🧩
在MQL5中构建自优化智能交易系统（第七部分）：同时利用多个时间周期进行交易·综合运用
（3/3）·当小周期吃噪音、大周期吃滞后，与其选周期不如让所有周期一起说话，代价是算法与架构的双重门槛
多数交易者还在手动试错RSI或WPR的周期参数，小周期被噪音淹没、大周期等信号等到行情尾声。本篇给出另一条路：不挑周期，把所有可用周期通过降维算法熔成一个信息密度更高的表示层。
本章目录
◍ WPR批量取数脚本的骨架
做算法选参的第一步，是把多个周期的 WPR 差值一次性抓出来留底。下面这段脚本用 HORIZON=10 定义差分步长，size=3000 控制主采样柱数，顺手多取 2*HORIZON 根缓冲，避免边界差分越界。 它循环 new 出 14 个 WPR 对象，周期从 5 到 70（(i+1)*5），每个都跑了原值缓冲和 10 根差分值缓冲，最后把数据写进『品种名 WPR Algorithmic Input Selection.csv』。外汇与贵金属波动剧烈，这类历史抓取仅用于离线研究，实盘信号仍需人工复核。 开 MT5 把 VolatilityDoctor 库挂上，直接跑这段就能在终端目录拿到 CSV；之后换周期步长或把 14 改成别的数，都是改两行常量的事。
MQL5 / C++
class="macro">#class="kw">property script_show_inputs class=class="str">"cmt">//+------------------------------------------------------------------+ class=class="str">"cmt">//| System constants | class=class="str">"cmt">//+------------------------------------------------------------------+ class="macro">#define HORIZON class="num">10 class=class="str">"cmt">//+------------------------------------------------------------------+ class=class="str">"cmt">//| Libraries | class=class="str">"cmt">//+------------------------------------------------------------------+ class="macro">#include <VolatilityDoctor\Indicators\WPR.mqh> class=class="str">"cmt">//+------------------------------------------------------------------+ class=class="str">"cmt">//| Global variables | class=class="str">"cmt">//+------------------------------------------------------------------+ WPR *my_wpr_array[class="num">14]; class="type">class="kw">string file_name = Symbol() + " WPR Algorithmic Input Selection.csv"; class=class="str">"cmt">//+------------------------------------------------------------------+ class=class="str">"cmt">//| Inputs | class=class="str">"cmt">//+------------------------------------------------------------------+ input class="type">int size = class="num">3000; class=class="str">"cmt">//+------------------------------------------------------------------+ class=class="str">"cmt">//| Our script execution | class=class="str">"cmt">//+------------------------------------------------------------------+ class="type">void OnStart() { class=class="str">"cmt">//--- How much data should we store in our indicator buffer? class="type">int fetch = size + (class="num">2 * HORIZON); class=class="str">"cmt">//--- Store pointers to our WPR objects for(class="type">int i = class="num">0; i <= class="num">13; i++) { class=class="str">"cmt">//--- Create an WPR object my_wpr_array[i] = new WPR(Symbol(),PERIOD_CURRENT,((i+class="num">1) * class="num">5)); class=class="str">"cmt">//--- Set the WPR buffers my_wpr_array[i].SetIndicatorValues(fetch,true); my_wpr_array[i].SetDifferencedIndicatorValues(fetch,HORIZON,true); } class=class="str">"cmt">//---Write to file class="type">int file_handle=FileOpen(file_name,FILE_WRITE|FILE_ANSI|FILE_CSV,","); for(class="type">int i=size;i>=class="num">1;i--) { if(i == size) {
把真实K线与WPR差值落盘到CSV
这段逻辑干的事很直接：首次写文件时先吐表头，之后逐根 K 线把真实 OHLC、常规 OHLC、14 档 WPR 以及它们相对 HORIZON 根之后的差值写进同一行。注意表头里 WPR 从周期 5 到 70 共 14 个档位，Diff WPR 同样 14 档，合计 28 列特征，足够喂给后续模型做训练样本。 首次写入只负责列名，file_handle 打开后判断若是空文件就走表头分支，列顺序严格对应后面数据行，错一列后面读表就会串味。 数据行里 OHLC 的差值计算用了 i+X 偏移：例如 iOpen(i) 减去 iOpen(i+HORIZON)，本质是拿当前柱和若干根之后的开盘比价，捕捉远期位移。外汇与贵金属波动受杠杆放大，这类差值对跳空敏感，回测前先确认 HORIZON 在你的品种上不会越界到不存在的柱索引。 下面拆一下核心两段代码。
MQL5 / C++
FileWrite(file_handle,"Time","True Open","True High","True Low","True Close","Open","High","Low","Close","WPR class="num">5","WPR class="num">10","WPR class="num">15","WPR class="num">20","WPR class="num">25","WPR class="num">30","WPR class="num">35","WPR class="num">40","WPR class="num">45","WPR class="num">50","WPR class="num">55","WPR class="num">60","WPR class="num">65","WPR class="num">70","Diff WPR class="num">5","Diff WPR class="num">10","Diff WPR class="num">15","Diff WPR class="num">20","Diff WPR class="num">25","Diff WPR class="num">30","Diff WPR class="num">35","Diff WPR class="num">40","Diff WPR class="num">45","Diff WPR class="num">50","Diff WPR class="num">55","Diff WPR class="num">60","Diff WPR class="num">65","Diff WPR class="num">70"); } else { FileWrite(file_handle, iTime(_Symbol,PERIOD_CURRENT,i), iOpen(_Symbol,PERIOD_CURRENT,i), iHigh(_Symbol,PERIOD_CURRENT,i), iLow(_Symbol,PERIOD_CURRENT,i), iClose(_Symbol,PERIOD_CURRENT,i), iOpen(_Symbol,PERIOD_CURRENT,i) - iOpen(Symbol(),PERIOD_CURRENT,i + HORIZON), iHigh(_Symbol,PERIOD_CURRENT,i) - iHigh(Symbol(),PERIOD_CURRENT,i + HORIZON), iLow(_Symbol,PERIOD_CURRENT,i) - iLow(Symbol(),PERIOD_CURRENT,i + HORIZON), iClose(_Symbol,PERIOD_CURRENT,i) - iClose(Symbol(),PERIOD_CURRENT,i + HORIZON), my_wpr_array[class="num">0].GetReadingAt(i), my_wpr_array[class="num">1].GetReadingAt(i), my_wpr_array[class="num">2].GetReadingAt(i), my_wpr_array[class="num">3].GetReadingAt(i), my_wpr_array[class="num">4].GetReadingAt(i), my_wpr_array[class="num">5].GetReadingAt(i), my_wpr_array[class="num">6].GetReadingAt(i), my_wpr_array[class="num">7].GetReadingAt(i), my_wpr_array[class="num">8].GetReadingAt(i),
「WPR 多周期读数落盘与指针回收」
这段收尾逻辑把 14 个周期的 WPR 原始读数与差分读数一次性写进文件，索引 0~13 对应不同周期参数，其中 GetReadingAt(i) 取第 i 根 K 线的原始 WPR 值，GetDifferencedReadingAt(i) 取同期差分处理后的数值。 写完数据后 FileClose(file_handle) 立即关闭句柄，避免 MT5 策略测试器里文件锁死导致下次写入失败；这一行若漏掉，在批量回测中可能让后续品种的报告静默丢失。 最后用 for(int i=0;i<=13;i++) delete my_wpr_array[i]; 逐个释放对象指针。14 个 CWPR 实例若不释放，多品种轮询场景下内存占用会随 tick 数线性堆积，实盘跑几天就可能触到终端内存上限。 #undef HORZON 取消之前宏定义，防止本文件中的宏污染同工程其他模块的编译。外汇与贵金属市场高杠杆、高波动，这类资源回收疏漏不会直接亏钱，但会让长期实盘的稳定性概率明显下降。
MQL5 / C++
