🧠

神经网络实验（第 4 部分）：模板·综合运用

（3/3）·从扇形平行四边形到三角形，手把手跑完 20 次优化与一年前向验证的全流程

新手友好第 3/3 篇

把价格硬塞进神经网络前，先想想你传进去的到底是信号还是噪声。很多交易者直接拿原始 K 线训练，结果散射太宽连简单感知器都不如。用形态模板收束输入范围，才是让 MQL5 神经网出稳定结果的实际起点。

本章目录

把多周期收盘价差压成百点整数栅格
用 1200 点偏移把 K 线距离整数化
用感知机把多周期价差压成整百点距
把蜡烛拆成四个偏移量喂给感知机
把K线距离折算成整百点距
感知机里的偏移与价格区间离散化
把 K 线各段占比量化成整百点值
感知机里的价格偏移量化
用百点取整压缩蜡烛形态特征
用偏移中枢量化多周期收盘落差
神经网络EA的前向验证实况
画得少，看得清

把多周期收盘价差压成百点整数栅格

这段逻辑在做一件事：把当前品种不同 K 线收盘价的相对位移，折算成以 100 点（Point）为步长的整数，塞进 xInputs 数组供后续感知机调用。注意它给远端收盘价统一叠加了 800*Point 的偏移，相当于在价格差里预置了一个 80 点（对多数外汇对是 8 个点）的缓冲带，避免贴近现价时噪声触发误判。以 a4 为例：用 13 根 K 前收盘减去（1 根前收盘 + 800 点），再除以 Point 取整，最后 (a4/100)*100 把结果吸附到最近的百点整数。xInputs[3] 到 [6] 都是类似结构，只是把 13/17/21 等不同滞后位置与 24 根前收盘做差值对照。下面这段 perceptron1 把外部传入的 x1~x4、y1~y4 各减 10.0 得到权重 w、v 序列，又补了一段 a1、a2 的计算：a1 取 1 根前与 48 根前收盘各自加 1200 点后的均值减现价，同样百点化；a2 则是 9 根前减（现价+1200点）。外汇与贵金属杠杆高，这类栅格化特征对点值敏感，直接上实盘前务必在 MT5 策略测试器用历史数据核对偏移量是否匹配你的品种小数点位数。

MQL5 / C++

class="type">int a4 = (class="type">int)((iClose(symbolS1.Name(),PERIOD_CURRENT,class="num">13)-(iClose(symbolS1.Name(),PERIOD_CURRENT,class="num">1)+(class="num">800*Point())))/Point());
 xInputs[class="num">3] = (class="type">int)(a4/class="num">100)*class="num">100;
 class="type">int g1 = (class="type">int)(((iClose(symbolS1.Name(),PERIOD_CURRENT,class="num">24)+(class="num">800*Point()))-iClose(symbolS1.Name(),PERIOD_CURRENT,class="num">13))/Point());
 xInputs[class="num">4] = (class="type">int)(g1/class="num">100)*class="num">100;
 class="type">int g2 = (class="type">int)(((iClose(symbolS1.Name(),PERIOD_CURRENT,class="num">24)+(class="num">800*Point()))-iClose(symbolS1.Name(),PERIOD_CURRENT,class="num">17))/Point());
 xInputs[class="num">5] = (class="type">int)(g2/class="num">100)*class="num">100;
 class="type">int g3 = (class="type">int)(((iClose(symbolS1.Name(),PERIOD_CURRENT,class="num">24)-(class="num">800*Point()))-iClose(symbolS1.Name(),PERIOD_CURRENT,class="num">21))/Point());
 xInputs[class="num">6] = (class="type">int)(g3/class="num">100)*class="num">100;
 class="type">int b4 = (class="type">int)((iClose(symbolS1.Name(),PERIOD_CURRENT,class="num">24)-((iClose(symbolS1.Name(),PERIOD_CURRENT,class="num">1)-(class="num">800*Point()))+(iClose(symbolS1.Name(),PERIOD_CURRENT,class="num">24)-(class="num">800*Point()))/class="num">2))/Point()) ;
 xInputs[class="num">7] = (class="type">int)(g4/class="num">100)*class="num">100;

 class="kw">return(class="num">1);
}
class=class="str">"cmt">//+------------------------------------------------------------------+
class=class="str">"cmt">//| The PERCEPRRON - a perceiving and recognizing function |
class=class="str">"cmt">//+------------------------------------------------------------------+
class="type">class="kw">double perceptron1()
 {
 class="type">class="kw">double w1 = x1 - class="num">10.0;
 class="type">class="kw">double w2 = x2 - class="num">10.0;
 class="type">class="kw">double w3 = x3 - class="num">10.0;
 class="type">class="kw">double w4 = x4 - class="num">10.0;

