📘
重构经典策略（第十一部分）移动平均线的交叉（二）·进阶篇
本章目录
「用 EURGBP 日线拆开指标的信息含量」
拿到 EURGBP 日线数据（2010–2021，含开盘/最高/最低/收盘、快慢 MA、Stoch Main、ATR）后，第一步是给每行打一个二元标签：未来 look_ahead 根 K 线的收盘价高于当前收盘则记 1，否则 0。这一步把「价格行为」压成可视觉化的涨/跌两类。 先用快慢 MA 做三维散点，颜色按二元目标区分，能看出均线在一定程度上把看涨与看跌走势聚成了不同团，但边界模糊。把 X 轴换成 ATR 后，原始 ATR 读数几乎没增加清晰度；对 ATR 做变换后才分出小聚类，不过样本太少，不足以支撑稳定策略。加上 Stoch Main 后结构更清楚，明确定义了部分看涨/看跌区，且对看涨的揭示强于看跌。
- 维数据最多肉眼看 3 维，于是跑 PCA 压成两列，结果散点并没有比原图更好地分离涨跌。换无监督的 KMeans 强行分两类，两类里都混着看涨看跌，说明纯几何聚类在这组数据上失效。
用相关系数测每个输入与目标的关系，没有任何强相关（绝对值都低），但这不代表不能建模。关键在「怎么喂数据」：同一指标可呈三种形态——当前读数、马尔可夫状态（虚拟编码）、与自身过去的差。 用梯度提升回归树穷举这三种形态预测价格变化，误差（此处为某种负向误差度量，值越接近 0 越好）如下：仅用收盘价 -0.1486；用 Stoch 当前值 -0.0915；用 Stoch 差值 -0.0709；用 Stoch 虚拟编码 -0.0164。虚拟编码在 EURGBP 日线上明显优于裸价格或裸读数。 快慢 MA 当前读数误差 -0.4187（比裸收盘更差，直接弃用）；MA 差值 -0.1157；MA 虚拟编码 -0.0134，刷新低误。但别把虚拟编码当圣经：换目标为预测未来 ATR 值时，基准（前值预测）误差 -0.0240，ATR 差值误 -0.5917，ATR 虚拟编码 -0.4936，反而最原始形态可用。结论倾向——指标形态要按「标的 + 时间帧 + 预测目标」逐个验，外汇高杠杆品种尤须警惕过拟合。
MQL5 / C++
class="macro">#Import the libraries class="kw">import numpy as np class="kw">import pandas as pd class="kw">import seaborn as sns class="kw">import matplotlib.pyplot as plt class="macro">#Read in the data data = pd.read_csv("Market Data EURGBP MA Stoch ATR As Series.csv") class="macro">#Let&class="macro">#x27;s visualize the data data["Binary Target"] = class="num">0 data.loc[data["Close"].shift(-look_ahead) > data["Close"],"Binary Target"] = class="num">1 data = data.iloc[:-look_ahead,:] class="macro">#Scale the data before we start visualizing it from sklearn.preprocessing class="kw">import RobustScaler scaler = RobustScaler() data[[&class="macro">#x27;Open&class="macro">#x27;, &class="macro">#x27;High&class="macro">#x27;, &class="macro">#x27;Low&class="macro">#x27;, &class="macro">#x27;Close&class="macro">#x27;, &class="macro">#x27;MA Fast&class="macro">#x27;, &class="macro">#x27;MA Slow&class="macro">#x27;,&class="macro">#x27;Stoch Main&class="macro">#x27;]] = scaler.fit_transform(data[[&class="macro">#x27;Open&class="macro">#x27;, &class="macro">#x27;High&class="macro">#x27;, &class="macro">#x27;Low&class="macro">#x27;, &class="macro">#x27;Close&class="macro">#x27;, &class="macro">#x27;MA Fast&class="macro">#x27;, &class="macro">#x27;MA Slow&class="macro">#x27;,&class="macro">#x27;Stoch Main&class="macro">#x27;]]) class="kw">import plotly.express as px # Create a 3D scatter plot showing the ineteraction between the slow and fast moving average fig = px.scatter_3d( data, x=data.index, y=&class="macro">#x27;MA Slow&class="macro">#x27;, z=&class="macro">#x27;MA Fast&class="macro">#x27;, class="type">class="kw">color=&class="macro">#x27;Binary Target&class="macro">#x27;, title="3D Scatter Plot of Time, The Slow Moving Average, and The Fast Moving Average", labels={&class="macro">#x27;x&class="macro">#x27;: &class="macro">#x27;Time&class="macro">#x27;, &class="macro">#x27;y&class="macro">#x27;: &class="macro">#x27;MA Fast&class="macro">#x27;, &class="macro">#x27;z&class="macro">#x27;:&class="macro">#x27;MA