基于Python和MQL5的特征工程（第二部分）：价格角度·进阶篇
（2/3）· 从价格差除以时间差到ONNX导出，拆解斜率策略在MT5里的真实误差陷阱
「建模前的数据切分与缩放陷阱」
做价格行为建模，第一步是把样本拆成训练与测试两块。这里用 train_test_split 且 shuffle=False，对 MT5 导出的时间序列而言很关键——乱序会把未来价格搅进过去，回测会虚高。
缩放器只能在训练集上 fit。若对整个数据集拟合，缩放参数会悄悄把测试集的分布泄露给模型，相当于用未来信息训过去，实盘外汇/贵金属这种高波动品种上容易直接崩。
候选模型池放了 12 个：从随机森林到 MLP。输入分三套测——纯 OHLC、仅斜率角度、全量特征。交叉验证用 5 折，评分取负均方误差均值。
结果里有个硬数据点：KNeighbors 在新特征下准确率提升 20%，是当时池子里最靓的；线性回归整体仍最优，但说明特征变换空间还没挖完，可能有别的方式反而拉低误差。
别把正态当圣经：StandardScaler 假设特征近似正态，OHLC 差值分布常有厚尾，实盘前最好对缩放后的训练集跑一遍分布检查。
from sklearn.model_selection class="kw">import train_test_split,cross_val_score from sklearn.preprocessing class="kw">import StandardScaler from sklearn.ensemble class="kw">import RandomForestRegressor, BaggingRegressor, GradientBoostingRegressor,AdaBoostRegressor from sklearn.svm class="kw">import LinearSVR from sklearn.linear_model class="kw">import LinearRegression, Ridge, Lasso, ElasticNet from sklearn.neighbors class="kw">import KNeighborsRegressor from sklearn.tree class="kw">import DecisionTreeRegressor from sklearn.neural_network class="kw">import MLPRegressor from sklearn.metrics class="kw">import mean_squared_error class="macro">#Let&class="macro">#x27;s split our data into train test splits train_data, test_data = train_test_split(data,test_size=class="num">0.5,shuffle=False) class="macro">#Scale the data scaler = StandardScaler() scaler.fit(train_data[all_inputs]) train_scaled= pd.DataFrame(scaler.transform(train_data[all_inputs]),columns=all_inputs) test_scaled = pd.DataFrame(scaler.transform(test_data[all_inputs]),columns=all_inputs) class="macro">#Create a dataframe to store our accuracy in training and testing columns = [ "Random Forest", "Bagging", "Gradient Boosting", "AdaBoost", "Linear SVR", "Linear Regression", "Ridge", "Lasso", "Elastic Net", "K Neighbors", "Decision Tree", "Neural Network" ] index = ["OHLC","Angle","All"] accuracy = pd.DataFrame(columns=columns,index=index) class="macro">#Store the models models = [ RandomForestRegressor(), BaggingRegressor(), GradientBoostingRegressor(), AdaBoostRegressor(), LinearSVR(), LinearRegression(), Ridge(), Lasso(), ElasticNet(), KNeighborsRegressor(), DecisionTreeRegressor(), MLPRegressor(hidden_layer_sizes=(class="num">4,class="num">6)) ] class="macro">#Cross validate the models class="macro">#First we have to iterate over the inputs for k in np.arange(class="num">0,len(cv_inputs)): current_inputs = cv_inputs[k] class="macro">#Then fit each model on that set of inputs for i in np.arange(class="num">0,len(models)): score = cross_val_score(models[i],train_scaled[current_inputs],train_data[target],cv=class="num">5,scoring="neg_mean_squared_error",n_jobs=-class="num">1) accuracy.iloc[k,i] = -score.mean()
◍ 用随机搜索给回归器挑参数