 class="type">class="kw">double v1 = y1 - class="num">10.0;
 class="type">class="kw">double v2 = y2 - class="num">10.0;
 class="type">class="kw">double v3 = y3 - class="num">10.0;
 class="type">class="kw">double v4 = y4 - class="num">10.0;

 class="type">int a1 = (class="type">int)((((iClose(symbolS1.Name(),PERIOD_CURRENT,class="num">1)+(class="num">1200*Point()))+(iClose(symbolS1.Name(),PERIOD_CURRENT,class="num">48)+(class="num">1200*Point()))/class="num">2)-iClose(symbolS1.Name(),PERIOD_CURRENT,class="num">1))/Point());
 a1 = (class="type">int)(a1/class="num">100)*class="num">100;
 class="type">int a2 = (class="type">int)((iClose(symbolS1.Name(),PERIOD_CURRENT,class="num">9)-(iClose(symbolS1.Name(),PERIOD_CURRENT,class="num">1)+(class="num">1200*Point())))/Point());
 a2 = (class="type">int)(a2/class="num">100)*class="num">100;

「用 1200 点偏移把 K 线距离整数化」

这段逻辑在做一件事：把不同周期收盘价的差值，先加减 1200*Point 的偏移，再除以 Point 转成整数点，最后用 /100*100 向下取整到百点粒度。偏移量 1200 点在 EURUSD 这类 Point=0.00001 的品种上等于 12 个基点，相当于人为给距离判读加了一道缓冲带，避免毛刺触发误判。以 a3 为例，它取第 17 根收盘价与「第 1 根收盘价减 1200 点」的差，再转点取整到百点；a4 则换成第 25 根与「第 1 根加 1200 点」的差。b1~b4 反过来用第 48 根加/减 1200 点去减更早的收盘价，覆盖 25/33/41/1 号 bar，构成一组对称的距离采样。 CandlePatterns 里把结果写进 xInputs[0]，但注意 a1 和 a2 连续两次覆盖同一槽位——后写的 a2 会直接冲掉 a1，这大概率是调试残留。开 MT5 把这段贴进 EA，打印 xInputs[0] 你会看到它只反映 a2 的百点值，外汇贵金属波动剧烈，这种覆盖 bug 可能让样本特征失真。

MQL5 / C++

class="type">int a3 = (class="type">int)((iClose(symbolS1.Name(),PERIOD_CURRENT,class="num">17)-(iClose(symbolS1.Name(),PERIOD_CURRENT,class="num">1)-(class="num">1200*Point())))/Point());
 a3 = (class="type">int)(a3/class="num">100)*class="num">100;
 class="type">int a4 = (class="type">int)((iClose(symbolS1.Name(),PERIOD_CURRENT,class="num">25)-(iClose(symbolS1.Name(),PERIOD_CURRENT,class="num">1)+(class="num">1200*Point())))/Point());
 a4 = (class="type">int)(a4/class="num">100)*class="num">100;
 class="type">int b1 = (class="type">int)(((iClose(symbolS1.Name(),PERIOD_CURRENT,class="num">48)+(class="num">1200*Point()))-iClose(symbolS1.Name(),PERIOD_CURRENT,class="num">25))/Point());
 b1 = (class="type">int)(b1/class="num">100)*class="num">100;
 class="type">int b2 = (class="type">int)(((iClose(symbolS1.Name(),PERIOD_CURRENT,class="num">48)+(class="num">1200*Point()))-iClose(symbolS1.Name(),PERIOD_CURRENT,class="num">33))/Point());
 b2 = (class="type">int)(b2/class="num">100)*class="num">100;
 class="type">int b3 = (class="type">int)(((iClose(symbolS1.Name(),PERIOD_CURRENT,class="num">48)-(class="num">1200*Point()))-iClose(symbolS1.Name(),PERIOD_CURRENT,class="num">41))/Point());
 b3 = (class="type">int)(b3/class="num">100)*class="num">100;
 class="type">int b4 = (class="type">int)((iClose(symbolS1.Name(),PERIOD_CURRENT,class="num">48)-((iClose(symbolS1.Name(),PERIOD_CURRENT,class="num">1)-(class="num">1200*Point()))+(iClose(symbolS1.Name(),PERIOD_CURRENT,class="num">48)-(class="num">1200*Point()))/class="num">2))/Point()) ;
 b4 = (class="type">int)(b4/class="num">100)*class="num">100;
 class="kw">return (w1 * a1 + w2 * a2 + w3 * a3 + w4 * a4   +   v1 * b1 + v2 * b2 + v3 * b3 + v4 * b4);
 }
class=class="str">"cmt">//+------------------------------------------------------------------+
class=class="str">"cmt">//|percentage of each part of the candle respecting total size       |
class=class="str">"cmt">//+------------------------------------------------------------------+
class="type">int CandlePatterns(class="type">class="kw">double &xInputs[])
  {
  class="type">int a1 = (class="type">int)((((iClose(symbolS1.Name(),PERIOD_CURRENT,class="num">1)+(class="num">1200*Point()))+(iClose(symbolS1.Name(),PERIOD_CURRENT,class="num">48)+(class="num">1200*Point()))/class="num">2)-iClose(symbolS1.Name(),PERIOD_CURRENT,class="num">1))/Point());
  xInputs[class="num">0] = (class="type">int)(a1/class="num">100)*class="num">100;
  class="type">int a2 = (class="type">int)((iClose(symbolS1.Name(),PERIOD_CURRENT,class="num">9)-(iClose(symbolS1.Name(),PERIOD_CURRENT,class="num">1)+(class="num">1200*Point())))/Point());
  xInputs[class="num">0] = (class="type">int)(a2/class="num">100)*class="num">100;

