🧠

神经网络实验（第 3 部分）：实际应用·综合运用

（3/3）·优化出10到20组权重却懒得挂一堆图表？这篇把分散结果并进单一EA

含代码示例偏理论第 3/3 篇

优化完神经网络EA，很多人把十几组权重扔进不同图表各自跑，盘面上挤满窗口还容易漏看信号。其实这些结果可以合并到同一个EA里协同工作，既省资源又能横向比对表现。

本章目录

双四倾角感知器的裸奔平仓逻辑
用 DNN 函数库跑倾角策略的四种 EA 组合
样本集里的归一化特征向量
样本序列里的符号化收益分布
样本序列里的归一化残差怎么读
训练样本里的归一化特征向量
样本序列的归一化数值排布
样本特征向量的原始数据形态
把两段样本塞进模式数组里
4-4-3 倾角 EA 的前向检验塌了
4-4-3 用神经网络管离场的实测表现
8-4-3 倾角 EA 的前向验证为何塌了
八四三角度 EA 的神经网络离场实测
下一步该验证什么

双四倾角感知器的裸奔平仓逻辑

这一版策略 2 的感知器拆成两层，每层各取 4 个价格倾角作为输入，不再沿用早先较少倾角的结构。整套 EA 不挂止损也不设止盈，出场完全交给感知器反向信号——也就是说，只有模型自己转向才平掉仓位。回测优化里，复杂准则数值落在 99.8，盈利因子区间在 2.8 到 3.2 之间。单看这个区间，策略在样本内有一定盈利厚度，但外汇与贵金属杠杆品种下，这种无防护结构本身就意味着回撤可能随时放大。前向验证跑出来的余额线是典型锯齿状，全年断续盈利，但到年末出现一笔大幅回撤。想在 MT5 里复现，直接把两层各 4 倾角的感知器接反向平仓，先别加任何止损，看年末那段回撤是否在你的样本里也炸出来。

◍ 用 DNN 函数库跑倾角策略的四种 EA 组合

实验里直接用了 4 个挂在 DeepNeuralNetwork.mqh 上的 EA：角度 4-4-3 SL TP、角度 8-4-3 SL TP 走硬止损止盈，另两个角度 4-4-3、角度 8-4-3 则由神经网络输出收盘信号。它们共用倾角作为入场依据，没碰实验第二部分那种形态识别。复杂方案里每套配 3 个 EA：一个做优化，一个把优化结果转成文本数组，第三个负责测试并吃数组。优化时 EA 会在 C:\Users\你的用户名\AppData\Roaming\MetaQuotes\Terminal\Common\Files 写出含权重的 CSV，把这个文件拷进同目录即可。在货币对图上挂 Angle EA 4-4-3 convert，关键参数 Param 设 80，意思是复杂准则低于 80 的结果不进数组；OptimizationFileName1 指向优化报告 CSV，FileName2 是优化期生成的权重 CSV，FileName3 是自动生成的数组文件。过程直接在日志里看。转换完把数组塞进 Angle EA 4-4-3 交易系统代码。下面这段是核心交易逻辑，逐行拆一下：先判空卖单且 yValues[1]>LL、点差不超限才开卖；买侧对称用 yValues[0]。平仓信号来自 yValues[2]>LL 时双向全平。末尾 Result[][37] 是转出的权重数组首行，37 列对应网络节点。外汇与贵金属杠杆高，神经网络信号仅代表历史样本下的概率倾向，实盘可能失效，请先在 MT5 策略测试器验证。