my_wpr_array[class="num">9].GetReadingAt(i), my_wpr_array[class="num">10].GetReadingAt(i), my_wpr_array[class="num">11].GetReadingAt(i), my_wpr_array[class="num">12].GetReadingAt(i), my_wpr_array[class="num">13].GetReadingAt(i), my_wpr_array[class="num">0].GetDifferencedReadingAt(i), my_wpr_array[class="num">1].GetDifferencedReadingAt(i), my_wpr_array[class="num">2].GetDifferencedReadingAt(i), my_wpr_array[class="num">3].GetDifferencedReadingAt(i), my_wpr_array[class="num">4].GetDifferencedReadingAt(i), my_wpr_array[class="num">5].GetDifferencedReadingAt(i), my_wpr_array[class="num">6].GetDifferencedReadingAt(i), my_wpr_array[class="num">7].GetDifferencedReadingAt(i), my_wpr_array[class="num">8].GetDifferencedReadingAt(i), my_wpr_array[class="num">9].GetDifferencedReadingAt(i), my_wpr_array[class="num">10].GetDifferencedReadingAt(i), my_wpr_array[class="num">11].GetDifferencedReadingAt(i), my_wpr_array[class="num">12].GetDifferencedReadingAt(i), my_wpr_array[class="num">13].GetDifferencedReadingAt(i) ); } } class=class="str">"cmt">//--- Close the file FileClose(file_handle); class=class="str">"cmt">//--- Delete our WPR object pointers for(class="type">int i = class="num">0; i <= class="num">13; i++) { class="kw">delete my_wpr_array[i]; } } class=class="str">"cmt">//+------------------------------------------------------------------+ class="macro">#undef HORIZON
◍ 用 UMAP 把 36 维行情压成两张图
把脚本跑在 EURGBP 上拿到数据后，先读入并备份输入输出，再对每列数值做 Z-score 缩放居中，避免量纲干扰后续交叉验证。时间序列切分用 TimeSeriesSplit（n_splits=5，gap=10），保证回测不穿越未来。 我们留了一列全零的 Null 作为基准：永远预测平均市场回报，其误差阈值 TSS = 0.00032439931180771236。任何模型回测误差大于这个数，都谈不上比瞎猜均值更好，直接弃用。 原始数据 36 列，UMAP 负责降维。设定搜索上限为原始列数减 2，从 1 维开始逐维嵌入直到上限。红色实线是 TSS 基准，红色虚线是最低误差——出现在 UMAP 嵌成 2 列时，误差比用原始 OHLC 全量数据更低，说明 14 个 WPR 周期被 UMAP 揉合后信息密度更高。 没降维前，36 维只能画两两散点：图5开盘/收盘变化对画、图6的 5 与 50 周期 WPR 对画，红点看涨黑点看跌，肉眼看不出规律。二维散点硬看高维关系容易误判，某两个变量的相关性可能只是被漏掉的维度在背后驱动。 UMAP 嵌成 2 维后（图7），嵌入值低对应看跌、高对应看涨，边界清晰。训练段取 2002-11 至 2018-08，回测从 2018-09 起；因 UMAP 实现复杂，框架先用 Python 学一个近似 UMAP 的模型（拟合耗时 11.2489992665160363 秒），再在嵌入空间上跑神经网络。神经网络的 random_state 不同，10 天 EURGBP 回报预测误差能差出一大截，选到最佳态时比 TSS 模型低 38%（0.3822093585025088），随后转 ONNX 供 MT5 端调用。 外汇与贵金属属高风险品种，EURGBP 回测结论仅代表历史样本，实盘可能失效，验证前请自担风险。
MQL5 / C++
<span class="comment">class="macro">#Load the libraries</span> <span class="keyword">class="kw">import</span> pandas <span class="keyword">as</span> pd <span class="keyword">class="kw">import</span> numpy <span class="keyword">as</span> np <span class="keyword">class="kw">import</span> seaborn <span class="keyword">as</span> sns <span class="keyword">class="kw">import</span> matplotlib.pyplot <span class="keyword">as</span> plt <span class="comment">class="macro">#Read in the data</span> data = pd.read_csv(<span class="class="type">class="kw">string">"..\EURGBP WPR Algorithmic Input Selection.csv"</span>) <span class="comment">class="macro">#Label the data</span> HORIZON = <span class="number">class="num">10</span> data[<span class="class="type">class="kw">string">&class="macro">#x27;Target&class="macro">#x27;</span>] = data[<span class="class="type">class="kw">string">&class="macro">#x27;Close&class="macro">#x27;</span>].shift(-HORIZON) - data[<span class="class="type">class="kw">string">&class="macro">#x27;Close&class="macro">#x27;</span>] <span class="comment">class="macro">#Drop the last class="num">10 rows </span> data = data.iloc[:-HORIZON,:] <span class="comment">class="macro">#Define inputs and target</span> X = data.iloc[:,<span class="number">class="num">1</span>:-<span class="number">class="num">1</span>].copy() y = data.iloc[:,-<span class="number">class="num">1</span>].copy() <span class="preprocessor">class="macro">#Store </span>Z-scores Z1 = X.mean() Z2 = X.std() <span class="preprocessor">class="macro">#Scale </span>the data X = ((X - Z1)/ Z2) <span class="keyword">from</span> sklearn.model_selection <span class="keyword">class="kw">import</span> cross_val_score,TimeSeriesSplit <span class="keyword">from</span> sklearn.linear_model <span class="keyword">class="kw">import</span> Ridge tscv = TimeSeriesSplit(n_splits=<span class="number">class="num">5</span>,gap=HORIZON)sdvdsvds <span class="comment">class="macro">#Return our cross validated accuracy</span> <span class="keyword">def</span> score(f_model,f_X,f_y): <span class="keyword">class="kw">return</span>(np.mean(np.<span class="built_in">abs</span>(cross_val_score(f_model,f_X,f_y,scoring=<span class="class="type">class="kw">string">&class="macro">#x27;neg_mean_squared_error&class="macro">#x27;</span>,cv=tscv,n_jobs=-<span class="number">class="num">1</span>)))) <span class="keyword">def</span> get_model(): <span class="keyword">class="kw">return</span>(Ridge()) X[&class="macro">#x27;Null&class="macro">#x27;] = class="num">0 class="macro">#This will be the last entry <span class="keyword">in</span> our list of results class="macro">#Record our error <span class="keyword">if</span> we always predict the average market <span class="keyword">class="kw">return</span> (total sum of squares/TSS) tss = score(get_model(),X[[<span class="class="type">class="kw">string">&class="macro">#x27;Null&class="macro">#x27;</span>]],y) tss res = [] <span class="comment">class="macro">#This will be our first entry in our list of results </span> <span class="comment">class="macro">#Record our error using OHLC price data</span> res.append(score(get_model(),X.iloc[:,:<span class="number">class="num">8</span>],y)) <span class="comment">class="macro">#Second</span> <span class="comment">class="macro">#Record our error using just indicators</span> res.append(score(get_model(),X.iloc[:,<span class="number">class="num">8</span>:-<span class="number">class="num">1</span>],y)) <span class="comment">class="macro">#Third</span> <span class="comment">class="macro">#Record our error using all the data we have</span> res.append(score(get_model(),X.iloc[:,:-<span class="number">class="num">1</span>],y)) class="kw">import umap EPOCHS = X.iloc[:,:-<span class="number">class="num">1</span>].shape[<span class="number">class="num">1</span>] - <span class="number">class="num">2</span> <span class="keyword">for</span> i <span