Shuo Cao
399 lines
12 KiB
Plaintext
399 lines
12 KiB
Plaintext
{
|
||
"cells": [
|
||
{
|
||
"cell_type": "markdown",
|
||
"id": "de76d792",
|
||
"metadata": {},
|
||
"source": [
|
||
"# LSTM外汇价格预测 - 精简版\n",
|
||
"\n",
|
||
"本Notebook包含LSTM模型训练的核心代码，用于预测EURUSD价格。\n",
|
||
"\n",
|
||
"## 主要步骤\n",
|
||
"1. 导入库和设置设备\n",
|
||
"2. 从MT5获取历史数据\n",
|
||
"3. 数据归一化\n",
|
||
"4. 构造训练/验证数据集\n",
|
||
"5. 定义LSTM模型\n",
|
||
"6. 训练模型（早停机制）\n",
|
||
"7. 可视化训练过程\n",
|
||
"8. 导出ONNX模型"
|
||
]
|
||
},
|
||
{
|
||
"cell_type": "markdown",
|
||
"id": "063db750",
|
||
"metadata": {},
|
||
"source": [
|
||
"## 1. 导入库和设置设备"
|
||
]
|
||
},
|
||
{
|
||
"cell_type": "code",
|
||
"execution_count": null,
|
||
"id": "fee98c7a",
|
||
"metadata": {},
|
||
"outputs": [],
|
||
"source": [
|
||
"import torch\n",
|
||
"import torch.nn as nn\n",
|
||
"import numpy as np\n",
|
||
"import pandas as pd\n",
|
||
"import MetaTrader5 as mt5\n",
|
||
"from sklearn.preprocessing import MinMaxScaler\n",
|
||
"from datetime import datetime, timedelta\n",
|
||
"import matplotlib.pyplot as plt\n",
|
||
"import os\n",
|
||
"\n",
|
||
"# 设置中文字体\n",
|
||
"plt.rcParams['font.sans-serif'] = ['SimHei']\n",
|
||
"plt.rcParams['axes.unicode_minus'] = False\n",
|
||
"\n",
|
||
"# 选择计算设备（GPU或CPU）\n",
|
||
"device = torch.device(\"cuda\" if torch.cuda.is_available() else \"cpu\")\n",
|
||
"print(f\"使用设备: {device}\")"
|
||
]
|
||
},
|
||
{
|
||
"cell_type": "markdown",
|
||
"id": "19c6b9d0",
|
||
"metadata": {},
|
||
"source": [
|
||
"## 2. 从MT5获取EURUSD历史数据"
|
||
]
|
||
},
|
||
{
|
||
"cell_type": "code",
|
||
"execution_count": null,
|
||
"id": "eee75337",
|
||
"metadata": {},
|
||
"outputs": [],
|
||
"source": [
|
||
"# 连接MT5\n",
|
||
"mt5.initialize()\n",
|
||
"\n",
|
||
"# 设置时间范围（最近5年）\n",
|
||
"end_date = datetime.now()\n",
|
||
"start_date = end_date - timedelta(days=5*365)\n",
|
||
"\n",
|
||
"# 获取EURUSD H1数据\n",
|
||
"rates = mt5.copy_rates_range(\"EURUSD\", mt5.TIMEFRAME_H1, start_date, end_date)\n",
|
||
"mt5.shutdown()\n",
|
||
"\n",
|
||
"# 转换为DataFrame\n",
|
||
"df = pd.DataFrame(rates)\n",
|
||
"print(f\"获取到 {len(df):,} 条数据\")\n",
|
||
"df.head()"
|
||
]
|
||
},
|
||
{
|
||
"cell_type": "markdown",
|
||
"id": "4fda9d49",
|
||
"metadata": {},
|
||
"source": [
|
||
"## 3. 数据归一化"
|
||
]
|
||
},
|
||
{
|
||
"cell_type": "code",
|
||
"execution_count": null,
|
||
"id": "c5dccb3f",
|
||
"metadata": {},
|
||
"outputs": [],
|
||
"source": [
|
||
"# 创建归一化器，将数据缩放到[0,1]范围\n",
|
||
"scaler = MinMaxScaler()\n",
|
||
"\n",
|
||
"# 对开、高、低、收、成交量进行归一化\n",
|
||
"df[['open', 'high', 'low', 'close', 'tick_volume']] = scaler.fit_transform(\n",
|
||
" df[['open', 'high', 'low', 'close', 'tick_volume']]\n",
|
||
")\n",
|
||
"\n",
|
||
"print(\"归一化完成\")\n",
|
||
"print(f\"最小值: {scaler.data_min_}\")\n",