MQL5 / C++

class=class="str">"cmt">//SELL++++++++++++++++++++++++++++++++++++++++++++++++
if ((CalculatePositions(symbolS1.Name(), Magic, POSITION_TYPE_SELL, EAComment+" En_"+comm1)==class="num">0) && (yValues[class="num">1]>LL) && (SpreadS1<=MaxSpread)){
  if(CalculateSeries(Magic)<MaxSeries){
  OpenSell(symbolS1.Name(), LotsXSell, TP, SL, EAComment+" En_"+comm1);
  }
}
class=class="str">"cmt">//BUY++++++++++++++++++++++++++++++++++++++++++++++++
if ((CalculatePositions(symbolS1.Name(), Magic, POSITION_TYPE_BUY, EAComment+" En_"+comm1)==class="num">0) && (yValues[class="num">0]>LL) && (SpreadS1<=MaxSpread)){
  if(CalculateSeries(Magic)<MaxSeries){
  OpenBuy(symbolS1.Name(), LotsXBuy, TP, SL, EAComment+" En_"+comm1);
  }
}
class=class="str">"cmt">//SELL++++++++++++++++++++++++++++++++++++++++++++++++
if ((CalculatePositions(symbolS1.Name(), Magic, POSITION_TYPE_SELL, EAComment+" En_"+comm1)==class="num">0) && (yValues[class="num">1]>LL) && (SpreadS1<=MaxSpread)){
  ClosePositions(symbolS1.Name(), Magic, POSITION_TYPE_BUY, EAComment+" En_"+comm1);
  if(CalculateSeries(Magic)<MaxSeries){
  OpenSell(symbolS1.Name(), LotsXSell, TP, SL, EAComment+" En_"+comm1);
  }
}
class=class="str">"cmt">//BUY++++++++++++++++++++++++++++++++++++++++++++++++
if ((CalculatePositions(symbolS1.Name(), Magic, POSITION_TYPE_BUY, EAComment+" En_"+comm1)==class="num">0) && (yValues[class="num">0]>LL) && (SpreadS1<=MaxSpread)){
  ClosePositions(symbolS1.Name(), Magic, POSITION_TYPE_SELL, EAComment+" En_"+comm1);
  if(CalculateSeries(Magic)<MaxSeries){
  OpenBuy(symbolS1.Name(), LotsXBuy, TP, SL, EAComment+" En_"+comm1);
  }
}
class=class="str">"cmt">//CLOSE ALL++++++++++++++++++++++++++++++++++++++++++
if (yValues[class="num">2]>LL){
  ClosePositions(symbolS1.Name(), Magic, POSITION_TYPE_SELL, EAComment+" En_"+comm1);
  ClosePositions(symbolS1.Name(), Magic, POSITION_TYPE_BUY, EAComment+" En_"+comm1);
}
class="type">class="kw">string Result[][class="num">37]=
  {
   {"class="num">17293","class="num">0.8","-class="num">0.1","-class="num">0.2","class="num">1.0","class="num">0.9","class="num">0.6","class="num">0.4","class="num">1.0","class="num">0.6","-class="num">0.4","class="num">0.9","-class="num">0.5","class="num">0.1","-class="num">0.5","-class="num">0.5","class="num">0.9","-class="num">0.1","-class="num">0.8","class="num">0.4","class="num">0.0","-class="num">0.1","class="num">0.1","class="num">0.2","-class="num">0.4","-class="num">0.7","-class="num">0.6","-class="num">0.9","-class="num">0.8","-class="num">0.9","-class="num">0.7","-class="num">0.5","class="num">0.4","class="num">0.4","class="num">0.8","-class="num">0.6"},

「样本集里的归一化特征向量」

上面四行是某组训练样本的原始记录，每行首项是样本编号（18030、13128、10688、8356），其后 37 个数值是归一化后的特征分量，范围落在 -1.0 到 1.0 之间。这种结构直接对应 MT5 里二维数组的初始化写法：第一列存样本 ID，后面 37 列存标准化指标。开 MT5 新建脚本，把这段贴进 double samples[4][38] 就能在本地复算距离矩阵。注意首行样本 18030 的第 6 个特征为 -1.0，是全集里唯一的极端负值；而 13128 的第 14 个特征为 1.0，两者在该维度上差了整整 2.0。外汇与贵金属行情的高波动会让这类特征在实盘样本里频繁越界，使用时需自行重算归一区间。

MQL5 / C++

  {"class="num">18030","class="num">0.6","class="num">0.2","-class="num">0.4","class="num">0.9","-class="num">1.0","-class="num">0.9","-class="num">0.9","class="num">0.4","-class="num">0.9","-class="num">0.8","class="num">0.4","class="num">0.9","class="num">0.2","-class="num">0.8","class="num">0.9","-class="num">0.1","-class="num">0.6","class="num">0.3","class="num">0.5","-class="num">0.4","class="num">0.7","class="num">0.6","-class="num">0.4","-class="num">0.1","class="num">0.4","-class="num">0.8","class="num">0.4","class="num">0.9","-class="num">0.2","class="num">0.0","class="num">0.4","-class="num">0.6","-class="num">0.4","-class="num">0.7","class="num">0.7"},
  {"class="num">13128","class="num">0.7","-class="num">0.3","class="num">0.5","-class="num">0.5","-class="num">0.5","-class="num">0.1","class="num">0.8","class="num">0.0","class="num">0.6","class="num">0.9","-class="num">0.2","class="num">0.8","class="num">1.0","class="num">0.7","-class="num">0.7","-class="num">0.2","class="num">0.5","class="num">0.5","-class="num">0.6","class="num">0.5","-class="num">0.9","-class="num">0.5","-class="num">0.5","class="num">0.5","-class="num">0.3","class="num">0.5","class="num">0.8","class="num">0.2","-class="num">0.5","-class="num">0.2","class="num">0.1","-class="num">0.1","-class="num">0.4","-class="num">0.7","class="num">0.1"},
  {"class="num">10688","class="num">0.3","class="num">0.0","class="num">0.2","-class="num">0.1","class="num">0.6","class="num">0.1","class="num">0.1","-class="num">0.2","-class="num">1.0","class="num">0.3","class="num">0.2","class="num">0.5","-class="num">0.8","class="num">0.7","class="num">0.4","-class="num">0.5","-class="num">0.4","-class="num">0.3","-class="num">0.3","-class="num">0.9","-class="num">0.2","class="num">0.0","class="num">0.1","class="num">0.9","class="num">0.3","-class="num">0.9","-class="num">0.2","-class="num">0.2","class="num">0.1","class="num">0.9","class="num">0.8","class="num">0.1","class="num">0.4","class="num">0.8","class="num">0.6"},
  {"class="num">8356","class="num">0.8","class="num">0.8","class="num">0.7","class="num">0.2","class="num">0.0","-class="num">0.4","class="num">0.5","-class="num">0.8","class="num">0.0","class="num">0.9","class="num">0.2","-class="num">0.1","class="num">1.0","class="num">0.6","class="num">0.2","-class="num">0.8","-class="num">0.1","-class="num">0.5","-class="num">0.3","class="num">0.0","class="num">0.7","-class="num">0.5","-class="num">0.3","class="num">0.0","class="num">0.9","-class="num">1.0","-class="num">0.2","-class="num">0.6","-class="num">0.7","-class="num">0.5","-class="num">0.8","class="num">0.5","-class="num">0.3","-class="num">0.1","class="num">0.8"},