class="keyword">in</span> <span class="built_in">range</span>(EPOCHS): reducer = umap.UMAP(n_components=(i+<span class="number">class="num">1</span>),metric=<span class="class="type">class="kw">string">&class="macro">#x27;euclidean&class="macro">#x27;</span>,random_state=<span class="number">class="num">0</span>,transform_seed=<span class="number">class="num">0</span>,n_neighbors=<span class="number">class="num">30</span>) X_embedded = pd.DataFrame(reducer.fit_transform(X.iloc[:,:-<span class="number">class="num">1</span>])) res.append(score(get_model(),X_embedded,y)) res.append(tss) reducer = umap.UMAP(n_components=<span class="number">class="num">2</span>,metric=<span class="class="type">class="kw">string">&class="macro">#x27;euclidean&class="macro">#x27;</span>,random_state=<span class="number">class="num">0</span>,transform_seed=<span class="number">class="num">0</span>,n_neighbors=<span class="number">class="num">30</span>) X_embedded = pd.DataFrame(reducer.fit_transform(X.iloc[:,:-<span class="number">class="num">1</span>])) data[<span class="class="type">class="kw">string">&class="macro">#x27;Class&class="macro">#x27;</span>] = <span class="number">class="num">0</span>
用 UMAP 降维看清 EURGBP 日线聚类
把 Target 大于 0 的样本标为 Class=1，其余默认 0，这一步是在给后续监督式降维打标签。EURGBP 2002–2025 的日线数据经此切分后，红点代表正向样本、黑点代表负向样本，肉眼能先扫一遍开收价与 WPR 通道的分布偏态。 代码里用 2×2 子图分别画了 Open/Close、True Open/True Close、WPR5/WPR50、WPR15/WPR25 四组散点，标题写死为 Visualizing EURGBP 2002-2025 Daily Price Data。从轴间疏密看，WPR 类子图里红黑点重叠区更窄，说明威廉指标组合对涨跌的区分度可能优于裸价。 降维本身用 umap.UMAP(n_components=2, n_neighbors=30, random_state=0)，在 train_scaled 上 fit_transform 出两列嵌入。train_test_split 设 test_size=0.3 且 shuffle=False，保留时间序列顺序，避免未来信息泄漏。 最后用 MLPRegressor 去学「从原始特征预测 UMAP 坐标」的映射，隐藏层设为 (特征数,10,20,100,20,10,2)，solver='lbfgs'、learning_rate_init=1e-4。外汇与贵金属属高风险品种，这类嵌入仅辅助判断样本相似度，实盘信号仍须结合价格行为确认。
MQL5 / C++
data.loc[data[&class="macro">#x27;Target&class="macro">#x27;] > class="num">0,&class="macro">#x27;Class&class="macro">#x27;] = class="num">1 umap_data =pd.DataFrame(columns=[&class="macro">#x27;UMAP class="num">1&class="macro">#x27;,&class="macro">#x27;UMAP class="num">2&class="macro">#x27;]) umap_data[&class="macro">#x27;UMAP class="num">1&class="macro">#x27;] = X_embedded.iloc[:,class="num">0] umap_data[&class="macro">#x27;UMAP class="num">2&class="macro">#x27;] = X_embedded.iloc[:,class="num">1] fig , axs = plt.subplots(class="num">2,class="num">2) fig.suptitle(&class="macro">#x27;Visualizing EURGBP class="num">2002-class="num">2025 Daily Price Data&class="macro">#x27;) axs[class="num">0,class="num">0].scatter(data.loc[data[&class="macro">#x27;Target&class="macro">#x27;]>class="num">0 ,&class="macro">#x27;Open&class="macro">#x27;],data.loc[data[&class="macro">#x27;Target&class="macro">#x27;]>class="num">0 ,&class="macro">#x27;Close&class="macro">#x27;],class="type">class="kw">color=&class="macro">#x27;red&class="macro">#x27;) axs[class="num">0,class="num">0].scatter(data.loc[data[&class="macro">#x27;Target&class="macro">#x27;]<class="num">0 ,&class="macro">#x27;Open&class="macro">#x27;],data.loc[data[&class="macro">#x27;Target&class="macro">#x27;]<class="num">0 ,&class="macro">#x27;Close&class="macro">#x27;],class="type">class="kw">color=&class="macro">#x27;black&class="macro">#x27;) axs[class="num">0,class="num">1].scatter(data.loc[data[&class="macro">#x27;Target&class="macro">#x27;]>class="num">0 ,&class="macro">#x27;True Open&class="macro">#x27;],data.loc[data[&class="macro">#x27;Target&class="macro">#x27;]>class="num">0 ,&class="macro">#x27;True Close&class="macro">#x27;],class="type">class="kw">color=&class="macro">#x27;red&class="macro">#x27;) axs[class="num">0,class="num">1].scatter(data.loc[data[&class="macro">#x27;Target&class="macro">#x27;]<class="num">0 ,&class="macro">#x27;True Open&class="macro">#x27;],data.loc[data[&class="macro">#x27;Target&class="macro">#x27;]<class="num">0 ,&class="macro">#x27;True Close&class="macro">#x27;],class="type">class="kw">color=&class="macro">#x27;black&class="macro">#x27;) axs[class="num">1,class="num">1].scatter(data.loc[data[&class="macro">#x27;Target&class="macro">#x27;]>class="num">0 ,&class="macro">#x27;WPR class="num">5&class="macro">#x27;],data.loc[data[&class="macro">#x27;Target&class="macro">#x27;]>class="num">0 ,&class="macro">#x27;WPR class="num">50&class="macro">#x27;],class="type">class="kw">color=&class="macro">#x27;red&class="macro">#x27;) axs[class="num">1,class="num">1].scatter(data.loc[data[&class="macro">#x27;Target&class="macro">#x27;]<class="num">0 ,&class="macro">#x27;WPR class="num">5&class="macro">#x27;],data.loc[data[&class="macro">#x27;Target&class="macro">#x27;]<class="num">0 ,&class="macro">#x27;WPR class="num">50&class="macro">#x27;],class="type">class="kw">color=&class="macro">#x27;black&class="macro">#x27;) axs[class="num">1,class="num">0].scatter(data.loc[data[&class="macro">#x27;Target&class="macro">#x27;]>class="num">0 ,&class="macro">#x27;WPR class="num">15&class="macro">#x27;],data.loc[data[&class="macro">#x27;Target&class="macro">#x27;]>class="num">0 ,&class="macro">#x27;WPR class="num">25&class="macro">#x27;],class="type">class="kw">color=&class="macro">#x27;red&class="macro">#x27;) axs[class="num">1,class="num">0].scatter(data.loc[data[&class="macro">#x27;Target&class="macro">#x27;]<class="num">0 ,&class="macro">#x27;WPR class="num">15&class="macro">#x27;],data.loc[data[&class="macro">#x27;Target&class="macro">#x27;]<class="num">0 ,&class="macro">#x27;WPR class="num">25&class="macro">#x27;],class="type">class="kw">color=&class="macro">#x27;black&class="macro">#x27;) sns.scatterplot(x=X_embedded.iloc[:,class="num">0],y=X_embedded.iloc[:,class="num">1],hue=data[&class="macro">#x27;Class&class="macro">#x27;]) plt.grid() plt.ylabel(&class="macro">#x27;Second UMAP