var s, int poles, var deviationMultiplier, var* upperBand, var* lowerBand) {
static vars f[10]; // Array to store previous filter values
var x = 1 - a;
var filt = s * a;
// Precompute powers of x to avoid redundant calculations
var x2 = x * x;
var x3 = x2 * x;
var x4 = x3 * x;
var x5 = x4 * x;
var x6 = x5 * x;
var x7 = x6 * x;
var x8 = x7 * x;
var x9 = x8 * x;
// Calculate the filter value iteratively
for (int i = 1; i <= poles; i++) {
switch (i) {
case 1:
filt += x * f[i];
break;
case 2:
filt -= 36 * x2 * f[i];
break;
case 3:
filt += 84 * x3 * f[i];
break;
case 4:
filt -= 126 * x4 * f[i];
break;
case 5:
filt += 126 * x5 * f[i];
break;
case 6:
filt -= 84 * x6 * f[i];
break;
case 7:
filt += 36 * x7 * f[i];
break;
case 8:
filt -= 9 * x8 * f[i];
break;
case 9:
filt += x9 * f[i];
break;
}
}
// Shift the previous values in the array
for (int j = 9; j > 0; j--) {
f[j] = f[j - 1];
}
f[0] = filt;
// Calculate standard deviation of filter values
var sum = 0, mean = 0, count = 10;
for (int j = 0; j < count; j++) {
sum += f[j];
}
mean = sum / count;
var varianceSum = 0;
for (int j = 0; j < count; j++) {
varianceSum += (f[j] - mean) * (f[j] - mean);
}
var stddev = sqrt(varianceSum / count);
// Calculate upper and lower bands
*upperBand = filt + deviationMultiplier * stddev;
*lowerBand = filt - deviationMultiplier * stddev;
return filt;
}
// Gaussian Filter Calculation
void calculateGaussianFilter(vars src, int poles, int period) {
vars filtSeries = series(0);
vars filtSeriesTR = series(0);
var lag = (period - 1) / (2 * poles);
vars srcData = series(0);
vars trData = series(0);
for(int i = 0; i < Bar; i++) {
if(i > lag) {
srcData[i] = src[i] + (src[i] - src[i - lag]);
trData[i] = TR() + (TR() - TR(lag));
} else {
srcData[i] = src[i];
trData[i] = TR();
}
}
for(int i = 0; i < Bar; i++) {
var upper, lower;
filtSeries[i] = filt9x_with_bands(alpha, srcData[i], poles, 1.414, &upper, &lower);
filtSeriesTR[i] = filt9x_with_bands(alpha, trData[i], poles, 1.414, &upper, &lower);
UpperBand[i] = upper;
LowerBand[i] = lower;
}
GaussianFiltered[0] = filtSeries[0];
GaussianFilteredTR[0] = filtSeriesTR[0];
hband = UpperBand[0];
lband = LowerBand[0];
}
// CrossOver and CrossUnder Functions
bool crossOver(vars data, var level) {
return (data[1] < level && data[0] >= level);
}
bool crossUnder(vars data, var level) {
return (data[1] > level && data[0] <= level);
}
// Main Trading Logic
void run() {
if(is(INITRUN)) {
init();
}
calculateGaussianFilter(PriceClose, 4, 144);
bool longCondition = crossOver(priceClose(), hband);
bool closeAllCondition = crossUnder(priceClose(), lband);
if (longCondition) {
enterLong();
}
if (closeAllCondition) {
exitLong();
}
}
|
|
|
Gaussian Decision Tree Hedging Strategy
[Re: TipmyPip]
#488305
07/11/24 05:43
07/11/24 05:43
|
Joined: Sep 2017
Posts: 147
TipmyPip
OP
Member
|
OP
Member
Joined: Sep 2017
Posts: 147
|