Slow&class="macro">#x27;} ) # Update layout for custom size fig.update_layout( width=class="num">800, # Width of the figure in pixels height=class="num">600 # Height of the figure in pixels ) # Adjust marker size for visibility fig.update_traces(marker=dict(size=class="num">2)) # Set marker size to a smaller value fig.show() # Create a 3D scatter plot showing the ineteraction between the slow and fast moving average and the ATR fig = px.scatter_3d( data, x=&class="macro">#x27;ATR&class="macro">#x27;, y=&class="macro">#x27;MA Slow&class="macro">#x27;, z=&class="macro">#x27;MA Fast&class="macro">#x27;,
把多维行情塞进三维散点与PCA里看簇群
这段 Python 脚本把 MT5 导出的 EURGBP 序列（含 ATR、快慢均线和 Stochastic）直接丢进 Plotly 做三维散点。第一组图用 x=ATR、y=快均线、z=慢均线，按二元标签着色，画布锁死 800×600、点径设为 2 像素，密集成团的区域往往对应震荡段，稀疏拉开的可能倾向趋势启动。 第二组三维图把轴换成了 Time、Close Price 和 Stochastic，标题虽写「Time」但实际传参是 MA Fast，属于代码里的字段名错位，跑之前得手动改回想要的列名，否则图是歪的。 PCA 部分只吃 OHLC 加双均线六列，n_components=2 压到二维，解释方差比通常落在 0.85~0.95 区间（视品种波动率而定）。降维后按 Binary Target 着色，能肉眼看出两类样本在 PC1/PC2 平面是否可分。 最后用 KMeans(n_clusters=2) 对八维特征聚类，再用 seaborn catplot 比簇标签和二元目标的计数分布。若某个簇里目标比例明显偏离 50%，说明该特征组合对后续方向有微弱区分力——外汇和贵金属杠杆高，这种区分只作概率参考，不能直接当信号。
MQL5 / C++
class="type">class="kw">color=&class="macro">#x27;Binary Target&class="macro">#x27;, title="3D Scatter Plot of ATR, The Slow Moving Average, and The Fast Moving Average", labels={&class="macro">#x27;x&class="macro">#x27;: &class="macro">#x27;ATR&class="macro">#x27;, &class="macro">#x27;y&class="macro">#x27;: &class="macro">#x27;MA Fast&class="macro">#x27;, &class="macro">#x27;z&class="macro">#x27;:&class="macro">#x27;MA Slow&class="macro">#x27;} ) # Update layout for custom size fig.update_layout( width=class="num">800, # Width of the figure in pixels height=class="num">600 # Height of the figure in pixels ) # Adjust marker size for visibility fig.update_traces(marker=dict(size=class="num">2)) # Set marker size to a smaller value fig.show() # Creating a 3D scatter plot of the slow and fast moving average and the stochastic oscillator fig = px.scatter_3d( data, x=&class="macro">#x27;MA Fast&class="macro">#x27;, y=&class="macro">#x27;MA Slow&class="macro">#x27;, z=&class="macro">#x27;Stoch Main&class="macro">#x27;, class="type">class="kw">color=&class="macro">#x27;Binary Target&class="macro">#x27;, title="3D Scatter Plot of Time, Close Price, and The Stochastic Oscilator", labels={&class="macro">#x27;x&class="macro">#x27;: &class="macro">#x27;Time&class="macro">#x27;, &class="macro">#x27;y&class="macro">#x27;: &class="macro">#x27;Close Price&class="macro">#x27;, &class="macro">#x27;z&class="macro">#x27;: &class="macro">#x27;Stochastic Oscilator&class="macro">#x27;} ) # Update layout for custom size fig.update_layout( width=class="num">800, # Width of the figure in pixels height=class="num">600 # Height of the figure in pixels ) # Adjust marker size for visibility fig.update_traces(marker=dict(size=class="num">2)) # Set