◍ 用感知机把多周期价差压成整百点距

这段逻辑在干一件事：把不同偏移量的收盘价落差换算成以『点』为单位的整数，再向下取整到百位，塞进 xInputs[0] 做后续感知机输入。注意它反复覆盖同一个数组元素，说明此处只是分步调试各周期差值的观测段，并非最终并行特征。以 a3 为例：取 17 根 K 线前收盘，减去（最新收盘减 1200 点），除以 Point() 转成整数点距。1200 点在外汇黄金里约等于常见品种 12 美元跳动，属于人为加宽的缓冲带，用来过滤毛刺。 g4 那行最绕：用 48 根前收盘，减去『最新减1200点』与『48根前减1200点』的均值，再除 Point()。它本质上在量 48 周期中轨和近端下沿的偏离，结果同样被百点化。 perceptron1() 里 w1~w4 各自减 10.0，再把 a1 算成『(最新+800点+24根前)/2 与最新的差』的百点整数。外汇和贵金属杠杆高，这类点距阈值直接上实盘前，务必在 MT5 策略测试器用对应品种点值核对，否则不同 broker 的 Point() 精度会让信号偏移。

MQL5 / C++

class="type">int a3 = (class="type">int)((iClose(symbolS1.Name(),PERIOD_CURRENT,class="num">17)-(iClose(symbolS1.Name(),PERIOD_CURRENT,class="num">1)-(class="num">1200*Point())))/Point());
 xInputs[class="num">0] = (class="type">int)(a3/class="num">100)*class="num">100;
 class="type">int a4 = (class="type">int)((iClose(symbolS1.Name(),PERIOD_CURRENT,class="num">25)-(iClose(symbolS1.Name(),PERIOD_CURRENT,class="num">1)+(class="num">1200*Point())))/Point());
 xInputs[class="num">0] = (class="type">int)(a4/class="num">100)*class="num">100;
 class="type">int g1 = (class="type">int)(((iClose(symbolS1.Name(),PERIOD_CURRENT,class="num">48)+(class="num">1200*Point()))-iClose(symbolS1.Name(),PERIOD_CURRENT,class="num">25))/Point());
 xInputs[class="num">0] = (class="type">int)(g1/class="num">100)*class="num">100;
 class="type">int g2 = (class="type">int)(((iClose(symbolS1.Name(),PERIOD_CURRENT,class="num">48)+(class="num">1200*Point()))-iClose(symbolS1.Name(),PERIOD_CURRENT,class="num">33))/Point());
 xInputs[class="num">0] = (class="type">int)(g2/class="num">100)*class="num">100;
 class="type">int g3 = (class="type">int)(((iClose(symbolS1.Name(),PERIOD_CURRENT,class="num">48)-(class="num">1200*Point()))-iClose(symbolS1.Name(),PERIOD_CURRENT,class="num">41))/Point());
 xInputs[class="num">0] = (class="type">int)(g3/class="num">100)*class="num">100;
 class="type">int g4 = (class="type">int)((iClose(symbolS1.Name(),PERIOD_CURRENT,class="num">48)-((iClose(symbolS1.Name(),PERIOD_CURRENT,class="num">1)-(class="num">1200*Point()))+(iClose(symbolS1.Name(),PERIOD_CURRENT,class="num">48)-(class="num">1200*Point()))/class="num">2))/Point()) ;
 xInputs[class="num">0] = (class="type">int)(g4/class="num">100)*class="num">100;