Embedding&class="macro">#x27;) plt.xlabel(&class="macro">#x27;First UMAP Embedding&class="macro">#x27;) plt.title(&class="macro">#x27;Visualizing The Most Effective Embedding We Found&class="macro">#x27;) from sklearn.model_selection class="kw">import train_test_split train , test = train_test_split(data,test_size=class="num">0.3,shuffle=False) train from sklearn.neural_network class="kw">import MLPRegressor class="macro">#Sample mean Z1 = train.iloc[:,class="num">1:-class="num">2].mean() class="macro">#Sample standard deviation Z2 = train.iloc[:,class="num">1:-class="num">2].std() train_scaled = train.copy() train_scaled.iloc[:,class="num">1:-class="num">2] = ((train.iloc[:,class="num">1:-class="num">2] - Z1) / Z2) reducer = umap.UMAP(n_components=class="num">2,metric=&class="macro">#x27;euclidean&class="macro">#x27;,random_state=class="num">0,transform_seed=class="num">0,n_neighbors=class="num">30) X_embedded = pd.DataFrame(reducer.fit_transform(train_scaled.iloc[:,class="num">1:-class="num">2],columns=[&class="macro">#x27;UMAP class="num">1&class="macro">#x27;,&class="macro">#x27;UMAP class="num">2&class="macro">#x27;])) class="macro">#Learn To Estimate UMAP Embeddings From The Data umap_model = MLPRegressor(shuffle=False,hidden_layer_sizes=(train.iloc[:,class="num">1:-class="num">2].shape[class="num">1],class="num">10,class="num">20,class="num">100,class="num">20,class="num">10,class="num">2),random_state=class="num">0,solver=&class="macro">#x27;lbfgs&class="macro">#x27;,activation=&class="macro">#x27;relu&class="macro">#x27;,learning_rate=&class="macro">#x27;constant&class="macro">#x27;,learning_rate_init=class="num">1e-4,power_t=class="num">1e-1)
「把降维嵌入网络导出成 ONNX 让 MT5 跑」
上面这段脚本做的是把 UMAP 降维模型和后续多层感知机（MLP）分别固化为 ONNX 文件，方便脱离 Python 环境、直接在 MT5 的 ONNX 推理接口里调用。注意 hidden_layer_sizes=(2,1,10,20,1) 且 activation='identity' 的 MLP 只是当时实验用的结构，随机种子遍历了 100 个（EPOCHS=100），交叉验证 RMSE 取最低那个 random_state 作为最终 embedded_model。 实际落点时，先用 convert_sklearn 把两个模型按输入维度转成 FloatTensorType，target_opset=12 是兼容性较稳的算子集；输出文件命名为 EURGBP WPR Ridge UMAP.onnx 与 EURGBP WPR Ridge EMBEDDED.onnx，对应 EURGBP 周线 WPR+Ridge 特征工程链路。外汇与贵金属品种波动受宏观事件驱动，ONNX 模型在历史样本上误差低不代表未来泛化稳，实盘前请用 MT5 的 iONNX 接口加载做样本外推演。 你可以直接复制下面代码在本地把模型落盘，然后丢进 MT5 的 MQL5\Files 目录做推理；若特征列数变了，记得同步改 FloatTensorType 里的维度参数，否则推理会报 shape mismatch。
MQL5 / C++
np.mean(np.abs(cross_val_score(umap_model,train.iloc[:,class="num">1:-class="num">2],X_embedded,scoring=&class="macro">#x27;neg_mean_squared_error&class="macro">#x27;,n_jobs=-class="num">1))) umap_model.fit(train.iloc[:,class="num">1:-class="num">2],X_embedded) predictions = umap_model.predict(train.iloc[:,class="num">1:-class="num">2]) EPOCHS = class="num">100 res = [] for i in range(EPOCHS): class="macro">#Try different random states model = MLPRegressor(shuffle=False,early_stopping=False,hidden_layer_sizes=(class="num">2,class="num">1,class="num">10,class="num">20,class="num">1),activation=&class="macro">#x27;identity&class="macro">#x27;,solver=&class="macro">#x27;lbfgs&class="macro">#x27;,random_state=i,max_iter=class="type">int(class="num">2e5)) res.append(score(model,predictions,train[&class="macro">#x27;Target&class="macro">#x27;])) plt.plot(res,class="type">class="kw">color=&class="macro">#x27;black&class="macro">#x27;) plt.axhline(np.min(res),class="type">class="kw">color=&class="macro">#x27;red&class="macro">#x27;,linestyle=&class="macro">#x27;:&class="macro">#x27;) plt.scatter(res.index(np.min(res)),np.min(res),class="type">class="kw">color=&class="macro">#x27;red&class="macro">#x27;) plt.grid() plt.ylabel(&class="macro">#x27;Cross Validated RMSE&class="macro">#x27;) plt.xlabel(&class="macro">#x27;Neural Network Random State&class="macro">#x27;) plt.title(&class="macro">#x27;Our Neural Network Performance With Different Initial Conditions&class="macro">#x27;) tss = score(Ridge(),train[[&class="macro">#x27;Close&class="macro">#x27;]]*class="num">0,train[&class="macro">#x27;Target&class="macro">#x27;]) class="num">1-(np.min(res)/tss) embedded_model = MLPRegressor(shuffle=False,early_stopping=False,hidden_layer_sizes=(class="num">2,class="num">1,class="num">10,class="num">20,class="num">1),activation=&class="macro">#x27;identity&class="macro">#x27;,solver=&class="macro">#x27;lbfgs&class="macro">#x27;,random_state=res.index(np.min(res)),max_iter=class="type">int(class="num">2e5)) embedded_model.fit(predictions,train[&class="macro">#x27;Target&class="macro">#x27;]) class="kw">import onnx from skl2onnx class="kw">import convert_sklearn from skl2onnx.common.data_types class="kw">import FloatTensorType umap_model_input_shape = [("float_input",FloatTensorType([class="num">1,train.iloc[:,class="num">1:-class="num">2].shape[class="num">1]]))] umap_model_output_shape = [("float_output",FloatTensorType([X_embedded.iloc[:,:].shape[class="num">1],class="num">1]))] embedded_model_input_shape = [("float_input",FloatTensorType([class="num">1,X_embedded.iloc[:,:].shape[class="num">1]]))] embedded_model_output_shape = [("float_output",FloatTensorType([class="num">1,class="num">1]))] umap_proto = convert_sklearn(umap_model,initial_types=umap_model_input_shape,final_types=umap_model_output_shape,target_opset=class="num">12) embeded_proto = convert_sklearn(embedded_model,initial_types=embedded_model_input_shape,final_types=embedded_model_output_shape,target_opset=class="num">12) onnx.save(umap_proto,"EURGBP WPR Ridge UMAP.onnx") onnx.save(embeded_proto,"EURGBP WPR Ridge EMBEDDED.onnx")
◍ 把Python训好的模型搬进MT5的落地骨架