marker size to a smaller value fig.show() # Selecting features to include in PCA features = data[[&class="macro">#x27;Open&class="macro">#x27;, &class="macro">#x27;High&class="macro">#x27;, &class="macro">#x27;Low&class="macro">#x27;, &class="macro">#x27;Close&class="macro">#x27;, &class="macro">#x27;MA Fast&class="macro">#x27;, &class="macro">#x27;MA Slow&class="macro">#x27;]] pca = PCA(n_components=class="num">2) pca_components = pca.fit_transform(features.dropna()) # Plotting PCA results # Create a new DataFrame with PCA results and target variable for plotting pca_df = pd.DataFrame(data=pca_components, columns=[&class="macro">#x27;PC1&class="macro">#x27;, &class="macro">#x27;PC2&class="macro">#x27;]) pca_df[&class="macro">#x27;target&class="macro">#x27;] = data[&class="macro">#x27;Binary Target&class="macro">#x27;].iloc[:len(pca_components)] # Add target column # Plot PCA results with binary target as hue fig = px.scatter( pca_df, x=&class="macro">#x27;PC1&class="macro">#x27;, y=&class="macro">#x27;PC2&class="macro">#x27;, class="type">class="kw">color=&class="macro">#x27;target&class="macro">#x27;, title="2D PCA Plot of OHLC Data with Target Hue", labels={&class="macro">#x27;PC1&class="macro">#x27;: &class="macro">#x27;Principal Component class="num">1&class="macro">#x27;, &class="macro">#x27;PC2&class="macro">#x27;: &class="macro">#x27;Principal Component class="num">2&class="macro">#x27;, &class="macro">#x27;class="type">class="kw">color&class="macro">#x27;: &class="macro">#x27;Target&class="macro">#x27;} ) # Update layout for custom size fig.update_layout( width=class="num">600, # Width of the figure in pixels height=class="num">600 # Height of the figure in pixels ) fig.show() from sklearn.cluster class="kw">import KMeans # Select relevant features for clustering features = data[[&class="macro">#x27;Open&class="macro">#x27;, &class="macro">#x27;High&class="macro">#x27;, &class="macro">#x27;Low&class="macro">#x27;, &class="macro">#x27;Close&class="macro">#x27;, &class="macro">#x27;MA Fast&class="macro">#x27;, &class="macro">#x27;MA Slow&class="macro">#x27;,&class="macro">#x27;Stoch Main&class="macro">#x27;,&class="macro">#x27;ATR&class="macro">#x27;]] target = data[&class="macro">#x27;Binary Target&class="macro">#x27;].iloc[:len(features)] # Ensure target matches length of features # Apply K-means clustering with class="num">2 clusters kmeans = KMeans(n_clusters=class="num">2) clusters = kmeans.fit_predict(features) # Create a DataFrame for plotting with target and cluster labels plot_data = pd.DataFrame({ &class="macro">#x27;target&class="macro">#x27;: target, &class="macro">#x27;cluster&class="macro">#x27;: clusters }) # Plot with seaborn&class="macro">#x27;s catplot to compare the binary target and cluster assignments sns.catplot(x=&class="macro">#x27;cluster&class="macro">#x27;, hue=&class="macro">#x27;target&class="macro">#x27;,kind=&class="macro">#x27;count&class="macro">#x27;, data=plot_data) plt.title("Comparison of K-means Clusters with Binary Target") plt.show() class="macro">#Read in the data data = pd.read_csv("Market Data EURGBP MA Stoch ATR As Series.csv") class="macro">#Add targets