默认参数往往只是通用起点，放到具体品种节奏里未必够用。这里拿 K 近邻回归器做例子，用随机搜索在训练集上扫一遍超参空间，看能不能捞出更贴合样本的配置。 搜索跑完给出一组最优参数：weights 取 uniform、p 为 1、n_neighbors 是 10、leaf_size 设 100、algorithm 用 ball_tree。训练集上的最优得分约 0.7107（71% 上下），但这个数只说明拟合程度，不代表拿去实盘能复现。 真正要紧的是模型在遇到没见过的数据时表现如何，训练误差低反而可能藏着过拟合。外汇与贵金属波动受消息面扰动大，这类回归信号仅作概率参考，杠杆品种高风险，别直接当方向依据。 下面这段是搜索过程的代码，逐行拆一下方便你直接抄去改范围： from sklearn.model_selection import RandomizedSearchCV # 引入随机搜索交叉验证工具 model = KNeighborsRegressor(n_jobs=-1) # 建 KNN 回归器，n_jobs=-1 吃满 CPU tuner = RandomizedSearchCV(model, # 把模型丢进搜索器 { "n_neighbors": [2,3,4,5,6,7,8,9,10], # 邻居数候选 "weights": ["uniform","distance"], # 均匀或距离加权 "algorithm": ["auto","ball_tree","kd_tree","brute"], # 近邻查找算法 "leaf_size": [1,2,3,4,5,10,20,30,40,50,60,100,200,300,400,500,1000], # 树叶子尺寸 "p": [1,2] # 曼哈顿或欧氏距离 }, n_iter = 100, # 随机抽 100 组组合 n_jobs=-1, # 并行跑 cv=5 # 5 折交叉验证 ) tuner.fit(train_scaled.loc[:,all_inputs],train_data[target]) # 用缩放后的特征和目标列拟合 best_params_ # 查看最优参数组合 tuner.best_score_ # 查看交叉验证最优得分
from sklearn.model_selection class="kw">import RandomizedSearchCV model = KNeighborsRegressor(n_jobs=-class="num">1) tuner = RandomizedSearchCV(model, { "n_neighbors": [class="num">2,class="num">3,class="num">4,class="num">5,class="num">6,class="num">7,class="num">8,class="num">9,class="num">10], "weights": ["uniform","distance"], "algorithm": ["auto","ball_tree","kd_tree","brute"], "leaf_size": [class="num">1,class="num">2,class="num">3,class="num">4,class="num">5,class="num">10,class="num">20,class="num">30,class="num">40,class="num">50,class="num">60,class="num">100,class="num">200,class="num">300,class="num">400,class="num">500,class="num">1000], "p": [class="num">1,class="num">2] }, n_iter = class="num">100, n_jobs=-class="num">1, cv=class="num">5 ) tuner.fit(train_scaled.loc[:,all_inputs],train_data[target]) tuner.best_params_ tuner.best_score_
用默认模型当照妖镜
过拟合的本质是模型在训练集里啃下了噪声而非规律。一种低成本验证法：拿一个没做任何参数定制的默认模型，和你的自定义模型同台比武。 若自定义版连默认实例都赢不了，那基本可判定训练阶段已吞下无意义信息，泛化能力存疑。外汇与贵金属市场噪声极高，这类过拟合在实盘里会迅速失效，属典型高风险陷阱。 本例实测数值：默认模型得分 0.0009797322460441842，自定义模型 0.0009697248896608824。自定义略优，暂未落入过拟合区间，但优势极薄，仅约 1% 量级，需持续用样本外数据盯防。
class="macro">#Testing for over fitting model = KNeighborsRegressor(n_jobs=-class="num">1) custom_model = KNeighborsRegressor(n_jobs=-class="num">1,weights= &class="macro">#x27;uniform&class="macro">#x27;,p=class="num">1,n_neighbors= class="num">10,leaf_size= class="num">100,algorithm=&class="macro">#x27;ball_tree&class="macro">#x27;) model.fit(train_scaled.loc[:,all_inputs],train_data[target]) custom_model.fit(train_scaled.loc[:,all_inputs],train_data[target])
class="macro">#Testing for over fitting model = KNeighborsRegressor(n_jobs=-class="num">1) custom_model = KNeighborsRegressor(n_jobs=-class="num">1,weights= &class="macro">#x27;uniform&class="macro">#x27;,p=class="num">1,n_neighbors= class="num">10,leaf_size= class="num">100,algorithm=&class="macro">#x27;ball_tree&class="macro">#x27;) model.fit(train_scaled.loc[:,all_inputs],train_data[target]) custom_model.fit(train_scaled.loc[:,all_inputs],train_data[target])
「把训练好的模型塞进ONNX给MT5用」
想让Python里跑出来的AI模型在MQL5里复现，第一步不是直接导模型，而是先把价格特征的预处理参数固化下来。这里以USDZAR的M1数据为例，我们把所有输入列的均值和标准差分别写进两个CSV，后续MT5端做标准化时必须用同一组数值，否则推理结果会和训练时偏掉。 标准化本身很简单，就是减均值除标准差，让每列特征接近零均值单位方差。这一步在Python侧做完，模型才用skl2onnx转成ONNX——注意target_opset=12，低版本MT5的ONNX推理引擎可能认不了。 下面这段是实际导出流程的核心代码，包含均值/标准差落盘、标准化、以及模型转换保存。外汇与贵金属杠杆高、点差跳变频繁，USDZAR这种 exotic 对流动性风险更突出，模型在历史数据上表现好不代表实盘概率占优，导入后务必用MT5策略测试器先跑一遍样本外。