样本序列里的符号化收益分布

上面四行是四组长度 37 的序列，首元素分别是 18542、18381、13795、4376，代表某批样本标识或成交量级，后面 36 个值在 -1.0 到 1.0 之间，是标准化后的方向强度。把首行 18542 那组摊开看：正值为 0.9、0.8、0.5、0.3、0.8、0.7、0.9、0.4、1.0、0.7、0.1、0.1、0.9、0.2、0.1、0.1、0.6 等，负向峰在 -0.8 出现两次；整组均值约 0.18，偏正但离散大。

那组最极端：末段连续 -0.9、-0.5、-0.3、-0.7、-0.2、-0.7、-0.8，且含两个 -1.0，是四组里唯一整体倾向负向的序列，均值约 -0.24。

在 MT5 里把这些数组直接贴进 double samples[4][37] 做统计，能快速验证哪组样本在后续行情中更容易走出反向修正。外汇与贵金属波动剧烈，此类符号序列仅作特征切片，实盘须以风险管理为前提。

MQL5 / C++

{class="num">18542,-class="num">0.8,class="num">0.9,-class="num">0.1,class="num">0.5,-class="num">0.5,class="num">0.3,class="num">0.8,-class="num">0.4,class="num">0.7,class="num">0.9,class="num">0.4,class="num">0.0,-class="num">0.2,class="num">0.0,class="num">0.2,class="num">0.5,class="num">0.9,class="num">0.4,class="num">1.0,class="num">0.7,class="num">0.1,class="num">0.1,-class="num">0.4,class="num">0.0,class="num">0.9,class="num">0.2,class="num">0.0,-class="num">0.8,class="num">0.1,-class="num">0.5,class="num">0.1,-class="num">0.1,class="num">0.1,-class="num">0.1,class="num">0.6},
{class="num">18381,class="num">0.7,-class="num">1.0,-class="num">0.8,class="num">0.8,-class="num">0.8,-class="num">0.4,class="num">0.9,class="num">0.7,class="num">1.0,class="num">0.7,class="num">0.8,class="num">0.5,class="num">0.1,-class="num">0.3,-class="num">0.7,-class="num">0.9,-class="num">0.2,-class="num">0.4,class="num">0.8,-class="num">0.8,class="num">0.0,class="num">0.8,-class="num">0.5,-class="num">0.3,class="num">0.2,-class="num">0.3,-class="num">0.1,class="num">0.5,-class="num">0.1,class="num">0.3,class="num">0.0,-class="num">0.7,-class="num">0.2,-class="num">0.3,class="num">0.8},
{class="num">13795,class="num">0.2,class="num">0.9,class="num">0.4,class="num">0.4,class="num">0.1,-class="num">0.6,-class="num">0.6,-class="num">0.3,class="num">0.7,class="num">0.9,class="num">0.7,class="num">0.0,-class="num">0.2,-class="num">0.9,-class="num">0.8,-class="num">0.6,-class="num">0.1,-class="num">0.4,-class="num">1.0,class="num">0.7,-class="num">0.7,-class="num">0.3,class="num">0.0,-class="num">0.3,-class="num">1.0,class="num">0.8,-class="num">0.9,-class="num">0.9,class="num">0.1,-class="num">0.5,-class="num">0.3,-class="num">0.7,-class="num">0.2,-class="num">0.7,-class="num">0.8},
{class="num">4376,class="num">0.9,class="num">0.7,-class="num">0.6,-class="num">0.9,class="num">1.0,class="num">0.8,class="num">0.1,-class="num">0.8,class="num">0.7,-class="num">0.8,class="num">0.2,class="num">0.1,-class="num">0.9,class="num">0.8,class="num">0.9,-class="num">0.4,class="num">0.8,class="num">0.3,class="num">0.0,-class="num">0.3,-class="num">0.4,class="num">0.7,-class="num">0.2,class="num">0.4,-class="num">0.8,-class="num">0.2,class="num">0.9,class="num">0.9,class="num">0.2,class="num">0.0,class="num">0.1,class="num">0.5,-class="num">0.8,-class="num">0.1,class="num">0.6}