 class="kw">return(class="num">1);
}
class=class="str">"cmt">//+------------------------------------------------------------------+
class=class="str">"cmt">//| The PERCEPRRON - a perceiving and recognizing function |
class=class="str">"cmt">//+------------------------------------------------------------------+
class="type">class="kw">double perceptron1()
 {
 class="type">class="kw">double w1 = x1 - class="num">10.0;
 class="type">class="kw">double w2 = x2 - class="num">10.0;
 class="type">class="kw">double w3 = x3 - class="num">10.0;
 class="type">class="kw">double w4 = x4 - class="num">10.0;

 class="type">int a1 = (class="type">int)((((iClose(symbolS1.Name(),PERIOD_CURRENT,class="num">1)+(class="num">800*Point())+iClose(symbolS1.Name(),PERIOD_CURRENT,class="num">24))/class="num">2)-iClose(symbolS1.Name(),PERIOD_CURRENT,class="num">1))/Point());
 a1 = (class="type">int)(a1/class="num">100)*class="num">100;

把蜡烛拆成四个偏移量喂给感知机

这段逻辑干的事很直接：用不同 bar 的收盘价差值，换算成以点为单位的整数偏移，再统一向下取整到百点整数倍。a1 到 a4 分别抓取了近一根、第 7 根、第 13 根、第 19 根与第 24 根收盘之间的相对距离，其中 a2 还叠了一个 800 点的平移补偿。 CandlePatterns 函数把同样的计算结果写进 xInputs 数组，下标 0~3 对应那四个偏移量，方便后续模块直接读取。注意它永远 return 1，真正的信号权重在另一个 perceptron1 里用 w1=x1-10 这类减法做偏移修正。外汇与贵金属市场杠杆高、滑点随机，这套特征构造在 EURUSD 这类点值为 0.0001 的品种上，800 点平移相当于 0.08 价格单位，回测时若换到 XAUUSD 必须重算 Point 量级，否则特征会整体失真。

MQL5 / C++

class="type">int a2 = (class="type">int)((((iClose(symbolS1.Name(),PERIOD_CURRENT,class="num">1)-(class="num">800*Point())+iClose(symbolS1.Name(),PERIOD_CURRENT,class="num">24))/class="num">2)-iClose(symbolS1.Name(),PERIOD_CURRENT,class="num">7))/Point());
 a2 = (class="type">int)(a2/class="num">100)*class="num">100;
 class="type">int a3 = (class="type">int)((iClose(symbolS1.Name(),PERIOD_CURRENT,class="num">24)-iClose(symbolS1.Name(),PERIOD_CURRENT,class="num">13))/Point());
 a3 = (class="type">int)(a3/class="num">100)*class="num">100;
 class="type">int a4 = (class="type">int)((iClose(symbolS1.Name(),PERIOD_CURRENT,class="num">24)-iClose(symbolS1.Name(),PERIOD_CURRENT,class="num">19))/Point());
 a4 = (class="type">int)(a4/class="num">100)*class="num">100;
 class="kw">return (w1 * a1 + w2 * a2 + w3 * a3 + w4 * a4);
}
class=class="str">"cmt">//+------------------------------------------------------------------+
class=class="str">"cmt">//|percentage of each part of the candle respecting total size          |
class=class="str">"cmt">//+------------------------------------------------------------------+
class="type">int CandlePatterns(class="type">class="kw">double &xInputs[])
  {
  class="type">int a1 = (class="type">int)((((iClose(symbolS1.Name(),PERIOD_CURRENT,class="num">1)+(class="num">800*Point())+iClose(symbolS1.Name(),PERIOD_CURRENT,class="num">24))/class="num">2)-iClose(symbolS1.Name(),PERIOD_CURRENT,class="num">1))/Point());
  xInputs[class="num">0] = (class="type">int)(a1/class="num">100)*class="num">100;
  class="type">int a2 = (class="type">int)((((iClose(symbolS1.Name(),PERIOD_CURRENT,class="num">1)-(class="num">800*Point())+iClose(symbolS1.Name(),PERIOD_CURRENT,class="num">24))/class="num">2)-iClose(symbolS1.Name(),PERIOD_CURRENT,class="num">7))/Point());
  xInputs[class="num">1] = (class="type">int)(a2/class="num">100)*class="num">100;
  class="type">int a3 = (class="type">int)((iClose(symbolS1.Name(),PERIOD_CURRENT,class="num">24)-iClose(symbolS1.Name(),PERIOD_CURRENT,class="num">13))/Point());
  xInputs[class="num">2] = (class="type">int)(a3/class="num">100)*class="num">100;
  class="type">int a4 = (class="type">int)((iClose(symbolS1.Name(),PERIOD_CURRENT,class="num">24)-iClose(symbolS1.Name(),PERIOD_CURRENT,class="num">19))/Point());
  xInputs[class="num">3] = (class="type">int)(a4/class="num">100)*class="num">100;
  class="kw">return(class="num">1);
}
class=class="str">"cmt">//+------------------------------------------------------------------+
class=class="str">"cmt">//|  The PERCEPRRON - a perceiving and recognizing function            |
class=class="str">"cmt">//+------------------------------------------------------------------+
class="type">class="kw">double perceptron1()
  {
  class="type">class="kw">double w1 = x1 - class="num">10.0;
  class="type">class="kw">double w2 = x2 - class="num">10.0;
  class="type">class="kw">double w3 = x3 - class="num">10.0;
  class="type">class="kw">double w4 = x4 - class="num">10.0;