This trading strategy leverages advanced signal processing and machine learning techniques to make informed trading decisions. We employ a Gaussian filter, which smooths the price series by reducing noise, and derive key indicators from it. These indicators include the filtered price and its bands, calculated by a non-recursive method. The signals are generated by comparing the filtered price to its bands, and identifying potential entry and exit points. To enhance decision-making, we integrate a decision tree algorithm. This machine learning model is trained on historical signals, capturing complex patterns and relationships in the data. The model predicts future price movements, guiding long and short positions. By combining signal processing with machine learning, the strategy aims to exploit market inefficiencies and improve trading performance. #include <default.c>
// Variables
var alpha, beta;
var filt, filttr, hband, lband;
vars PriceClose, GaussianFiltered, GaussianFilteredTR;
vars UpperBand, LowerBand;
// Initialization
void init() {
BarPeriod = 60; // 1-hour bars
LookBack = 150; // Lookback period for the Gaussian filter
PriceClose = series(priceClose());
GaussianFiltered = series(0);
GaussianFilteredTR = series(0);
UpperBand = series(0);
LowerBand = series(0);
alpha = calculateAlpha(144, 4); // Sampling Period and Poles
beta = (1 - cos(2 * PI / 144)) / (pow(1.414, 2 / 4) - 1);
adviseLong(DTREE, 0, NULL, 0); // Initialize the Decision Tree model
}
// Calculate Alpha
var calculateAlpha(int period, int poles) {
var beta = (1 - cos(2 * PI / period)) / (pow(1.414, 2 / poles) - 1);
return -beta + sqrt(beta * beta + 2 * beta);
}
// Modified Gaussian Filter with Bands
var filt9x_with_bands(var a, var s, int poles, var deviationMultiplier, var* upperBand, var* lowerBand) {
static vars f[10]; // Array to store previous filter values
var x = 1 - a;
var filt = s * a;
// Precompute powers of x to avoid redundant calculations
var x2 = x * x;
var x3 = x2 * x;
var x4 = x3 * x;
var x5 = x4 * x;
var x6 = x5 * x;
var x7 = x6 * x;
var x8 = x7 * x;
var x9 = x8 * x;
// Calculate the filter value iteratively
for (int i = 1; i <= poles; i++) {
switch (i) {
case 1:
filt += x * f[i];
break;
case 2:
filt -= 36 * x2 * f[i];
break;
case 3:
filt += 84 * x3 * f[i];
break;
case 4:
filt -= 126 * x4 * f[i];
break;
case 5:
filt += 126 * x5 * f[i];
break;
case 6:
filt -= 84 * x6 * f[i];
break;
case 7:
filt += 36 * x7 * f[i];
break;
case 8:
filt -= 9 * x8 * f[i];
break;
case 9:
filt += x9 * f[i];
break;
}
}
// Shift the previous values in the array
for (int j = 9; j > 0; j--) {
f[j] = f[j - 1];
}
f[0] = filt;
// Calculate standard deviation of filter values
var sum = 0, mean = 0, count = 10;
for (int j = 0; j < count; j++) {
sum += f[j];
}
mean = sum / count;
var varianceSum = 0;
for (int j = 0; j < count; j++) {
varianceSum += (f[j] - mean) * (f[j] - mean);
}
var stddev = sqrt(varianceSum / count);
// Calculate upper and lower bands
*upperBand = filt + deviationMultiplier * stddev;
*lowerBand = filt - deviationMultiplier * stddev;
return filt;
}
// Gaussian Filter Calculation
void calculateGaussianFilter(vars src, int poles, int period) {
vars filtSeries = series(0);
vars filtSeriesTR = series(0);
var lag = (period - 1) / (2 * poles);
vars srcData = series(0);
vars trData = series(0);
for (int i = 0; i < Bar; i++) {
if (i > lag) {
srcData[i] = src[i] + (src[i] - src[i - lag]);
trData[i] = TR() + (TR() - TR(lag));
} else {
srcData[i] = src[i];
trData[i] = TR();
}
}
for (int i = 0; i < Bar; i++) {
var upper, lower;
filtSeries[i] = filt9x_with_bands(alpha, srcData[i], poles, 1.414, &upper, &lower);
filtSeriesTR[i] = filt9x_with_bands(alpha, trData[i], poles, 1.414, &upper, &lower);