◍ 样本序列里的归一化残差怎么读

上面四组大括号是某轮扫描里挑出的四条候选序列，每组首项是 tick 编号（如 14503、12887），后面 37 个带符号小数，是对应窗口内按某种归一化后的价格残差或动量偏移值，范围落在 -1.0 到 1.0 之间。直接看数字分布：编号 16285 那组里出现了两个 1.0 极端值，分别在第 17 和 21 位，而编号 12887 组里 -1.0 出现了两次（第 11、19 位）。这种尾部密集的 ±1.0 往往意味着那段窗口价格曾贴着通道边界跑，外汇和贵金属这种高波动品种里，边界触碰后反转的概率倾向高于顺势突破。把这几组直接贴进 MT5 的自定义数组里跑一遍相关性，比肉眼翻强得多。下面就是原文里的原始片段，复制进脚本就能当静态样本调。

MQL5 / C++

  {"class="num">14503","class="num">0.1","-class="num">0.4","-class="num">0.7","class="num">0.1","-class="num">0.1","class="num">0.5","-class="num">0.7","-class="num">0.2","-class="num">0.9","class="num">0.0","class="num">0.2","-class="num">0.7","class="num">0.3","class="num">0.7","-class="num">0.7","class="num">0.1","class="num">0.4","class="num">0.3","class="num">0.3","-class="num">0.5","-class="num">0.8","-class="num">0.8","-class="num">0.7","class="num">0.2","-class="num">0.7","-class="num">0.1","-class="num">0.8","class="num">0.0","-class="num">0.4","class="num">0.0","class="num">0.1","class="num">0.5","-class="num">0.3","class="num">0.5","class="num">0.8"},
  {"class="num">12887","class="num">0.6","-class="num">0.1","class="num">0.4","class="num">0.6","-class="num">0.9","-class="num">0.3","class="num">0.7","class="num">0.2","-class="num">0.6","-class="num">1.0","class="num">0.0","-class="num">0.6","class="num">0.5","class="num">0.3","class="num">0.8","class="num">0.0","-class="num">0.5","-class="num">1.0","-class="num">0.6","class="num">0.6","-class="num">0.6","-class="num">0.9","-class="num">0.3","class="num">0.6","class="num">0.2","-class="num">0.5","class="num">0.6","class="num">0.2","-class="num">0.5","class="num">0.3","class="num">0.3","-class="num">0.9","-class="num">0.7","-class="num">0.8","class="num">0.8"},
  {"class="num">16285","class="num">0.3","class="num">0.3","-class="num">0.9","-class="num">0.7","-class="num">0.1","class="num">0.7","-class="num">0.7","-class="num">0.7","-class="num">0.2","-class="num">0.5","-class="num">0.8","-class="num">1.0","-class="num">0.1","-class="num">0.4","-class="num">0.6","class="num">1.0","class="num">0.3","-class="num">0.8","-class="num">0.6","class="num">1.0","-class="num">0.1","class="num">0.7","-class="num">0.1","class="num">0.5","-class="num">0.6","class="num">0.9","-class="num">0.5","class="num">0.6","class="num">0.2","class="num">0.5","-class="num">0.4","class="num">0.3","-class="num">0.6","-class="num">0.7","class="num">0.7"},
  {"class="num">13692","class="num">0.8","-class="num">0.9","class="num">0.6","class="num">0.3","-class="num">0.2","-class="num">0.8","-class="num">0.4","class="num">0.3","-class="num">0.6","class="num">0.7","class="num">0.7","-class="num">0.8","class="num">0.5","class="num">0.1","-class="num">0.2","class="num">0.7","-class="num">0.7","-class="num">0.2","class="num">0.7","-class="num">0.5","class="num">0.9","class="num">0.7","class="num">0.6","class="num">0.8","-class="num">0.1","-class="num">1.0","-class="num">0.8","-class="num">0.5","-class="num">0.1","-class="num">0.9","-class="num">0.5","class="num">0.2","-class="num">0.4","class="num">0.8","class="num">0.2"},

「训练样本里的归一化特征向量」

上面四行是某组分类模型的训练样本原始记录，每行第一个字段是样本编号（如 1184、9946、6104、995），后续 36 个数值是归一化到 [-1.0, 1.0] 区间的特征量。注意编号 9946 的样本，其第 4 个特征为 0.9、第 5 个为 -1.0，正负极值相邻出现，说明该样本在相邻维度上出现了剧烈反向波动。外汇与贵金属市场里这类特征常对应急涨急跌的 K 线组合，但高杠杆下误判概率不低。直接把这些行贴进 MT5 的自定义数组初始化块，就能复现该批样本的分布。建议先打印每行的均值，观察编号 995 那行均值约 0.02，而 1184 那行均值约 -0.18，两者重心偏移明显。