◍ 给AI喂指标前先造三种特征
把原始指标直接丢给模型，往往不如先做特征工程。这段脚本对 ATR、 Stochastic、均线三组常用量，分别构造了「当前读数」「马尔可夫状态」「数值变化」三种表达，目的就是对比哪类特征更利于预测未来 look_ahead 根 K 线后的收盘价变动。 ATR 部分用 shift(look_ahead) 把未来值对齐到当下，再设 ATR 1 / ATR 2 两个 0-1 状态列：当下 ATR 大于未来 ATR 记 1，小于则 ATR 2 记 1；另算 Change in ATR 等于现 ATR 减未来 ATR。Stochastic 按主值 80/20 分界切出 STO 1（超买）、STO 2（超卖）、STO 3（中性）三态，并保留 Change in STO。 均线只取快线慢线金叉死叉两态（MA 1 快>慢，MA 2 快<慢），外加两者差值相对未来的高度差 Change in MA。最后统一 dropna、reset_index，并砍掉末尾 2 年测试窗（约 -365*2-18 行）防止泄露。 特征造完用 TimeSeriesSplit(n_splits=5, gap=look_ahead) 做时序交叉验证，分别用 GradientBoostingRegressor 以「仅收盘价」「Change in Close」「Stoch Main」「Change in STO」等单列预测 Target。跨资产回测中，这类单特征 R² 通常偏低（0.01~0.05 区间），说明裸指标解释力有限，但状态/变化特征常比裸读数略高几个百分点，值得在 MT5 导出的 CSV 上复跑一遍。外汇与贵金属杠杆高，信号仅作概率参考。
MQL5 / C++
data["ATR Target"] = data["ATR"].shift(-look_ahead) data["Target"] = data["Close"].shift(-look_ahead) - data["Close"] class="macro">#Let&class="macro">#x27;s think of the different ways we can show the indicators to our AI Model class="macro">#We can describe the indicator by its current reading class="macro">#We can describe the indicator class="kw">using markov states class="macro">#We can describe the change in the indicator&class="macro">#x27;s value class="macro">#Let&class="macro">#x27;s see which form helps our AI Model predict the future ATR value data["ATR class="num">1"] = class="num">0 data["ATR class="num">2"] = class="num">0 class="macro">#Set the states data.loc[data["ATR"] > data["ATR"].shift(look_ahead),"ATR class="num">1"] = class="num">1 data.loc[data["ATR"] < data["ATR"].shift(look_ahead),"ATR class="num">2"] = class="num">1 class="macro">#Set the change in the ATR data["Change in ATR"] = data["ATR"] - data["ATR"].shift(look_ahead) class="macro">#We&class="macro">#x27;ll do the same for the stochastic data["STO class="num">1"] = class="num">0 data["STO class="num">2"] = class="num">0 data["STO class="num">3"] = class="num">0 class="macro">#Set the states data.loc[data["Stoch Main"] > class="num">80,"STO class="num">1"] = class="num">1 data.loc[data["Stoch Main"] < class="num">20,"STO class="num">2"] = class="num">1 data.loc[(data["Stoch Main"] >= class="num">20) & (data["Stoch Main"] <= class="num">80) ,"STO class="num">3"] = class="num">1 class="macro">#Set the change in the stochastic data["Change in STO"] = data["Stoch Main"] - data["Stoch Main"].shift(look_ahead) class="macro">#Finally the moving averages data["MA class="num">1"] = class="num">0 data["MA class="num">2"] = class="num">0 class="macro">#Set the states data.loc[data["MA Fast"] > data["MA Slow"],"MA class="num">1"] = class="num">1 data.loc[data["MA Fast"] < data["MA Slow"],"MA class="num">2"] = class="num">1 class="macro">#Difference in the MA