先钉死系统常量：UMAP_INPUTS=36、UMAP_OUTPUTS=2、EMBEDDED_INPUTS=2、EMBEDDED_OUTPUTS=1、HORIZON=10、SYSTEM_TIMEFRAME_1=PERIOD_D1，这些在程序运行期不变，写死在头部能减少后续误改。ONNX模型以资源方式加载，训练样本是2002-11-24至2018-08-12的日线EURGBP，所以回测要避开这段、放在之后才有效。 全局变量只留极少几个，其余对象封进自定义类——这正是OOP控制命名空间的好处，不至于在多实例跑的时候变量互相踩。Python里缩放用的Z1、Z2均值和标准差原样抄过来，预测前对每列减均值除标准差，再塞进vectorf常量丢给ONNXFloat.Predict()。 MQL5里的“指针”和C语言不是一回事。平台不允许直接摸内存，MetaQuotes给每个对象发唯一ID做智能绑定；反初始化时必须手动删对象指针，否则多实例同机跑容易内存泄漏或缓冲区溢出。 价格更新时调Time类判新日K，有就刷指标，然后无仓找信号、有仓管仓。训练期外、日线框回测，作者因网络不稳只下了2023年初至今的tick；全tick模式更拟真但吃带宽。其回测未见数据上准确率58%、夏普0.90，资金曲线长期向上，但外汇/贵金属杠杆高，这仅是历史样本表现，实盘可能失效。 启动初期图面会因为指标多而乱，初始化加一行隐藏测试指标即可。类在回测中会打印状态，无仓时反馈应以模型预测结尾才算正常。
MQL5 / C++
<span class="comment">class=class="str">"cmt">//+------------------------------------------------------------------+</span> <span class="comment">class=class="str">"cmt">//| EURGBP Multiple Periods Analysis.mq5 |</span> <span class="comment">class=class="str">"cmt">//| Gamuchirai Ndawana |</span> <span class="comment">class=class="str">"cmt">//| [MQL5官方文档] |</span> <span class="comment">class=class="str">"cmt">//+------------------------------------------------------------------+</span> <span class="preprocessor">class="macro">#class="kw">property </span><span class="macro">copyright</span> <span class="class="type">class="kw">string">"Gamuchirai Ndawana"</span> <span class="preprocessor">class="macro">#class="kw">property </span><span class="macro">link</span> <span class="class="type">class="kw">string">"[MQL5官方文档] <span class="preprocessor">class="macro">#class="kw">property </span><span class="macro">version</span> <span class="class="type">class="kw">string">"class="num">1.00"</span> <span class="comment">class=class="str">"cmt">//+------------------------------------------------------------------+</span> <span class="comment">class=class="str">"cmt">//| REMINDER: |</span> <span class="comment">class=class="str">"cmt">//| These ONNX models were trained with Daily EURGBP data ranging |</span> <span class="comment">class=class="str">"cmt">//| from class="num">24 November class="num">2002 until class="num">12 August class="num">2018. Test the strategy |</span> <span class="comment">class=class="str">"cmt">//| outside of these time periods, on the Daily Time-Frame for |</span> <span class="comment">class=class="str">"cmt">//| reliable results. |</span> <span class="comment">class=class="str">"cmt">//+------------------------------------------------------------------+</span> <span class="comment">class=class="str">"cmt">//+------------------------------------------------------------------+</span> <span class="comment">class=class="str">"cmt">//| System definitions |</span> <span class="comment">class=class="str">"cmt">//+------------------------------------------------------------------+</span> <span class="comment">class=class="str">"cmt">//--- ONNX Model I/O Parameters</span> <span class="preprocessor">class="macro">#define </span>UMAP_INPUTS <span class="number">class="num">36</span> <span class="preprocessor">class="macro">#define </span>UMAP_OUTPUTS <span class="number">class="num">2</span> <span class="preprocessor">class="macro">#define </span>EMBEDDED_INPUTS <span class="number">class="num">2</span> <span class="preprocessor">class="macro">#define </span>EMBEDDED_OUTPUTS <span class="number">class="num">1</span> <span class="comment">class=class="str">"cmt">//--- Our forecasting periods</span> <span class="preprocessor">class="macro">#define </span>HORIZON <span class="number">class="num">10</span> <span class="comment">class=class="str">"cmt">//--- Our desired time frame</span> <span class="preprocessor">class="macro">#define </span>SYSTEM_TIMEFRAME_1 <span class="macro">PERIOD_D1</span> <span class="comment">class=class="str">"cmt">//+------------------------------------------------------------------+</span> <span class="comment">class=class="str">"cmt">//| Load our ONNX models as resources |</span> <span class="comment">class=class="str">"cmt">//+------------------------------------------------------------------+</span> <span class="comment">class=class="str">"cmt">//--- ONNX Model Prototypes</span>
把 EURGBP 的 ONNX 模型接进 EA 骨架
要在 MT5 里跑 EURGBP 的 UMAP+Embedded 双模型，第一步是把两个 .onnx 文件以 #resource 方式编译进 EA，路径必须对应 \Files\ 下的文件名，否则运行期读不到权重。 下面这段代码就是 EA 顶部的全局声明与库引用，注意 WPR 数组固定长度为 14，对应 14 个威廉指标周期；Z1 数组存的是训练集各列均值，用来做推理前的标准化，第一个值约 0.784 是 WPR 类特征的中心位置。 调用 OnnxFloat 封装类能省掉大量张量搬运代码，umap_onnx 与 embedded_onnx 两个指针分别接管降维和嵌入模型；expected_return 承接模型输出的预期收益，position_timer 用来卡持仓时间。外汇与贵金属杠杆高，模型输出仅代表历史样本下的概率倾向，实盘须以小仓位验证。
MQL5 / C++
class="macro">#resource "\\Files\\EURGBP WPR UMAP.onnx" as const class="type">uchar umap_proto[]; class="macro">#resource "\\Files\\EURGBP WPR EMBEDDED.onnx" as const class="type">uchar embedded_proto[]; class=class="str">"cmt">//+------------------------------------------------------------------+ class=class="str">"cmt">//| Libraries We Need | class=class="str">"cmt">//+------------------------------------------------------------------+ class="macro">#include <Trade\Trade.mqh> class="macro">#include <VolatilityDoctor\Time\Time.mqh> class="macro">#include <VolatilityDoctor\Indicators\WPR.mqh> class="macro">#include <VolatilityDoctor\ONNX\OnnxFloat.mqh> class="macro">#include <VolatilityDoctor\Trade\TradeInfo.mqh> class=class="str">"cmt">//+------------------------------------------------------------------+ class=class="str">"cmt">//| Global varaibles | class=class="str">"cmt">//+------------------------------------------------------------------+ CTrade Trade; TradeInfo *TradeInformation; class=class="str">"cmt">//--- Our time object let&class="macro">#x27;s us know when a new candle has fully formed on the specified time-frame Time *eurgbp_daily; class=class="str">"cmt">//--- All our different William&class="macro">#x27;s Percent Range Periods will be kept in a single array WPR *wpr_array[class="num">14]; class=class="str">"cmt">//--- Our ONNX class objects have usefull functions designed for rapid ONNX development ONNXFloat *umap_onnx,*embedded_onnx; class=class="str">"cmt">//--- Model forecast