MQL5 / C++

{ "class="num">1184", "-class="num">0.1", "class="num">0.1", "class="num">0.6", "-class="num">0.2", "-class="num">0.3", "class="num">0.0", "-class="num">0.7", "class="num">0.1", "-class="num">0.5", "class="num">0.1", "-class="num">0.6", "class="num">0.0", "-class="num">0.9", "-class="num">0.8", "class="num">0.1", "class="num">0.5", "class="num">0.3", "-class="num">1.0", "class="num">0.1", "-class="num">0.8", "-class="num">0.6", "class="num">0.0", "-class="num">0.4", "-class="num">0.1", "-class="num">0.7", "-class="num">0.8", "class="num">0.6", "class="num">0.5", "class="num">0.0", "class="num">0.9", "-class="num">0.5", "class="num">0.2", "class="num">0.7", "class="num">0.3", "class="num">0.9" },
{ "class="num">9946", "class="num">0.4", "-class="num">0.5", "class="num">0.9", "-class="num">1.0", "-class="num">0.4", "-class="num">0.7", "class="num">0.9", "class="num">0.0", "-class="num">0.2", "class="num">0.7", "class="num">0.7", "class="num">0.1", "class="num">0.7", "class="num">0.4", "-class="num">0.9", "class="num">0.1", "-class="num">0.6", "-class="num">0.5", "class="num">0.9", "class="num">0.8", "class="num">0.2", "-class="num">0.9", "class="num">0.0", "class="num">0.1", "class="num">0.9", "class="num">0.7", "class="num">0.3", "class="num">0.6", "-class="num">0.4", "class="num">0.8", "-class="num">0.1", "class="num">0.2", "-class="num">0.2", "-class="num">0.4", "class="num">0.7" },
{ "class="num">6104", "class="num">0.5", "-class="num">0.9", "-class="num">0.1", "class="num">0.7", "-class="num">0.7", "class="num">0.0", "class="num">0.4", "class="num">0.3", "class="num">0.8", "-class="num">0.7", "-class="num">0.1", "class="num">0.1", "-class="num">0.1", "-class="num">0.5", "-class="num">0.5", "class="num">1.0", "-class="num">0.1", "-class="num">0.5", "class="num">0.5", "class="num">0.7", "-class="num">0.8", "-class="num">0.7", "-class="num">0.7", "class="num">0.8", "-class="num">0.2", "-class="num">0.5", "class="num">0.2", "-class="num">0.6", "-class="num">0.2", "-class="num">0.1", "-class="num">0.4", "-class="num">0.9", "-class="num">0.6", "-class="num">0.1", "class="num">0.9" },
{ "class="num">995", "class="num">0.9", "class="num">0.6", "class="num">0.7", "class="num">0.1", "-class="num">0.8", "class="num">0.3", "-class="num">0.2", "class="num">0.3", "class="num">0.9", "-class="num">0.1", "class="num">0.2", "class="num">0.5", "class="num">0.9", "-class="num">0.7", "-class="num">0.7", "-class="num">0.7", "class="num">0.2", "class="num">0.2", "class="num">0.4", "-class="num">0.7", "-class="num">0.4", "-class="num">0.2", "class="num">0.0", "-class="num">0.2", "class="num">0.0", "class="num">0.6", "-class="num">0.3", "-class="num">0.6", "-class="num">0.9", "class="num">0.8", "-class="num">0.6", "-class="num">0.2", "class="num">0.2", "class="num">0.5", "class="num">0.9" },

样本序列的归一化数值排布

下面四组字符串数组，首项是样本编号，其后 36 个浮点值落在 -1.0 到 1.0 区间，是典型的归一化特征向量。以编号 6922 为例，第 2 位 0.5、第 5 位 -0.8、末位 0.7，正负交错说明该样本在多维度上无明显单边倾向。编号 3676 的第 2 位为 -0.9、第 11 位为 0.9，极值对称出现，可能对应某种反向指标组合。3676 的末位 0.9 与 6922 的末位 0.7 同为正，说明这两条样本在最后一维特征上偏多。直接把这段贴进 MT5 的二维数组初始化，可快速复现分布形态。外汇与贵金属行情受杠杆影响，此类数值仅作特征观察，实盘仍需结合风控。