|
||
"print(f\"最大值: {scaler.data_max_}\")"
|
||
]
|
||
},
|
||
{
|
||
"cell_type": "markdown",
|
||
"id": "15ddac0b",
|
||
"metadata": {},
|
||
"source": [
|
||
"## 4. 构造时间序列数据集"
|
||
]
|
||
},
|
||
{
|
||
"cell_type": "code",
|
||
"execution_count": null,
|
||
"id": "4887abce",
|
||
"metadata": {},
|
||
"outputs": [],
|
||
"source": [
|
||
"def create_dataset(data, lookback=10):\n",
|
||
" \"\"\"将时间序列转换为监督学习数据集\"\"\"\n",
|
||
" X, y = [], []\n",
|
||
" for i in range(len(data) - lookback - 1):\n",
|
||
" X.append(data[i:i+lookback]) # 输入：前10根K线\n",
|
||
" y.append(data[i+lookback, 3]) # 输出：第11根K线的收盘价\n",
|
||
" return np.array(X, dtype=np.float32), np.array(y, dtype=np.float32).reshape(-1, 1)\n",
|
||
"\n",
|
||
"# 提取特征\n",
|
||
"features = df[['open', 'high', 'low', 'close', 'tick_volume']].values\n",
|
||
"X, y = create_dataset(features)\n",
|
||
"\n",
|
||
"# 划分训练集（80%）和验证集（20%）\n",
|
||
"train_size = int(len(X) * 0.8)\n",
|
||
"X_train = torch.from_numpy(X[:train_size]).to(device)\n",
|
||
"y_train = torch.from_numpy(y[:train_size]).to(device)\n",
|
||
"X_val = torch.from_numpy(X[train_size:]).to(device)\n",
|
||
"y_val = torch.from_numpy(y[train_size:]).to(device)\n",
|
||
"\n",
|
||
"print(f\"训练集: {X_train.shape}\")\n",
|
||
"print(f\"验证集: {X_val.shape}\")"
|
||
]
|
||
},
|
||
{
|
||
"cell_type": "markdown",
|
||
"id": "208c234b",
|
||
"metadata": {},
|
||
"source": [
|
||
"## 5. 定义LSTM模型"
|
||
]
|
||
},
|
||
{
|
||
"cell_type": "code",
|
||
"execution_count": null,
|
||
"id": "c2916d2d",
|
||
"metadata": {},
|
||
"outputs": [],
|
||
"source": [
|
||
"class SimpleLSTM(nn.Module):\n",
|
||
" \"\"\"简单的LSTM神经网络\"\"\"\n",
|
||
" def __init__(self):\n",
|
||
" super().__init__()\n",
|
||
" # LSTM层：输入5个特征，隐藏层20个神经元\n",
|
||
" self.lstm = nn.LSTM(input_size=5, hidden_size=20, num_layers=1, batch_first=True)\n",
|
||
" # 全连接层：将20维输出映射到1维（预测值）\n",
|
||
" self.fc = nn.Linear(20, 1)\n",
|
||
"\n",
|
||
" def forward(self, x):\n",
|
||
" lstm_out, _ = self.lstm(x)\n",
|
||
" return self.fc(lstm_out[:, -1, :]) # 取最后一个时间步的输出\n",
|
||
"\n",
|
||
"# 创建模型\n",
|
||
"model = SimpleLSTM().to(device)\n",
|
||
"optimizer = torch.optim.Adam(model.parameters(), lr=0.001)\n",
|
||
"criterion = nn.MSELoss()\n",
|
||
"\n",
|
||
"print(\"模型创建完成\")"
|
||
]
|
||
},
|
||
{
|
||
"cell_type": "markdown",
|
||
"id": "9e6f4eeb",
|
||
"metadata": {},
|
||
"source": [
|
||
"## 6. 训练模型（带早停机制）"
|
||
]
|
||
},
|
||
{
|
||
"cell_type": "code",
|
||
"execution_count": null,
|
||
"id": "04481e99",
|
||
"metadata": {},
|
||
"outputs": [],
|
||
"source": [
|
||
"# 训练参数\n",
|
||
"max_epochs = 500\n",
|
||
"patience = 20 # 验证Loss不降20轮则停止\n",
|
||
"\n",
|
||
"# 记录训练过程\n",
|
||
"train_losses = []\n",
|
||
"val_losses = []\n",
|
||
"best_val_loss = float('inf')\n",
|
||