data.loc[:,all_inputs].mean().to_csv("USDZAR M1 MEAN.csv") data.loc[:,all_inputs].std().to_csv("USDZAR M1 STD.csv") data.loc[:,all_inputs] = ((data.loc[:,all_inputs] - data.loc[:,all_inputs].mean())/ data.loc[:,all_inputs].std()) 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 shape initial_type = [(&class="macro">#x27;float_input&class="macro">#x27;, FloatTensorType([class="num">1, len(all_inputs)]))] class="macro">#Fit the model on all the data we have custom_model.fit(data.loc[:,all_inputs],data.loc[:,"Target"]) class="macro">#Convert the model to ONNX format onnx_model = convert_sklearn(model, initial_types=initial_type,target_opset=class="num">12) class="macro">#Save the ONNX model onnx.save(onnx_model,"USDZAR M1 OHLC Angle.onnx")
◍ 把ONNX模型塞进EA的初始化骨架
做AI辅助交易，第一步是把训练好的ONNX模型加载进MT5的EA里，而不是在图表上手动看信号。上面这段源码用 #resource 把 USDZAR M1 OHLC Angle.onnx 以字节数组 onnx_buffer[] 的形式编进程序，运行时再喂给推理接口，省掉了外部文件路径依赖。
全局变量里有两个数组直接决定了模型输入的标准化方式：mean_values[5] 和 std_values[5]，数值分别是 {18.143698, 18.145870, 18.141644, 18.143724, 0.608216} 和 {0.112957, 0.113113, 0.112835, 0.112970, 0.580481}。这5个维度对应M1的OHLC角度特征，推理前必须用它们做减均值除标准差，否则模型输出会偏移。
仓位与风控参数写死在全局段：vol = 0.3 手、profit_target = 10（应该是报价点或小额度盈利目标，具体看品种小数点）、system_state = 0 表示初始空闲。外汇与贵金属杠杆高，0.3手在USDZAR这种宽幅品种上也可能在单根日线回调中吃掉远超10点浮盈的回撤，实盘前务必在策略测试器里改小 vol 验证。
指标句柄预留了 macd_handle 和两组快慢均线（美元指数与USDZAR各一对），但初始化函数里只加载ONNX和申请句柄，真正 CopyBuffer 抓数据是在每次新报价事件里做。你可以直接复制这段全局声明到自己的EA头部，把 onnx 文件名换成你导出的模型，就能在MT5里跑通加载环节。
class=class="str">"cmt">//+------------------------------------------------------------------+ class=class="str">"cmt">//| Slope 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 ONNX files | class=class="str">"cmt">//+------------------------------------------------------------------+ class="macro">#resource "\Files\USDZAR M1 OHLC Angle.onnx" as class="kw">const class="type">uchar onnx_buffer[]; class=class="str">"cmt">//+------------------------------------------------------------------+ class=class="str">"cmt">//| Libraries | class=class="str">"cmt">//+------------------------------------------------------------------+ class="macro">#include <Trade\Trade.mqh> CTrade Trade; class=class="str">"cmt">//+------------------------------------------------------------------+ class=class="str">"cmt">//| Global variables | class=class="str">"cmt">//+------------------------------------------------------------------+ class="type">class="kw">double mean_values[class="num">5] = {class="num">18.143698,class="num">18.145870,class="num">18.141644,class="num">18.143724,class="num">0.608216}; class="type">class="kw">double std_values[class="num">5] = {class="num">0.112957,class="num">0.113113,class="num">0.112835,class="num">0.112970,class="num">0.580481}; class="type">long onnx_model; class="type">int macd_handle; class="type">int usd_ma_slow,usd_ma_fast; class="type">int usd_zar_slow,usd_zar_fast; class="type">class="kw">double macd_s[],macd_m[],usd_zar_s[],usd_zar_f[],usd_s[],usd_f[]; class="type">class="kw">double bid,ask; class="type">class="kw">double vol = class="num">0.3; class="type">class="kw">double profit_target = class="num">10; class="type">int system_state = class="num">0; vectorf model_forecast = vectorf::Zeros(class="num">1); class=class="str">"cmt">//+------------------------------------------------------------------+