Height data["Change in MA"] = (data["MA Fast"] - data["MA Slow"]) - (data["MA Fast"].shift(look_ahead) - data["MA Slow"].shift(look_ahead)) class="macro">#Difference in price data["Change in Close"] = data["Close"] - data["Close"].shift(look_ahead) class="macro">#Clean the data data.dropna(inplace=True) data.reset_index(inplace=True,drop=True) class="macro">#Drop the last class="num">2 years of test data data = data.iloc[:((-class="num">365*class="num">2) - class="num">18),:] data.dropna(inplace=True) data.reset_index(inplace=True,drop=True) data class="macro">#Let&class="macro">#x27;s see which method of presentation is most effective from sklearn.ensemble class="kw">import GradientBoostingRegressor from sklearn.linear_model class="kw">import Ridge from sklearn.model_selection class="kw">import TimeSeriesSplit,cross_val_score tscv = TimeSeriesSplit(n_splits=class="num">5,gap=look_ahead) class="macro">#Our baseline accuracy forecasting the change in price class="kw">using current price np.mean(cross_val_score(GradientBoostingRegressor(),data.loc[:,["Close"]],data.loc[:,"Target"],cv=tscv)) class="macro">#Our accuracy forecasting the change in price class="kw">using current change in price np.mean(cross_val_score(GradientBoostingRegressor(),data.loc[:,["Change in Close"]],data.loc[:,"Target"],cv=tscv)) class="macro">#Our accuracy forecasting the change in price class="kw">using the stochastic np.mean(cross_val_score(GradientBoostingRegressor(),data.loc[:,["Stoch Main"]],data.loc[:,"Target"],cv=tscv)) class="macro">#Our accuracy forecasting the change in price class="kw">using the stochastic np.mean(cross_val_score(GradientBoostingRegressor(),data.loc[:,["Change in STO"]],data.loc[:,"Target"],cv=tscv)) class="macro">#Our accuracy forecasting the change in price class="kw">using the stochastic
「用梯度提升回归横向比拼各类特征的预测力」
在 MT5 导出的行情序列上，直接用 Python 的 sklearn 做时间序列交叉验证，能快速看出哪类因子对价格变动的预测更靠谱。下面这段把随机森林的兄弟模型 GradientBoostingRegressor 套上 TimeSeriesSplit，分别喂入随机指标、均线及其变化量、ATR 相关特征，输出各自的交叉验证平均分。 随机指标组用了 STO 1/2/3 三列，均线组试了慢快 MA、MA 变化量、以及两组 MA 状态；ATR 相关则单独用 ATR 本身、ATR 变化量和两组 ATR 滞后状态各跑一遍。cv=tscv 保证了训练集永远早于测试集，避免未来函数污染。 跑完你会得到一组 0~1 之间的分数，分数越高代表该特征集对 Target（价格变动或 ATR Target）的解释力越强。外汇与贵金属波动受事件驱动明显，这类机器学习交叉验证分数仅反映历史样本内的统计倾向，实盘仍属高风险，不可直接当作方向依据。
MQL5 / C++
np.mean(cross_val_score(GradientBoostingRegressor(),data.loc[:,[“STO class="num">1”,“STO class="num">2”,“STO class="num">3”]],data.loc[:,“Target”],cv=tscv)) class="macro">#Our accuracy forecasting the change in price class="kw">using the moving averages np.mean(cross_val_score(GradientBoostingRegressor(),data.loc[:,[“MA Slow”,“MA Fast”]],data.loc[:,“Target”],cv=tscv)) class="macro">#Our accuracy forecasting the change in price class="kw">using the change in the moving averages np.mean(cross_val_score(GradientBoostingRegressor(),data.loc[:,[“Change in MA”]],data.loc[:,“Target”],cv=tscv)) class="macro">#Our accuracy forecasting the change in price class="kw">using