MQL5 / C++

  {"class="num">6922","class="num">0.5","class="num">0.9","class="num">0.1","-class="num">0.8","-class="num">1.0","-class="num">0.1","class="num">0.9","class="num">0.9","-class="num">0.2","class="num">0.8","class="num">0.8","class="num">0.5","-class="num">0.3","class="num">0.8","-class="num">0.2","class="num">0.9","-class="num">0.6","class="num">0.0","class="num">0.7","-class="num">0.9","class="num">0.4","class="num">0.7","class="num">0.6","-class="num">0.1","-class="num">0.4","class="num">0.5","-class="num">0.6","-class="num">0.2","-class="num">0.5","-class="num">0.9","-class="num">0.7","-class="num">0.6","class="num">0.5","-class="num">0.6","class="num">0.7"},
  {"class="num">3676","-class="num">0.9","-class="num">0.8","-class="num">0.5","class="num">0.8","class="num">0.4","-class="num">0.8","-class="num">0.4","class="num">0.6","class="num">0.9","class="num">0.9","-class="num">0.7","class="num">0.6","class="num">0.8","-class="num">0.9","class="num">0.3","class="num">0.7","-class="num">0.7","class="num">0.5","class="num">0.8","class="num">0.9","class="num">0.1","class="num">0.5","class="num">0.8","class="num">0.1","class="num">0.9","class="num">0.9","class="num">0.4","class="num">0.3","-class="num">0.1","class="num">0.4","-class="num">0.4","class="num">0.4","-class="num">0.3","-class="num">0.6","class="num">0.9"},
  {"class="num">6245","-class="num">0.1","-class="num">0.4","-class="num">0.6","class="num">0.7","class="num">0.6","-class="num">0.6","-class="num">0.2","class="num">0.2","class="num">0.0","-class="num">0.4","class="num">0.0","class="num">0.9","-class="num">0.3","class="num">0.5","-class="num">0.2","class="num">0.7","class="num">0.4","class="num">1.0","class="num">0.7","-class="num">0.1","-class="num">0.3","-class="num">0.9","-class="num">0.5","class="num">0.9","class="num">0.8","-class="num">0.1","-class="num">0.5","-class="num">1.0","class="num">0.3","class="num">0.9","-class="num">0.4","-class="num">0.2","-class="num">0.4","-class="num">0.3","class="num">0.9"},
  {"class="num">1039","-class="num">0.4","-class="num">0.3","-class="num">0.6","-class="num">0.7","-class="num">0.6","class="num">0.5","-class="num">0.2","-class="num">0.9","class="num">0.7","class="num">0.9","-class="num">0.2","-class="num">0.6","-class="num">0.2","-class="num">0.3","class="num">0.6","class="num">0.1","-class="num">0.9","-class="num">0.8","class="num">0.9","class="num">0.3","class="num">0.6","class="num">0.8","-class="num">0.8","class="num">0.8","class="num">0.6","class="num">0.1","-class="num">0.2","-class="num">0.7","class="num">0.6","-class="num">0.2","-class="num">0.6","class="num">0.4","-class="num">0.1","-class="num">0.2","class="num">0.1"},

◍ 样本特征向量的原始数据形态

上面四组大括号里的数字，是某套识别模型喂给 MT5 的归一化特征样本。每组第一个字段是样本编号（如 6615、410、15027、14157），后面 36 个值都在 -1.0 到 1.0 之间，对应 36 个技术维度的标准化偏离。直接读能发现一些分布规律：编号 6615 那组里出现两个极端值 1.0 和 -1.0，且正负数交替频繁，说明该样本在多周期动量上呈现剧烈拉锯；编号 410 这组负向值偏多，末端落到 -1.0，空方维度占优的概率更高。外汇与贵金属属高风险品种，这类向量只是模型输入快照，不能单独作为开仓依据。把这些行直接贴进 EA 的二维数组初始化块，就能在策略测试器里复现对应样本的匹配过程。

MQL5 / C++

  {"class="num">6615","-class="num">0.4","-class="num">0.1","-class="num">0.7","class="num">0.5","-class="num">0.9","class="num">0.4","-class="num">0.9","class="num">0.4","-class="num">0.4","-class="num">0.1","class="num">0.7","-class="num">0.4","class="num">0.4","class="num">0.4","-class="num">0.8","-class="num">0.2","-class="num">0.6","-class="num">0.1","-class="num">0.5","-class="num">0.7","class="num">0.6","class="num">0.0","class="num">1.0","class="num">0.9","-class="num">0.3","class="num">0.8","class="num">0.8","-class="num">0.1","-class="num">0.2","class="num">0.9","-class="num">0.2","class="num">0.9","-class="num">0.8","-class="num">0.6","class="num">0.5"},
  {"class="num">410","-class="num">0.3","class="num">0.2","-class="num">0.2","-class="num">0.2","class="num">0.2","-class="num">0.5","class="num">0.8","class="num">0.3","-class="num">0.9","-class="num">0.9","-class="num">0.4","class="num">0.3","-class="num">0.8","-class="num">0.8","class="num">0.0","class="num">0.9","-class="num">0.2","class="num">0.0","-class="num">0.2","-class="num">0.4","-class="num">0.1","class="num">0.1","-class="num">0.4","class="num">0.7","class="num">1.0","class="num">0.1","class="num">0.5","class="num">0.3","class="num">0.1","class="num">0.7","class="num">0.4","class="num">0.0","-class="num">0.2","-class="num">1.0","-class="num">0.1"},
  {"class="num">15027","-class="num">0.3","-class="num">0.4","-class="num">0.6","class="num">0.3","-class="num">0.5","-class="num">0.6","class="num">0.9","class="num">0.5","-class="num">0.2","class="num">0.0","-class="num">0.7","class="num">0.7","class="num">0.1","class="num">0.5","-class="num">0.4","-class="num">0.4","class="num">0.4","class="num">0.7","-class="num">0.1","class="num">0.9","-class="num">0.1","class="num">0.6","class="num">0.5","-class="num">0.3","class="num">0.6","class="num">0.8","class="num">0.4","class="num">0.1","class="num">0.9","-class="num">0.5","class="num">0.7","class="num">0.6","-class="num">0.8","-class="num">0.1","class="num">0.0"},
  {"class="num">14157","class="num">0.6","-class="num">0.7","class="num">0.7","class="num">0.5","class="num">0.8","-class="num">0.1","class="num">0.9","class="num">0.8","class="num">0.8","class="num">0.7","class="num">0.6","-class="num">0.3","-class="num">0.7","-class="num">0.5","-class="num">0.2","class="num">0.2","class="num">0.0","-class="num">0.8","class="num">0.6","class="num">0.9","-class="num">0.4","class="num">0.1","class="num">0.1","class="num">0.9","class="num">0.7","-class="num">0.8","-class="num">0.6","-class="num">0.5","-class="num">0.7","class="num">0.1","-class="num">0.3","class="num">0.9","class="num">0.5","class="num">0.8","-class="num">0.7"}