"best_epoch = 0\n",
|
||
"patience_counter = 0\n",
|
||
"best_model_state = None\n",
|
||
"\n",
|
||
"print(\"开始训练...\")\n",
|
||
"for epoch in range(max_epochs):\n",
|
||
" # 训练阶段\n",
|
||
" model.train()\n",
|
||
" optimizer.zero_grad()\n",
|
||
" train_output = model(X_train)\n",
|
||
" train_loss = criterion(train_output, y_train)\n",
|
||
" train_loss.backward()\n",
|
||
" optimizer.step()\n",
|
||
" \n",
|
||
" # 验证阶段\n",
|
||
" model.eval()\n",
|
||
" with torch.no_grad():\n",
|
||
" val_output = model(X_val)\n",
|
||
" val_loss = criterion(val_output, y_val)\n",
|
||
" \n",
|
||
" # 记录Loss\n",
|
||
" train_losses.append(train_loss.item())\n",
|
||
" val_losses.append(val_loss.item())\n",
|
||
" \n",
|
||
" # 每10轮打印一次\n",
|
||
" if (epoch + 1) % 10 == 0:\n",
|
||
" print(f\"Epoch {epoch+1:3d} | 训练Loss: {train_loss.item():.6f} | 验证Loss: {val_loss.item():.6f}\")\n",
|
||
" \n",
|
||
" # 早停检查\n",
|
||
" if val_loss.item() < best_val_loss:\n",
|
||
" best_val_loss = val_loss.item()\n",
|
||
" best_epoch = epoch + 1\n",
|
||
" patience_counter = 0\n",
|
||
" best_model_state = {k: v.cpu().clone() for k, v in model.state_dict().items()}\n",
|
||
" else:\n",
|
||
" patience_counter += 1\n",
|
||
" \n",
|
||
" if patience_counter >= patience:\n",
|
||
" print(f\"\\n早停触发！最佳模型在Epoch {best_epoch}\")\n",
|
||
" break\n",
|
||
"\n",
|
||
"# 恢复最佳模型\n",
|
||
"if best_model_state:\n",
|
||
" model.load_state_dict(best_model_state)\n",
|
||
" model = model.to(device)\n",
|
||
" print(f\"已恢复最佳模型 (Epoch {best_epoch}, 验证Loss: {best_val_loss:.6f})\")"
|
||
]
|
||
},
|
||
{
|
||
"cell_type": "markdown",
|
||
"id": "1669bc75",
|
||
"metadata": {},
|
||
"source": [
|
||
"## 7. 可视化训练过程"
|
||
]
|
||
},
|
||
{
|
||
"cell_type": "code",
|
||
"execution_count": null,
|
||
"id": "cd1011f3",
|
||
"metadata": {},
|
||
"outputs": [],
|
||
"source": [
|
||
"# 绘制训练和验证Loss曲线\n",
|
||
"plt.figure(figsize=(12, 5))\n",
|
||
"\n",
|
||
"# 完整曲线\n",
|
||
"plt.subplot(1, 2, 1)\n",
|
||
"plt.plot(train_losses, label='训练Loss', color='blue', linewidth=2)\n",
|
||
"plt.plot(val_losses, label='验证Loss', color='orange', linewidth=2)\n",
|
||
"plt.axvline(x=best_epoch-1, color='red', linestyle='--', label=f'最佳模型 (Epoch {best_epoch})')\n",
|
||
"plt.xlabel('Epoch')\n",
|
||
"plt.ylabel('Loss')\n",
|
||
"plt.title('训练和验证损失曲线')\n",
|
||
"plt.legend()\n",
|
||
"plt.grid(True, alpha=0.3)\n",
|
||
"\n",
|
||
"# 后80%放大曲线\n",
|
||
"plt.subplot(1, 2, 2)\n",
|
||
"start_idx = int(len(train_losses) * 0.2)\n",
|
||
"plt.plot(range(start_idx, len(train_losses)), train_losses[start_idx:], label='训练Loss', color='blue', linewidth=2)\n",
|
||
"plt.plot(range(start_idx, len(val_losses)), val_losses[start_idx:], label='验证Loss', color='orange', linewidth=2)\n",
|
||