EA 初始化与逐笔更新的骨架
把 ONNX 模型接进 MT5 专家顾问，第一步是在 OnInit 里把模型和跨品种句柄一次性加载完。代码先调 onnx_load()，失败直接返回 INIT_FAILED，避免后续空跑；随后用 iMACD 抓 EURUSD 当前周期的 12/26/9 参数柱，再分别向 USDZAR 日线和自定义 DXY_Z4 日线要 20 与 60 周期 EMA——这几条线是后面模型特征的基础。 OnDeinit 里挨个 Release 模型和指标句柄，不然切换周期或重编译会漏内存。OnTick 的逻辑很薄：先 update() 刷最新 bid/ask 与 MACD 缓冲，再 model_predict() 取模型输出；无持仓就 find_entry()，有持仓就 manage_positions()，把进出场决策完全拆到别的函数里。 update() 内部用 SymbolInfoDouble 取当前品种买卖价，CopyBuffer(macd_handle,0,0,1,macd_m) 只拉最新一根 MACD 主线值。实盘里若 USDZAR 或 DXY_Z4 报价中断，iMA 返回的句柄会是 INVALID_HANDLE，EA 不会崩但特征会 stale，建议在 update 里加一句句柄有效性检查再喂模型。外汇与贵金属杠杆高，这类跨品种 ONNX 策略回测顺滑不代表 live 能复现，先开 MT5 用策略测试器跑一周 tick 级验证再上资金。
class=class="str">"cmt">//| Expert initialization function | class=class="str">"cmt">//+------------------------------------------------------------------+ class="type">int OnInit() { class=class="str">"cmt">//--- Load the ONNX file if(!onnx_load()) { class=class="str">"cmt">//--- We failed to load the ONNX file class="kw">return(INIT_FAILED); } class=class="str">"cmt">//--- Load the MACD Indicator macd_handle = iMACD("EURUSD",PERIOD_CURRENT,class="num">12,class="num">26,class="num">9,PRICE_CLOSE); usd_zar_fast = iMA("USDZAR",PERIOD_D1,class="num">20,class="num">0,MODE_EMA,PRICE_CLOSE); usd_zar_slow = iMA("USDZAR",PERIOD_D1,class="num">60,class="num">0,MODE_EMA,PRICE_CLOSE); usd_ma_fast = iMA("DXY_Z4",PERIOD_D1,class="num">20,class="num">0,MODE_EMA,PRICE_CLOSE); usd_ma_slow = iMA("DXY_Z4",PERIOD_D1,class="num">60,class="num">0,MODE_EMA,PRICE_CLOSE); class=class="str">"cmt">//--- Everything went fine class="kw">return(INIT_SUCCEEDED); } 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">//--- Release the handles don&class="macro">#x27;t need OnnxRelease(onnx_model); IndicatorRelease(macd_handle); IndicatorRelease(usd_zar_fast); IndicatorRelease(usd_zar_slow); IndicatorRelease(usd_ma_fast); IndicatorRelease(usd_ma_slow); } class=class="str">"cmt">//+------------------------------------------------------------------+ class=class="str">"cmt">//| Expert tick function | class=class="str">"cmt">//+------------------------------------------------------------------+ class="type">void OnTick() { class=class="str">"cmt">//--- Update our market data update(); class=class="str">"cmt">//--- Get a prediction from our model model_predict(); if(PositionsTotal() == class="num">0) { find_entry(); } if(PositionsTotal() > class="num">0) { manage_positions(); } } class=class="str">"cmt">//+------------------------------------------------------------------+ class=class="str">"cmt">//+------------------------------------------------------------------+ class=class="str">"cmt">//| Update our market data | class=class="str">"cmt">//+------------------------------------------------------------------+ class="type">void update(class="type">void) { bid = SymbolInfoDouble(Symbol(),SYMBOL_BID); ask = SymbolInfoDouble(Symbol(),SYMBOL_ASK); CopyBuffer(macd_handle,class="num">0,class="num">0,class="num">1,macd_m);