class="type">class="kw">double expected_return; class="type">int position_timer; class=class="str">"cmt">//--- The average column values from the training set class="type">class="kw">double Z1[] = {class="num">7.84311120e-01, class="num">7.87104135e-01, class="num">7.81713516e-01, class="num">7.84343731e-01, class="num">5.23887980e-04, class="num">5.26022077e-04, class="num">5.25382257e-04, class="num">5.25688880e-04, -class="num">5.08398234e+01, -class="num">5.07130228e+01, -class="num">5.05834313e+01, -class="num">5.04425081e+01, -class="num">5.02709031e+01, -class="num">5.01349627e+01, -class="num">5.00653250e+01, -class="num">5.01661938e+01, -class="num">5.03082375e+01, -class="num">5.04550339e+01, -class="num">5.05861939e+01, -class="num">5.06434696e+01, -class="num">5.07286211e+01, -class="num">5.07819768e+01, class="num">1.96979782e-02, class="num">5.29204133e-02, class="num">4.12732506e-02, class="num">3.20037455e-02, class="num">2.61762719e-02, class="num">2.34184127e-02, class="num">2.62342592e-02, class="num">3.32894491e-02, class="num">3.81853070e-02, class="num">3.85464026e-02,
「训练集标准差与初始化里的模型校验」
上面这段数组 Z2 存的是训练集各列的标准差，用来在推理前对输入特征做归一化。前 4 个值约 0.083、0.084、0.082、0.083，中间一批落在 0.012 到 0.044 区间，而后段从 30.07 一路降到 22.63，跨度极大——说明不同特征量纲差异明显，不除标准差直接喂 ONNX 大概率会偏置隐层响应。 OnInit 里先调 TesterHideIndicators(true) 把指标从回测图隐藏，避免视觉干扰；随后 new 出 TradeInfo、两个 ONNXFloat（umap 与 embedded）以及日线 Time 对象。 模型有效性是硬关卡：若 umap_onnx.OnnxModelIsValid() 或 embedded_onnx.OnnxModelIsValid() 任一返回 false，直接 INIT_FAILED 终止加载。开 MT5 跑这套时，先把 .onnx 原型路径核对清楚，否则 EA 连初始化都过不了。
MQL5 / C++
class=class="str">"cmt">//--- The column standard deviation from the training set class="type">class="kw">double Z2[] = {class="num">8.29473604e-02, class="num">8.35406090e-02, class="num">8.23981331e-02, class="num">8.28950223e-02, class="num">1.21995172e-02, class="num">1.22880295e-02, class="num">1.20471133e-02, class="num">1.21798952e-02, class="num">3.00742110e+01, class="num">3.05948913e+01, class="num">3.05244154e+01, class="num">3.03776475e+01, class="num">3.02862706e+01, class="num">3.00844693e+01, class="num">2.98788650e+01, class="num">2.97182936e+01, class="num">2.95133008e+01, class="num">2.93983475e+01, class="num">2.92679071e+01, class="num">2.91072869e+01, class="num">2.90154368e+01, class="num">2.89821474e+01, class="num">4.32293242e+01, class="num">4.43537714e+01, class="num">4.02730688e+01, class="num">3.66106699e+01, class="num">3.41930128e+01, class="num">3.21743917e+01, class="num">3.03647897e+01, class="num">2.87462989e+01, class="num">2.73771066e+01, class="num">2.63857585e+01, class="num">2.54625376e+01, class="num">2.43656339e+01, class="num">2.33983568e+01, class="num">2.26334633e+01 }; class=class="str">"cmt">//+------------------------------------------------------------------+ class=class="str">"cmt">//| Expert initialization function | class=class="str">"cmt">//+------------------------------------------------------------------+ class="type">int OnInit() { class=class="str">"cmt">//--- Do no display the indicators, they will clutter our view TesterHideIndicators(true); class=class="str">"cmt">//--- Setup our pointers to our WPR objects update_indicators(); class=class="str">"cmt">//--- Get trade information on the symbol TradeInformation = new TradeInfo(Symbol(),SYSTEM_TIMEFRAME_1); class=class="str">"cmt">//--- Create our ONNXFloat objects umap_onnx = new ONNXFloat(umap_proto); embedded_onnx = new ONNXFloat(embedded_proto); class=class="str">"cmt">//--- Create our Time management object eurgbp_daily = new Time(Symbol(),SYSTEM_TIMEFRAME_1); class=class="str">"cmt">//--- Check if the models are valid if(!umap_onnx.OnnxModelIsValid()) class="kw">return(INIT_FAILED); if(!embedded_onnx.OnnxModelIsValid()) class="kw">return(INIT_FAILED); class=class="str">"cmt">//--- Reset our position timer
◍ OnTick 里的持仓计时与每日触发逻辑
EA 在 OnDeinit 里把 umap_onnx、embedded_onnx、eurgbp_daily 三个指针以及 wpr_array[0..13] 共 14 个 WPR 对象指针逐一 delete，避免 MT5 退出时内存泄漏；注意循环上界是 13，说明 WPR 数组固定长度为 14。 OnTick 不每 tick 跑策略，而是用 eurgbp_daily.NewCandle() 卡每日新 K 线：新日线出现时先 Print 自增计数 i，再 update_indicators() 刷新指标。若 PositionsTotal()==0 就清零 position_timer 并调 find_setup() 找形态；若已有持仓且 position_timer<HORIZON 则每新日 +1，到了 HORIZON-1 就 Trade.PositionClose(Symbol()) 强制平掉。 这套计时把持仓寿命锁在 HORIZON 根日线内，外汇与贵金属杠杆高、跳空频发，硬平机制可能截断趋势但也限制隔夜风险，实盘前应在 MT5 策略测试器把 HORIZON 从默认值改小做对照。 find_setup() 开头再次 update_indicators() 并构造长度为 UMAP_INPUTS 的 market_state 向量，输入准备和 OnTick 的刷新有重复调用，复制代码时可合并以减少冗余。
MQL5 / C++
position_timer = class="num">0; class=class="str">"cmt">//--- Specify the models I/O shapes if(!umap_onnx.DefineOnnxInputShape(class="num">0,class="num">1,UMAP_INPUTS)) class="kw">return(INIT_FAILED); if(!embedded_onnx.DefineOnnxInputShape(class="num">0,class="num">1,EMBEDDED_INPUTS)) class="kw">return(INIT_FAILED); if(!umap_onnx.DefineOnnxOutputShape(class="num">0,class="num">1,UMAP_OUTPUTS)) class="kw">return(INIT_FAILED); if(!embedded_onnx.DefineOnnxOutputShape(class="num">0,class="num">1,EMBEDDED_OUTPUTS)) class="kw">return(INIT_FAILED); class=class="str">"cmt">//--- class="kw">return(INIT_SUCCEEDED); } class=class="str">"cmt">//+------------------------------------------------------------------+ class=class="str">"cmt">//| Expert deinitialization function | class=class="str">"cmt">//+------------------------------------------------------------------+ class="type">void OnDeinit(const class="type">int reason) { class=class="str">"cmt">//--- Delete the pointers for our custom objects class="kw">delete umap_onnx; class="kw">delete embedded_onnx; class="kw">delete eurgbp_daily; class=class="str">"cmt">//--- Delete all pointers to our WPR objects for(class="type">int i = class="num">0; i <= class="num">13; i++) { class="kw">delete wpr_array[i]; } } class=class="str">"cmt">//+------------------------------------------------------------------+ class=class="str">"cmt">//| Expert tick function | class=class="str">"cmt">//+------------------------------------------------------------------+ class="type">void OnTick() { class=class="str">"cmt">//--- Do we have a new daily candle? if(eurgbp_daily.NewCandle()) { class="kw">static class="type">int i = class="num">0; Print(i+=class="num">1); update_indicators(); if(PositionsTotal() == class="num">0) { position_timer =class="num">0; find_setup(); } else if((PositionsTotal() > class="num">0) && (position_timer < HORIZON)) position_timer += class="num">1; else if((PositionsTotal() > class="num">0) && (position_timer >= (HORIZON -class="num">1))) Trade.PositionClose(Symbol()); Comment("Position Timer: ",position_timer); } } class=class="str">"cmt">//+------------------------------------------------------------------+ class=class="str">"cmt">//| Find A Trading Setup For Us | class=class="str">"cmt">//+------------------------------------------------------------------+ class="type">void find_setup(class="type">void) { class=class="str">"cmt">//--- Update our indicators update_indicators(); class=class="str">"cmt">//--- Prepare our input vector vectorf market_state(UMAP_INPUTS);
把行情切片塞进降维向量
做 UMAP 嵌入前，得先把当前市场状态压成一个定长浮点数组。下面这段逻辑把 27 个特征写进 market_state[]，前 8 个是价格与跨期差值，后面 19 个来自 14 组 WPR 读数及其差分。 前 4 位直接取 SYSTEM_TIMEFRAME_1 的当根 K 线开高低收；第 4~7 位用当根减去 HORIZON 根之前的同价位，得到绝对位移。这类跨期差在 EURUSD 的 M15 上，HORIZON=20 时单根振幅常落在 0~30 点区间，能粗略反映中短周期动量。 WPR 部分先用 wpr_array[0..13] 取即时读数（第 8~21 位），再取 GetDifferencedReadingAt(0) 的差分（第 22~26 位）。差分项对拐点更敏感，回测里常比裸读数更早脱离超买超卖带。外汇与贵金属杠杆高，这类特征仅作状态描述，信号失效时亏损可能快速放大。 代码里所有赋值都强转 (float)，说明后续嵌入层吃单精度输入；若你自己在 MT5 里改特征维度，记得同步调整 market_state 的数组长度，否则越界会直接报数组错误。
MQL5 / C++
class=class="str">"cmt">//--- Fill in the Market Data that has to embedded into UMAP form market_state[class="num">0] = (class="type">class="kw">float) iOpen(_Symbol,SYSTEM_TIMEFRAME_1,class="num">0); market_state[class="num">1] = (class="type">class="kw">float) iHigh(_Symbol,SYSTEM_TIMEFRAME_1,class="num">0); market_state[class="num">2] = (class="type">class="kw">float) iLow(_Symbol,SYSTEM_TIMEFRAME_1,class="num">0); market_state[class="num">3] = (class="type">class="kw">float) iClose(_Symbol,SYSTEM_TIMEFRAME_1,class="num">0); market_state[class="num">4] = (class="type">class="kw">float)(iOpen(_Symbol,SYSTEM_TIMEFRAME_1,class="num">0) - iOpen(Symbol(),SYSTEM_TIMEFRAME_1,HORIZON)); market_state[class="num">5] = (class="type">class="kw">float)(iHigh(_Symbol,SYSTEM_TIMEFRAME_1,class="num">0) - iHigh(Symbol(),SYSTEM_TIMEFRAME_1,HORIZON)); market_state[class="num">6] = (class="type">class="kw">float)(iLow(_Symbol,SYSTEM_TIMEFRAME_1,class="num">0) - iLow(Symbol(),SYSTEM_TIMEFRAME_1,HORIZON)); market_state[class="num">7] = (class="type">class="kw">float)(iClose(_Symbol,SYSTEM_TIMEFRAME_1,class="num">0) - iClose(Symbol(),SYSTEM_TIMEFRAME_1,HORIZON)); market_state[class="num">8] = (class="type">class="kw">float) wpr_array[class="num">0].GetReadingAt(class="num">0); market_state[class="num">9] = (class="type">class="kw">float) wpr_array[class="num">1].GetReadingAt(class="num">0); market_state[class="num">10] = (class="type">class="kw">float) wpr_array[class="num">2].GetReadingAt(class="num">0); market_state[class="num">11] = (class="type">class="kw">float) wpr_array[class="num">3].GetReadingAt(class="num">0); market_state[class="num">12] = (class="type">class="kw">float) wpr_array[class="num">4].GetReadingAt(class="num">0); market_state[class="num">13] = (class="type">class="kw">float) wpr_array[class="num">5].GetReadingAt(class="num">0); market_state[class="num">14] = (class="type">class="kw">float) wpr_array[class="num">6].GetReadingAt(class="num">0); market_state[class="num">15] = (class="type">class="kw">float) wpr_array[class="num">7].GetReadingAt(class="num">0); market_state[class="num">16] = (class="type">class="kw">float) wpr_array[class="num">8].GetReadingAt(class="num">0); market_state[class="num">17] = (class="type">class="kw">float) wpr_array[class="num">9].GetReadingAt(class="num">0); market_state[class="num">18] = (class="type">class="kw">float) wpr_array[class="num">10].GetReadingAt(class="num">0); market_state[class="num">19] = (class="type">class="kw">float) wpr_array[class="num">11].GetReadingAt(class="num">0); market_state[class="num">20] = (class="type">class="kw">float) wpr_array[class="num">12].GetReadingAt(class="num">0); market_state[class="num">21] = (class="type">class="kw">float) wpr_array[class="num">13].GetReadingAt(class="num">0); market_state[class="num">22] = (class="type">class="kw">float) wpr_array[class="num">0].GetDifferencedReadingAt(class="num">0); market_state[class="num">23] = (class="type">class="kw">float) wpr_array[class="num">1].GetDifferencedReadingAt(class="num">0); market_state[class="num">24] = (class="type">class="kw">float) wpr_array[class="num">2].GetDifferencedReadingAt(class="num">0); market_state[class="num">25] = (class="type">class="kw">float) wpr_array[class="num">3].GetDifferencedReadingAt(class="num">0); market_state[class="num">26] = (class="type">class="kw">float) wpr_array[class="num">4].GetDifferencedReadingAt(class="num">0);
「把降维嵌入接进反转信号的实盘触发」
这段逻辑干的事很直接：先把 9 个威廉指标差分读数（索引 27 到 35）塞进 market_state 数组，再统一用 Z1 均值、Z2 标准差做标准化，喂给 UMAP 的 ONNX 模型拿到低维嵌入，接着用第二个 ONNX 模型把嵌入转成对 EURGBP 的未来回报预期。 标准化那步是个 for 循环，遍历 UMAP_INPUTS 长度，每行都是 (market_state[i] - Z1[i]) / Z2[i]。Z1、Z2 若是训练集统计出的固定值，换品种或换周期就可能偏移，MT5 里 Print 一下 umap_predictions 和 expected_eurgbp_return 能直接看到数值是否合理。 信号触发只看 SYSTEM_TIMEFRAME_1 上最近 HORIZON 根 K 线的开收均值关系：o.Mean() < c.Mean() 判为 bullish_reversal，反之 bearish_reversal。若空头反转形态出现且模型预期回报为正，就按最小手数两倍 Buy，否则最小手数 Buy；多头反转且预期为负则对称地 Sell。 外汇与贵金属杠杆高，这种由 ML 嵌入驱动的反转单属于概率倾向而非确定性，实盘前务必在策略测试器用真实点差回测，重点观察 expected_return 符号与反转方向冲突时的胜率衰减。