UpperBand[i] = upper;
LowerBand[i] = lower;
}
GaussianFiltered[0] = filtSeries[0];
GaussianFilteredTR[0] = filtSeriesTR[0];
hband = UpperBand[0];
lband = LowerBand[0];
}
// Main Trading Logic
void run() {
BarPeriod = 60; // 1-hour bars
LookBack = 150; // Lookback period
TradesPerBar = 2;
if (Train) {
Hedge = 2; // Hedge during training
}
set(RULES | TESTNOW); // Set rules and test mode
if (is(INITRUN)) {
init();
}
calculateGaussianFilter(PriceClose, 4, 144);
// Generate some signals from GaussianFiltered in the -100..100 range
var Signals[2];
Signals[0] = (GaussianFiltered[0] - GaussianFilteredTR[0]) / PIP;
Signals[1] = 100 * FisherN(priceClose(), 100);
// Train and trade the signals using the decision tree model
if (adviseLong(DTREE, 0, Signals, 2) > 0)
enterLong();
if (adviseShort(DTREE, 0, Signals, 2) > 0)
enterShort();
}
Last edited by TipmyPip; 07/11/24 05:45.
|
|
|
Gaussian-Enhanced Hybrid Ensemble Strategy
[Re: TipmyPip]
#488318
07/20/24 05:00
07/20/24 05:00
|
Joined: Sep 2017
Posts: 147
TipmyPip
OP
Member
|
OP
Member
Joined: Sep 2017
Posts: 147
|
The following is an optimized version of the Gaussian Filter Strategy, which includes Decision Trees, and Neural Networks to monitor error correction within the signals. This code combines decision trees and neural networks to enhance previous versions of the implementation. It initializes and trains both models, uses Gaussian filters for signal processing, and employs a meta-model neural network to integrate their predictions for final trading actions, aiming to reduce errors and improve trading performance in a systematic, data-driven manner. #include <default.c>
// Variables
var alpha, beta;
var filt, filttr, hband, lband;
vars PriceClose, GaussianFiltered, GaussianFilteredTR;
vars UpperBand, LowerBand;
var Signals[2];
var MetaSignals[2]; // Signals for the meta-model
// Initialization
void init() {
BarPeriod = 60; // 1-hour bars
LookBack = 150; // Lookback period for the Gaussian filter
PriceClose = series(priceClose());
GaussianFiltered = series(0);
GaussianFilteredTR = series(0);
UpperBand = series(0);
LowerBand = series(0);
alpha = calculateAlpha(144, 4); // Sampling Period and Poles
beta = (1 - cos(2 * PI / 144)) / (pow(1.414, 2 / 4) - 1);
adviseLong(DTREE, 0, NULL, 0); // Initialize the Decision Tree model
adviseLong(NN, 0, NULL, 0); // Initialize the Neural Network model
adviseLong(NN, 1, NULL, 0); // Initialize the Meta-Model Neural Network
}
// Calculate Alpha
var calculateAlpha(int period, int poles) {
var beta = (1 - cos(2 * PI / period)) / (pow(1.414, 2 / poles) - 1);
return -beta + sqrt(beta * beta + 2 * beta);
}
// Optimized Gaussian Filter with Bands
var filt9x_with_bands(var a, var s, int poles, var deviationMultiplier, var* upperBand, var* lowerBand) {
static vars f[10]; // Array to store previous filter values
var x = 1 - a;
var filt = s * a;
// Precompute powers of x to avoid redundant calculations
var x_powers[10];
x_powers[1] = x;
for (int i = 2; i <= 9; i++) {
x_powers[i] = x_powers[i - 1] * x;
}
// Calculate the filter value iteratively
var coefficients[10] = {0, 1, -36, 84, -126, 126, -84, 36, -9, 1};
for (int i = 1; i <= poles; i++) {
filt += coefficients[i] * x_powers[i] * f[i];
}
// Shift the previous values in the array
for (int j = 9; j > 0; j--) {