"plt.axvline(x=best_epoch-1, color='red', linestyle='--', label=f'最佳模型')\n",
|
||
"plt.xlabel('Epoch')\n",
|
||
"plt.ylabel('Loss')\n",
|
||
"plt.title('后期收敛细节')\n",
|
||
"plt.legend()\n",
|
||
"plt.grid(True, alpha=0.3)\n",
|
||
"\n",
|
||
"plt.tight_layout()\n",
|
||
"plt.savefig('training_loss.png', dpi=150)\n",
|
||
"plt.show()\n",
|
||
"\n",
|
||
"print(\"训练曲线已保存为 training_loss.png\")"
|
||
]
|
||
},
|
||
{
|
||
"cell_type": "markdown",
|
||
"id": "80b9c1c2",
|
||
"metadata": {},
|
||
"source": [
|
||
"## 8. 导出ONNX模型和归一化参数"
|
||
]
|
||
},
|
||
{
|
||
"cell_type": "code",
|
||
"execution_count": null,
|
||
"id": "c5b82ee2",
|
||
"metadata": {},
|
||
"outputs": [],
|
||
"source": [
|
||
"# 导出ONNX模型\n",
|
||
"model.eval()\n",
|
||
"model_cpu = model.cpu()\n",
|
||
"dummy_input = torch.randn(1, 10, 5)\n",
|
||
"\n",
|
||
"torch.onnx.export(\n",
|
||
" model_cpu,\n",
|
||
" dummy_input,\n",
|
||
" \"lstm_model.onnx\",\n",
|
||
" export_params=True,\n",
|
||
" opset_version=11,\n",
|
||
" input_names=['input'],\n",
|
||
" output_names=['output']\n",
|
||
")\n",
|
||
"\n",
|
||
"print(\"✓ ONNX模型已导出: lstm_model.onnx\")\n",
|
||
"\n",
|
||
"# 保存归一化参数\n",
|
||
"np.save('scaler_params.npy', {'min': scaler.data_min_, 'max': scaler.data_max_})\n",
|
||
"print(\"✓ 归一化参数已保存: scaler_params.npy\")\n",
|
||
"\n",
|
||
"# 输出MQL5格式的归一化参数\n",
|
||
"print(\"\\n=\" * 60)\n",
|
||
"print(\"归一化参数 (复制到LSTM_EA.mq5):\")\n",
|
||
"print(\"=\" * 60)\n",
|
||
"print(f\"double data_min[5] = {{{', '.join([f'{x:.5f}' for x in scaler.data_min_])}}};\")\n",
|
||
"print(f\"double data_max[5] = {{{', '.join([f'{x:.5f}' for x in scaler.data_max_])}}};\")\n",
|
||
"print(\"=\" * 60)"
|
||
]
|
||
},
|
||
{
|
||
"cell_type": "markdown",
|
||
"id": "cb21d2d4",
|
||
"metadata": {},
|
||
"source": [
|
||
"## 9. 训练总结"
|
||
]
|
||
},
|
||
{
|
||
"cell_type": "code",
|
||
"execution_count": null,
|
||
"id": "b6ac5bc9",
|
||
"metadata": {},
|
||
"outputs": [],
|
||
"source": [
|
||
"print(\"=\" * 60)\n",
|
||
"print(\"训练总结\")\n",
|
||
"print(\"=\" * 60)\n",
|
||
"print(f\"总训练轮数: {len(train_losses)} epochs\")\n",
|
||
"print(f\"最佳模型: Epoch {best_epoch}\")\n",
|
||
"print(f\"最佳验证Loss: {best_val_loss:.6f}\")\n",
|
||
"print(f\"最终训练Loss: {train_losses[-1]:.6f}\")\n",
|
||
"print(f\"最终验证Loss: {val_losses[-1]:.6f}\")\n",
|
||
"\n",
|
||
"loss_diff = abs(train_losses[-1] - val_losses[-1])\n",
|
||
"print(f\"\\nLoss差异: {loss_diff:.6f}\")\n",
|
||
"if loss_diff < 0.0001:\n",
|
||
" print(\"✓ 模型状态: 良好\")\n",
|
||
"elif loss_diff < 0.001:\n",
|
||
" print(\"⚠ 模型状态: 尚可，有轻微过拟合\")\n",
|
||
"else:\n",
|
||
" print(\"✗ 模型状态: 可能过拟合\")\n",
|
||
"print(\"=\" * 60)"
|
||
]
|
||
}
|
||
],
|
||
"metadata": {
|
||
"language_info": {
|
||
"name": "python"
|
||
}
|
||
},
|
||
"nbformat": 4,
|
||
"nbformat_minor": 5
|
||
}
|