MQL5 / C++
market_state[class="num">27] = (class="type">class="kw">float) wpr_array[class="num">5].GetDifferencedReadingAt(class="num">0); market_state[class="num">27] = (class="type">class="kw">float) wpr_array[class="num">6].GetDifferencedReadingAt(class="num">0); market_state[class="num">29] = (class="type">class="kw">float) wpr_array[class="num">7].GetDifferencedReadingAt(class="num">0); market_state[class="num">30] = (class="type">class="kw">float) wpr_array[class="num">8].GetDifferencedReadingAt(class="num">0); market_state[class="num">31] = (class="type">class="kw">float) wpr_array[class="num">9].GetDifferencedReadingAt(class="num">0); market_state[class="num">32] = (class="type">class="kw">float) wpr_array[class="num">10].GetDifferencedReadingAt(class="num">0); market_state[class="num">33] = (class="type">class="kw">float) wpr_array[class="num">11].GetDifferencedReadingAt(class="num">0); market_state[class="num">34] = (class="type">class="kw">float) wpr_array[class="num">12].GetDifferencedReadingAt(class="num">0); market_state[class="num">35] = (class="type">class="kw">float) wpr_array[class="num">13].GetDifferencedReadingAt(class="num">0); class=class="str">"cmt">//--- Standardize and scale each input for(class="type">int i =class="num">0; i < UMAP_INPUTS;i++) { market_state[i] = (class="type">class="kw">float)((market_state[i] - Z1[i]) / Z2[i]); }; const vectorf onnx_inputs = market_state; const vectorf umap_predictions = umap_onnx.Predict(onnx_inputs); Print("UMAP Model Returned Embeddings: ",umap_predictions); const vectorf expected_eurgbp_return = embedded_onnx.Predict(umap_predictions); Print("Embeddings Model Expects EURGBP Returns: ",expected_eurgbp_return); expected_return = expected_eurgbp_return[class="num">0]; vector o,c; o.CopyRates(Symbol(),SYSTEM_TIMEFRAME_1,COPY_RATES_OPEN,class="num">0,HORIZON); c.CopyRates(Symbol(),SYSTEM_TIMEFRAME_1,COPY_RATES_CLOSE,class="num">0,HORIZON); class="type">bool bullish_reversal = o.Mean() < c.Mean(); class="type">bool bearish_reversal = o.Mean() > c.Mean(); if(bearish_reversal) { if(expected_return > class="num">0) { Trade.Buy((TradeInformation.MinVolume()*class="num">2),Symbol(),TradeInformation.GetAsk(),class="num">0,class="num">0,""); class="kw">return; } Trade.Buy(TradeInformation.MinVolume(),Symbol(),TradeInformation.GetAsk(),class="num">0,class="num">0,""); class="kw">return; } else if(bullish_reversal) { if(expected_return < class="num">0) { Trade.Sell((TradeInformation.MinVolume()*class="num">2),Symbol(),TradeInformation.GetBid(),class="num">0,class="num">0,""); } Trade.Sell(TradeInformation.MinVolume(),Symbol(),TradeInformation.GetBid(),class="num">0,class="num">0,""); class="kw">return; } class=class="str">"cmt">//+------------------------------------------------------------------+
◍ 用循环批量装载多周期 WPR 读数
这段函数把 14 组不同周期的威廉指标（WPR）一次性塞进数组，省去手写十几个对象的麻烦。循环变量 i 从 0 跑到 13，每组周期按 (i+1)*5 计算，也就是 5、10、15……一直到 70 根 K 线。 wpr_array[i] = new WPR(Symbol(), SYSTEM_TIMEFRAME_1, ((i+1) * 5)); 这一行在当前品种、系统时间框架下新建 WPR 实例。紧接着两行 SetIndicatorValues(60, true) 与 SetDifferencedIndicatorValues(60, HORIZON, true) 把主值和差分值的缓冲长度都设为 60，差分跨度用宏 HORIZON 控制。 函数末尾用 #undef 把 EMBEDDED_INPUTS、UMAP_INPUTS、HORIZON 等临时宏全部撤销，避免污染后续编译单元。最后 TesterHideIndicators(true) 让回测时这些指标不画在图上，界面清爽但数据仍在跑。 开 MT5 把这段直接贴进 EA 的 update_indicators 函数，改 SYSTEM_TIMEFRAME_1 就能验证不同周期组合对信号密度的影响；外汇与贵金属杠杆高，回测结论仅代表历史概率，实盘须控仓。
MQL5 / C++
class=class="str">"cmt">//| Update our indicator readings | class=class="str">"cmt">//+------------------------------------------------------------------+ class="type">void update_indicators(class="type">void) { class=class="str">"cmt">//--- Store pointers to our WPR objects for(class="type">int i = class="num">0; i <= class="num">13; i++) { class=class="str">"cmt">//--- Create an WPR object wpr_array[i] = new WPR(Symbol(),SYSTEM_TIMEFRAME_1,((i+class="num">1) * class="num">5)); class=class="str">"cmt">//--- Set the WPR buffers wpr_array[i].SetIndicatorValues(class="num">60,true); wpr_array[i].SetDifferencedIndicatorValues(class="num">60,HORIZON,true); } } class=class="str">"cmt">//+------------------------------------------------------------------+ class=class="str">"cmt">//| Undefine system constants we no longer need | class=class="str">"cmt">//+------------------------------------------------------------------+ class="macro">#undef EMBEDDED_INPUTS class="macro">#undef EMBEDDED_OUTPUTS class="macro">#undef UMAP_INPUTS class="macro">#undef UMAP_OUTPUTS class="macro">#undef HORIZON class="macro">#undef SYSTEM_TIMEFRAME_1 class=class="str">"cmt">//+------------------------------------------------------------------+ class=class="str">"cmt">//--- Do no display the indicators, they will clutter our view TesterHideIndicators(true);
替换指标就能复用这套降维思路
把 36 列 WPR 特征压到 2 列 UMAP 嵌入后，回测里模型表现反而好于喂全量原始数据；这说明降维不是信息丢失，而是把非线性结构留了下来。你可以直接把 WPR 换成自己常用的指标组合，再跑一遍 embedding+ONNX 预测的两段式流程，大概率能为原有策略找到更紧凑的表示。 UMAP 和 PCA 的本质差异在几何假设：PCA 默认世界是平的、只用直线距离切数据，UMAP 把样本当黎曼流形上的点在弯曲空间里顺势贴合。遇到汇率或贵金属这种高噪声、非线性占优的盘面，后者灵活性更高，但别误读成“更强大”——它只是不强制线性前提。 随文给的 EURGBP_WPR_EMBEDDED.onnx 只有 2.33 KB、EURGBP_WPR_UMAP.onnx 26.5 KB，加载进 MT5 的 EA 里就能离线推理；Fetch_Data_Algorithmic_Input_Selection.mq5（5.42 KB）负责拉数据，EURGBP_Multiple_Periods_Analysis.mq5（12.41 KB）同时挂了 14 个 WPR 周期做验证。外汇和贵金属杠杆高、滑点突变频繁，任何模型输出都只是概率倾向，实盘前务必用历史 Tick 重跑一遍。 OOP 在 MQL5 里不算新潮，但把控制流和异常处理收进一个类文件，确实省掉大量样板代码；改参数、换品种时编译结果可预测。打开 MT5 把上面几个文件拖进对应目录，先跑通 EURGBP 再换你自己的标的，比空谈架构来得实在。
把多周期诊断交给小布盯盘
这些跨周期一致性与降维后的聚类状态，小布盯盘的AIGC已内置，打开对应品种页即可看到实时面板，你只管判断是否入场。
常见问题
WPR固定在0到-100区间且对价格位置敏感，不同周期读数可直接横向比较；RSI虽类似但平滑方式不同，本篇改用WPR是为统一缓冲区结构。
UMAP倾向保留局部流形结构，市场突变点通常在低维空间仍成簇；但属于概率性保留，极端跳空仍需结合原始周期复核。
可以，小布盯盘支持上传ONNX推理文件并绑定品种页，多周期特征提取由内置类自动完成，无需重写MQL5。
不需要，子类仅继承并编译即可；新增通用方法只动父类，这是本篇OOP设计的核心收益。
杠杆与隔夜风险放大了模型误判成本，建议先在模拟盘验证低维表示稳定性，实盘仓位倾向轻仓分步。