f[j] = f[j - 1];
}
f[0] = filt;
// Calculate mean and variance in a single pass
var sum = 0, varianceSum = 0, count = 10;
for (int j = 0; j < count; j++) {
sum += f[j];
}
var mean = sum / count;
for (int j = 0; j < count; j++) {
varianceSum += (f[j] - mean) * (f[j] - mean);
}
var stddev = sqrt(varianceSum / count);
// Calculate upper and lower bands
*upperBand = filt + deviationMultiplier * stddev;
*lowerBand = filt - deviationMultiplier * stddev;
return filt;
}
// Gaussian Filter Calculation
void calculateGaussianFilter(vars src, int poles, int period) {
vars filtSeries = series(0);
vars filtSeriesTR = series(0);
var lag = (period - 1) / (2 * poles);
vars srcData = series(0);
vars trData = series(0);
for (int i = 0; i < Bar; i++) {
if (i > lag) {
srcData[i] = src[i] + (src[i] - src[i - lag]);
trData[i] = TR() + (TR() - TR(lag));
} else {
srcData[i] = src[i];
trData[i] = TR();
}
}
for (int i = 0; i < Bar; i++) {
var upper, lower;
filtSeries[i] = filt9x_with_bands(alpha, srcData[i], poles, 1.414, &upper, &lower);
filtSeriesTR[i] = filt9x_with_bands(alpha, trData[i], poles, 1.414, &upper, &lower);
UpperBand[i] = upper;
LowerBand[i] = lower;
}
GaussianFiltered[0] = filtSeries[0];
GaussianFilteredTR[0] = filtSeriesTR[0];
hband = UpperBand[0];
lband = LowerBand[0];
}
// Meta-Model for Combining Predictions using Neural Network
var metaModel(var dtreePrediction, var nnPrediction) {
MetaSignals[0] = dtreePrediction;
MetaSignals[1] = nnPrediction;
return advise(NN, 1, MetaSignals, 2); // Use another neural network (NN) model as meta-model
}
// Main Trading Logic
void run() {
BarPeriod = 60; // 1-hour bars
LookBack = 150; // Lookback period
TradesPerBar = 2;
if (Train) {
Hedge = 2; // Hedge during training
}
set(RULES | TESTNOW); // Set rules and test mode
if (is(INITRUN)) {
init();
}
calculateGaussianFilter(PriceClose, 4, 144);
// Generate some signals from GaussianFiltered in the -100..100 range
Signals[0] = (GaussianFiltered[0] - GaussianFilteredTR[0]) / PIP;
Signals[1] = 100 * FisherN(priceClose(), 100);
// Train and trade the signals using decision tree and neural network models
var dtreePrediction = advise(DTREE, 0, Signals, 2);
var nnPrediction = advise(NN, 0, Signals, 2);
var finalDecision = metaModel(dtreePrediction, nnPrediction);
if (finalDecision > 0)
enterLong();
else if (finalDecision < 0)
enterShort();
else {
if (isLong()) exitLong();
if (isShort()) exitShort();
}
}
|
|
|
Re: Gaussian-Enhanced Hybrid Ensemble Strategy
[Re: TipmyPip]
#488319
07/20/24 06:46
07/20/24 06:46
|
Joined: Sep 2017
Posts: 147
TipmyPip
OP
Member
|
OP
Member
Joined: Sep 2017
Posts: 147
|
I must confess, my fascination with automation knows no bounds. While resolving coding problems can be a daunting task, the correct implementation of automation can significantly simplify much of this work.
Imagine the incredible efficiency we could achieve by quickly resolving problems and eliminating the need for error corrections in our code. This would not only save us time but also enrich our lives by allowing us to focus on more creative and fulfilling pursuits.
It's true that addressing errors in automated code generation can be a slow process. However, developing and debugging code manually without automation demands just as much effort and time.