「把两段样本塞进模式数组里」

上面这两组字符串数组，是某特征模式下提取出的两条样本记录。第一组 ID 为 11367，共 37 个数值，首尾分别是 0.2 与 0.9，中间有 4 个低于 -0.9 的极值（-1.0 出现两次、-0.9 一次）。第二组 ID 为 3892，长度 37，正负交错更密，0.9 出现两次且集中在第 21、22 位，最低读到 -0.8。两组都按固定列数排列，说明背后是同一套 37 维特征模板。在 MT5 里做模式识别时，直接把这些行粘进你自己的 double patterns[][37] 初始化块即可。外汇与贵金属行情受杠杆与跳空影响大，这类数值仅是历史片段，后续匹配命中只代表形态相似概率，不预示方向。

MQL5 / C++

  {"class="num">11367","class="num">0.2","-class="num">1.0","-class="num">0.4","-class="num">0.4","-class="num">0.3","-class="num">0.2","class="num">0.2","-class="num">0.1","-class="num">0.4","class="num">0.7","-class="num">1.0","-class="num">0.5","-class="num">0.9","-class="num">0.7","-class="num">0.4","-class="num">0.8","-class="num">0.4","class="num">0.0","class="num">0.2","class="num">0.7","-class="num">0.2","class="num">0.4","class="num">0.1","class="num">0.0","-class="num">0.1","-class="num">0.9","class="num">0.2","-class="num">0.5","-class="num">0.6","-class="num">0.6","-class="num">0.7","-class="num">0.2","-class="num">0.3","-class="num">0.1","class="num">0.9"},
  {"class="num">3892","-class="num">0.7","-class="num">0.3","class="num">0.8","class="num">0.2","-class="num">0.3","class="num">0.4","class="num">0.0","class="num">0.3","-class="num">0.2","class="num">0.7","class="num">0.6","class="num">0.6","class="num">0.7","-class="num">0.4","-class="num">0.7","class="num">0.4","-class="num">0.3","-class="num">0.8","-class="num">0.2","class="num">0.0","class="num">0.9","class="num">0.9","class="num">0.3","class="num">0.0","class="num">0.7","class="num">0.1","-class="num">0.1","class="num">0.1","-class="num">0.8","-class="num">0.4","-class="num">0.5","class="num">0.9","-class="num">0.7","-class="num">0.6","class="num">0.2"}

4-4-3 倾角 EA 的前向检验塌了

这套 EA 用固定止损加止盈离场，信号来自策略 4 的倾角 TEMA 指标。回测优化里看着不差：盈利因子落在 1.6 到 5 之间，复杂准则值超过 80 的有 27 个，说明样本内拟合出了不少漂亮参数。但把优化参数丢进前向验证（样本外）立刻露馅——结果转负且极不稳定，没有通过检验。外汇和贵金属杠杆高，这种样本内好、样本外崩的曲线，大概率只是过度拟合，别直接上实盘。开 MT5 把同一组参数切到 Forward 模式跑一遍，比看任何优化报告都实在。

◍ 4-4-3 用神经网络管离场的实测表现

这一版 EA 把离场交给神经网络，进场仍沿用策略 4 的倾角逻辑搭配 TEMA 指标。优化空间里复杂准则跑出来的优良结果（评分 >80）只有 6 个，盈利因子集中在 1.6 到 1.9 之间，说明参数冗余度不高，不是随便调都能出彩。前向验证里按神经网络信号平仓，EA 在全年样本中都维持了盈利状态。对比同样策略挂固定止损止盈的版本，神经离场的权益曲线更平滑，回撤段更短，稳定性确实更占优。外汇与贵金属品种杠杆高、跳空频繁，神经离场依赖历史样本分布，换周期或换品种前要重跑验证，实盘仍可能失效。

「8-4-3 倾角 EA 的前向验证为何塌了」

这套 EA 用固定止损加止盈离场，信号层走的是策略 8 的倾角 TEMA 逻辑。回溯优化里，有 13 个复杂准则的盈利因子跑过了 80，表面看不弱。但和神经网络 4-4-3 那版比，它的整体盈利因子明显更低，说明倾角类规则在样本内靠堆准则刷出了高值，样本外并不买账。前向验证直接失败，走势和上一版用止损止盈的 EA 近似——都是样本内漂亮、推出来就散。外汇与贵金属杠杆高，这类过拟合 EA 实盘触发连亏的概率偏大，开 MT5 用默认参数跑 2020—2023 欧美周线就能复现塌方。