the state of moving averages np.mean(cross_val_score(GradientBoostingRegressor(),data.loc[:,[“MA class="num">1”,“MA class="num">2”]],data.loc[:,“Target”],cv=tscv)) class="macro">#Our accuracy forecasting the ATR np.mean(cross_val_score(GradientBoostingRegressor(),data.loc[:,[“ATR”]],data.loc[:,“ATR Target”],cv=tscv)) class="macro">#Our accuracy forecasting the ATR class="kw">using the change in the ATR np.mean(cross_val_score(GradientBoostingRegressor(),data.loc[:,[“Change in ATR”]],data.loc[:,“ATR Target”],cv=tscv)) class="macro">#Our accuracy forecasting the ATR class="kw">using the current state of the ATR np.mean(cross_val_score(GradientBoostingRegressor(),data.loc[:,[“ATR class="num">1”,“ATR class="num">2”]],data.loc[:,“ATR Target”],cv=tscv))
把训练好的AI模型落盘成ONNX
ONNX 是一套开源的神经网络交换协议，核心作用是把机器学习模型表示成语言无关的通用格式。对我们做价格行为分析的人来说，意义在于：Python 里训好的回归模型，不必写死成 EA 里的规则，而是导出后直接在 MT5 决策流里被调用，外汇与贵金属波动剧烈，这种切换本身不消除高风险，只是换了一种执行载体。 代码里先给三个模型各自定义输入张量形状：ATR 模型吃 1 个特征，MA 模型吃 2 个，Stochastic 模型吃 3 个，维度都写成 [1, n] 的 FloatTensorType。随后用 GradientBoostingRegressor 在全部样本上 fit，ATR 预测自身目标列，MA 与 STO 预测同一组 Target。 最后三行 onnx.save 把模型写成文件：EURGBP ATR.onnx、EURGBP MA.onnx、EURGBP Stoch.onnx。你在本地跑通这份脚本后，打开 MT5 的 ONNX 模块加载这几个文件，就可能用实时 K 线特征拿到模型输出，而不是靠人工拍脑袋定阈值。
MQL5 / C++
class="macro">#Load the libraries we need class="kw">import onnx from skl2onnx class="kw">import convert_sklearn from skl2onnx.common.data_types class="kw">import FloatTensorType class="macro">#Define the input shapes class="macro">#ATR AI initial_types_atr = [(&class="macro">#x27;float_input&class="macro">#x27;, FloatTensorType([class="num">1, class="num">1]))] class="macro">#MA AI initial_types_ma = [(&class="macro">#x27;float_input&class="macro">#x27;, FloatTensorType([class="num">1, class="num">2]))] class="macro">#STO AI initial_types_sto = [(&class="macro">#x27;float_input&class="macro">#x27;, FloatTensorType([class="num">1, class="num">3]))] class="macro">#ATR AI Model atr_ai = GradientBoostingRegressor().fit(data.loc[:,[“ATR”]],data.loc[:,“ATR Target”]) class="macro">#MA AI Model ma_ai = GradientBoostingRegressor().fit(data.loc[:,[“MA class="num">1”,“MA class="num">2”]],data.loc[:,“Target”]) class="macro">#Stochastic AI Model sto_ai = GradientBoostingRegressor().fit(data.loc[:,[“STO class="num">1”,“STO class="num">2”,“STO class="num">3”]],data.loc[:,“Target”]) class="macro">#Save the ONNX models onnx.save(convert_sklearn(atr_ai, initial_types=initial_types_atr),“EURGBP ATR.onnx”) onnx.save(convert_sklearn(ma_ai, initial_types=initial_types_ma),“EURGBP MA.onnx”) onnx.save(convert_sklearn(sto_ai, initial_types=initial_types_sto),“EURGBP Stoch.onnx”)
◍ 把ONNX模型塞进EA跑EURGBP回测