My enthusiasm for automation is matched by my eagerness to share its benefits with others. Please forgive my insistence, but I wholeheartedly encourage you to embrace automation as I have. Together, we can transform our manual trading tasks into streamlined, automated solutions. :-) Hahahah!
I surely, will need to many errors, in my code, But if anyone, feels like sharing any solutions, or ideas, please do.
Last edited by TipmyPip; 07/22/24 19:04.
|
|
|
Re: Gaussian-Enhanced Hybrid Ensemble Strategy
[Re: firecrest]
#488429
11/09/24 16:42
11/09/24 16:42
|
Joined: Sep 2017
Posts: 147
TipmyPip
OP
Member
|
OP
Member
Joined: Sep 2017
Posts: 147
|
Let me give you an example of answers from ZorroGPT: Train the PyTorch Model Separately First, you’ll use PyTorch to train your DRL model based on historical data. After training, save the model so Zorro can use it for predictions. import torch
from torch import nn, optim
# Example DRL Model (Adjust according to your architecture)
class TradingModel(nn.Module):
def __init__(self):
super(TradingModel, self).__init__()
self.fc1 = nn.Linear(NumSignals, Neurons)
self.relu = nn.ReLU()
self.fc2 = nn.Linear(Neurons, 1)
self.sigmoid = nn.Sigmoid()
def forward(self, x):
x = self.relu(self.fc1(x))
x = self.sigmoid(self.fc2(x))
return x
# Training function
def train_model(data_path, model_save_path):
data = np.loadtxt(data_path, delimiter=',') # Load data from CSV
X, y = data[:, :-1], data[:, -1]
X = torch.tensor(X, dtype=torch.float32)
y = torch.tensor(y, dtype=torch.float32).view(-1, 1)
model = TradingModel()
criterion = nn.BCELoss()
optimizer = optim.Adam(model.parameters(), lr=Rate)
for epoch in range(Epochs):
optimizer.zero_grad()
predictions = model(X)
loss = criterion(predictions, y)
loss.backward()
optimizer.step()
# Save model for use in Zorro
torch.save(model.state_dict(), model_save_path)Train the model on your historical data and save it: train_model("Data/Signals.csv", "model.pth")2. Load and Use the Model in Zorro is possible but with different code: Since Zorro cannot directly run PyTorch, you’ll need to create a separate Python script (e.g., predict.py) that: Loads the trained model. Receives data from Zorro for predictions. Outputs the prediction results back to Zorro. import torch
import numpy as np
from TradingModel import TradingModel # Import your model class
# Initialize model
model = TradingModel()
model.load_state_dict(torch.load("model.pth")) # Load saved model
model.eval() # Set model to evaluation mode
def predict(input_data):
input_tensor = torch.tensor(input_data, dtype=torch.float32)
with torch.no_grad():
prediction = model(input_tensor)
return prediction.item()
# Example usage for testing
# Replace [0.5, 1.2, -0.7, ...] with real data passed from Zorro
print(predict([0.5, 1.2, -0.7, ...])) # Output a prediction for testing3. Integrate Prediction with Zorro in lite-C Code In your Zorro script, the neural function will send input data to predict.py, retrieve the prediction result, and make trading decisions based on this prediction. To link Zorro with Python, you’ll use Zorro’s Rx() function to run the predict.py script and get the result back. function run()
{
vars InputData = series(priceClose());
double Data[NumSignals];
for(int i = 0; i < NumSignals; i++)
Data[i] = InputData[i]; // Gather recent data for model input
// Save input data to CSV format expected by predict.py
file_write("Data\\input_data.csv", Data, 0);
// Call Python predict script and retrieve prediction
Rx("Y <- predict('Data/input_data.csv')");
var prediction = Rd("Y[1]"); // Assuming Y[1] holds the prediction result
// Based on prediction, send buy or sell order
if(prediction > 0.5)
enterLong(); // Buy if prediction is above 0.5 threshold
else
enterShort(); // Sell if prediction is 0.5 or below
}
Now, there might be some errors, which you will need to ask ZorroGPT how do resolve them, let me assure you that it took me at least 40-80 exchanges of error handling to create the Gaussian Band indicator, of which I am still not sure the values are working properly in all time frames.