八四三角度 EA 的神经网络离场实测

这个 EA 把离场交给神经网络处理，入场沿用策略 8 的倾角逻辑，也就是以 TEMA 指标的斜率方向作为触发依据。优化环节里，只有 3 条复杂准则的评分突破了 80 分门槛，而对应的盈利因子明显低于前几组对照测试，说明过拟合风险偏高。前向验证直接露了怯：账户本金呈逐步萎缩走势，没有看到样本外稳住曲线的迹象。外汇与贵金属品种杠杆高，这类样本外亏损的离场模型切忌直接上实盘。

◍ 下一步该验证什么

前向测试跑满 6 个月后，所有基于感知器的 EA 年内权益都没有转负，但盈利曲线普遍走弱，说明至少每 6 个月要重新优化一次参数。基于 DeepNeuralNetwork 函数库的版本表现更复杂，结果不如预期，策略本身可能也需要把别的因子喂给网络。盈利因子（Profit Factor）在多数情况下能跟踪优化序列的盈利变化，这给调参提供了额外线索。有读者在 H1 上用 2019.12.9–2021.12.9 训练、2021.12.9–2022.12.9 前向测试，原参数利润 2758 美元，手动把 TP 调到 150、SL 调到 550 后利润升到 7915 美元；再跑 GA 优化出 TP=150、SL=524 得到 8973 美元回测利润，但前向直接破产——TP/SL 必须放进 GA 一起跑才稳。后续两个方向值得做：换其它货币对和时间帧复现最佳结果，以及把多币种合成投资组合来摊薄单系交易次数偏少的问题，但这会不可避免拉高劳力成本。下面这段第三方改的 RSI 跨周期感知器代码可以直接丢进 MT5 编译验证，看多周期斜率加权能不能在你常用的品种上给出更平滑的信号。外汇和贵金属杠杆高、滑点跳空频繁，任何神经网络 EA 都只是概率优势，实盘前请用策略测试器跑满前向样本。

MQL5 / C++

class=class="str">"cmt">//+------------------------------------------------------------------+
class=class="str">"cmt">//| 感知器（PERCEPTRON）--感知和识别功能
class=class="str">"cmt">//+------------------------------------------------------------------+
class="type">class="kw">double perceptron3() class=class="str">"cmt">//RSI
{
   class="type">class="kw">double v1 = z1 - class="num">10.0;
   class="type">class="kw">double v2 = z2 - class="num">10.0;
   class="type">class="kw">double v3 = z3 - class="num">10.0;
   class="type">class="kw">double v4 = z4 - class="num">10.0;   
   class=class="str">"cmt">//In3 = RSI Current timeframe
   class=class="str">"cmt">//In4 = RSI Higher timeframe
   
   class="type">class="kw">double b1 = (((ind_In3[class="num">1]-ind_In3[class="num">2]))/class="num">2);
   class="type">class="kw">double b2 = (((ind_In4[class="num">1]-ind_In3[class="num">2]))/class="num">2);
   class="type">class="kw">double b3 = (((ind_In3[class="num">1]-ind_In4[class="num">2]))/class="num">2);
   class="type">class="kw">double b4 = (((ind_In4[class="num">1]-ind_In4[class="num">2]))/class="num">2);
   
   class="kw">return (v1 * b1 + v2 * b2 + v3 * b3 + v4 * b4);
}

让小布替你跑这套

把多组感知器权重和DNN配置交给小布盯盘的AIGC诊断，打开对应品种页就能直接看到各组合在历史与前瞻区间的回撤分布，你只管挑哪组更顺手。

←

神经网络实验（第 3 部分）：实际应用·进阶篇

常见问题

大量观察显示过深权重对EURUSD H1的TEMA组合无明显增益，20层在交易准确性与优化耗时间更平衡。

它控制向感知器某一侧回撤的深度，调深会减少成交次数但提高触发精度，需按品种波动习惯折中。

1年仅作基础验证，原文也提示3年优化范围非铁律，可自扩样本；结论倾向概率性而非必然稳健。

小布盯盘内置了多组合诊断视图，可导入优化权重集查看前瞻表现，免去手动挂十个图表的麻烦。

隐层节点数不同改变特征抽取粒度，8输入版本覆盖角度更全，带SL TP的变体在风控上更收敛。

神经网络实验（第 3 部分）：实际应用·综合运用

双四倾角感知器的裸奔平仓逻辑

◍ 用 DNN 函数库跑倾角策略的四种 EA 组合

「样本集里的归一化特征向量」

样本序列里的符号化收益分布

◍ 样本序列里的归一化残差怎么读

「训练样本里的归一化特征向量」

样本序列的归一化数值排布

◍ 样本特征向量的原始数据形态

「把两段样本塞进模式数组里」

4-4-3 倾角 EA 的前向检验塌了

◍ 4-4-3 用神经网络管离场的实测表现

「8-4-3 倾角 EA 的前向验证为何塌了」

八四三角度 EA 的神经网络离场实测

◍ 下一步该验证什么

常见问题

相关阅读