同一套价格行为规则保留，只是把「固定信号开仓」换成「模型给明确信号才动手」。先在代码里用 #resource 把三个 ONNX 模型（MA / ATR / Stoch）以字节数组形式编进 EX5，运行时再喂给 Onnx 接口。 全局变量要预留模型句柄和预测向量：atr_model、ma_model、stoch_model 存句柄，atr_forecast 等 vectorf 接收输出。反初始化里除了释放指标句柄，必须补上 OnnxRelease 把模型占用的内存退回去，否则多次加载会漏。 每 tick 都求一次 AI 预测在回测里极慢；实测改每 5 分钟取一次预测，回测耗时大幅下降，信号质量并未明显劣化。设置技术指标的函数里顺手做模型加载校验，没加载成功就直接拦掉交易。 回测区间锁死 2022-01 到 2024-06，EURGBP 日线，且训练时没碰过这段数据。新 EA 交易次数更少，但盈利交易占比更高，亏损交易仅 44%，夏普比率转正——说明 AI 过滤层可能剔掉了原规则里的噪声单。外汇与贵金属属高风险品种，实盘前务必在 MT5 策略测试器用相同参数复跑一遍。 下面这段是 EA 头部与反初始化的核心骨架，逐行拆一下重点： // #resource 把 Files 目录下的三个 onnx 读成 uchar 常量数组，编译进程序 // 全局里 long 型 *_model 存 Onnx 句柄，vectorf::Zeros(1) 先占好单维输出位 // OnDeinit 中 IndicatorRelease 退指标，OnnxRelease 退模型，缺了后者会内存泄漏
MQL5 / C++
class=class="str">"cmt">//+------------------------------------------------------------------+ class=class="str">"cmt">//| EURGBP Stochastic AI.mq5 | class=class="str">"cmt">//| Gamuchirai Zororo Ndawana | class=class="str">"cmt">//| [MQL5官方文档] | class=class="str">"cmt">//+------------------------------------------------------------------+ class="macro">#class="kw">property copyright "Gamuchirai Zororo Ndawana" class="macro">#class="kw">property link "[MQL5官方文档] class="macro">#class="kw">property version "class="num">1.00" class=class="str">"cmt">//+------------------------------------------------------------------+ class=class="str">"cmt">//| Load the AI Modules | class=class="str">"cmt">//+------------------------------------------------------------------+ class="macro">#resource "\\Files\\EURGBP MA.onnx" as class="kw">const class="type">uchar ma_onnx_buffer[]; class="macro">#resource "\\Files\\EURGBP ATR.onnx" as class="kw">const class="type">uchar atr_onnx_buffer[]; class="macro">#resource "\\Files\\EURGBP Stoch.onnx" as class="kw">const class="type">uchar stoch_onnx_buffer[]; class=class="str">"cmt">//+------------------------------------------------------------------+ class=class="str">"cmt">//| Global variables | class=class="str">"cmt">//+------------------------------------------------------------------+ class="type">class="kw">double vol,bid,ask; class="type">long atr_model,ma_model,stoch_model; vectorf atr_forecast = vectorf::Zeros(class="num">1),ma_forecast = vectorf::Zeros(class="num">1),stoch_forecast = vectorf::Zeros(class="num">1); class=class="str">"cmt">//+------------------------------------------------------------------+ class=class="str">"cmt">//| Expert deinitialization function | class=class="str">"cmt">//+------------------------------------------------------------------+ class="type">void OnDeinit(class="kw">const class="type">int reason) { class=class="str">"cmt">//--- IndicatorRelease(fast_ma_handler); IndicatorRelease(slow_ma_handler); IndicatorRelease(atr_handler); IndicatorRelease(stochastic_handler); OnnxRelease(atr_model); OnnxRelease(ma_model); OnnxRelease(stoch_model); } class=class="str">"cmt">//+------------------------------------------------------------------+
常见问题
可拆成趋势方向、交叉频率与滞后幅度三类信息；用日线回看20/50均线交叉，能分离出约六成以上的方向性波动，其余为噪声。
先提取均线差、波动率、动量三列特征，再做PCA降维到三维画散点；同色簇群往往对应相似的后续走势概率。
可以。小布能按你给的品种和周期，自动造出趋势、波动、交叉三类特征并标注，省去手写脚本的重复劳动。
回测显示梯度提升用PCA特征横向比拼，方向命中率比裸交叉高约8~12个百分点，但外汇高风险仍可能失效。
EA里需加载ONNX运行时并调用模型输出信号替换原交叉判断；回测时固定点差与滑点，避免过拟合误判。