Last edited by TipmyPip; 11/09/24 16:47.
|
|
|
PyTorch DRL with ZorroGPT
[Re: TipmyPip]
#488430
11/09/24 16:52
11/09/24 16:52
|
Joined: Sep 2017
Posts: 147
TipmyPip
OP
Member
|
OP
Member
Joined: Sep 2017
Posts: 147
|
Zorro can indeed use PyTorch directly if you configure it to run Python scripts through the pynet.cpp interface. This method allows you to train and use a PyTorch model directly from Zorro without needing to write a separate data passing script. Here’s how to do it using Zorro’s pynet.cpp and a Python script with PyTorch: Install Required Python Libraries: Make sure you have Python 64-bit installed. Install the necessary libraries by running: pip3 install torch numpy math Open Zorro.ini and set the PythonPath64 variable to the path of your Python installation. For example: PythonPath64 = "C:\Users\YourName\AppData\Local\Programs\Python\Python312" Write a PyTorch Model Script: In the same directory as your Zorro script, create a Python file with the same name as your Zorro strategy file, but with a .py extension (e.g., if your Zorro script is MyStrategy.c, name the Python file MyStrategy.py). This Python script will contain the PyTorch model, functions for prediction, and optionally for training. import torch
import torch.nn as nn
import numpy as np
class TradingModel(nn.Module):
def __init__(self, input_size, hidden_size):
super(TradingModel, self).__init__()
self.fc1 = nn.Linear(input_size, hidden_size)
self.relu = nn.ReLU()
self.fc2 = nn.Linear(hidden_size, 1)
self.sigmoid = nn.Sigmoid()
def forward(self, x):
x = self.relu(self.fc1(x))
x = self.sigmoid(self.fc2(x))
return x
# Initialize the model with input size and hidden layer size
model = TradingModel(input_size=8, hidden_size=256) # Adjust based on your input features
# Load trained model weights if available
try:
model.load_state_dict(torch.load("model.pth"))
model.eval()
except FileNotFoundError:
print("Model weights file not found. Train the model first.")
def predict(input_data):
input_tensor = torch.tensor(input_data, dtype=torch.float32)
with torch.no_grad():
prediction = model(input_tensor)
return prediction.item()Using pynet.cpp to Call Python from Zorro: pynet.cpp allows Zorro to call Python functions, making it possible to run PyTorch directly for predictions. To use the model in Zorro, write the lite-C code that sends data to predict in your Python script, fetches the result, and executes trades based on the output. function run()
{
vars InputData = series(priceClose());
double Data[8]; // Adjust based on your input features
for(int i = 0; i < 8; i++)
Data[i] = InputData[i]; // Gather recent data for model input
// Call the predict function in MyStrategy.py
python("Y = predict", Data, 8); // Sends data to the Python function `predict`
var prediction = python("Y"); // Retrieve the prediction result from Python
// Trade decision based on prediction
if(prediction > 0.5) enterLong();
else enterShort();
}Compile and Run with Zorro: Use the 64-bit version of Zorro to run this script, as pynet.cpp requires a 64-bit Python setup. This setup calls the predict function in the MyStrategy.py script, passing Data as input, and receives Y, which is used to decide whether to enter a long or short position. Explanation of the Code Zorro’s python() Command: python() is used to send and retrieve data between Zorro and Python. Trade Logic: Based on the prediction returned by python("Y"), the script enters a long or short position. Model Reusability: You can save and load the trained model in the Python script, so Zorro will use the latest saved model for predictions. This setup provides a clean integration that allows Zorro to use a PyTorch model directly for predictions without needing extra data transfer scripts. Please note : these scripts are 80% logical and need to be refined so that it will work well. While deeper experimentation is always needed to make a trading strategy perfect.
Last edited by TipmyPip; 11/09/24 16:54.
|
|
|
|