「把K线距离折算成整百点距」

这段逻辑干的事很直白：用当前价和若干根历史收盘价的差，除以 Point() 得到以点为单位的数值，再向下取整到百点整数倍。a1 取的是「(现价+1200点+48根前收盘)/2」与现价的差，a2 换成减 1200 点后和 13 根前收盘比，a3、a4 则纯看 48 根前收盘相对 25、37 根前收盘的落点。注意 1200*Point() 这个偏移量，在常规外汇品种上约等于 1200 点（若是五位数报价就是 120 点，需按品种确认），它把中枢线人为外推，避免和现价贴太死。除以 Point() 转成整数点距后，(int)(x/100)*100 把噪声削到百点粒度——你开 MT5 把这段贴进 EA，改 100 为 50 就能拿到更细的档位。 CandlePatterns 函数把同样的四个量写进 xInputs 数组而非直接返回加权值，调用方便但容易让人忽略：a1~a4 的计算式和前一个函数完全一致，只是出口不同。外汇与贵金属波动剧烈，这类点距特征仅描述形态偏离，不预示方向，实盘须自担高风险。

MQL5 / C++

  class="type">int a1 = (class="type">int)((((iClose(symbolS1.Name(),PERIOD_CURRENT,class="num">1)+(class="num">1200*Point())+iClose(symbolS1.Name(),PERIOD_CURRENT,class="num">48))/class="num">2)-iClose(symbolS1.Name(),PERIOD_CURRENT,class="num">1))/Point());
  a1 = (class="type">int)(a1/class="num">100)*class="num">100;
  class="type">int a2 = (class="type">int)((((iClose(symbolS1.Name(),PERIOD_CURRENT,class="num">1)-(class="num">1200*Point())+iClose(symbolS1.Name(),PERIOD_CURRENT,class="num">48))/class="num">2)-iClose(symbolS1.Name(),PERIOD_CURRENT,class="num">13))/Point());
  a2 = (class="type">int)(a2/class="num">100)*class="num">100;
  class="type">int a3 = (class="type">int)((iClose(symbolS1.Name(),PERIOD_CURRENT,class="num">48)-iClose(symbolS1.Name(),PERIOD_CURRENT,class="num">25))/Point());
  a3 = (class="type">int)(a3/class="num">100)*class="num">100;
  class="type">int a4 = (class="type">int)((iClose(symbolS1.Name(),PERIOD_CURRENT,class="num">48)-iClose(symbolS1.Name(),PERIOD_CURRENT,class="num">37))/Point());
  a4 = (class="type">int)(a4/class="num">100)*class="num">100;
  class="kw">return (w1 * a1 + w2 * a2 + w3 * a3 + w4 * a4);
  }
class=class="str">"cmt">//+------------------------------------------------------------------+
class=class="str">"cmt">//|percentage of each part of the candle respecting total size        |
class=class="str">"cmt">//+------------------------------------------------------------------+
class="type">int CandlePatterns(class="type">class="kw">double &xInputs[])
  {
  class="type">int a1 = (class="type">int)((((iClose(symbolS1.Name(),PERIOD_CURRENT,class="num">1)+(class="num">1200*Point())+iClose(symbolS1.Name(),PERIOD_CURRENT,class="num">48))/class="num">2)-iClose(symbolS1.Name(),PERIOD_CURRENT,class="num">1))/Point());
  xInputs[class="num">0] = (class="type">int)(a1/class="num">100)*class="num">100;
  class="type">int a2 = (class="type">int)((((iClose(symbolS1.Name(),PERIOD_CURRENT,class="num">1)-(class="num">1200*Point())+iClose(symbolS1.Name(),PERIOD_CURRENT,class="num">48))/class="num">2)-iClose(symbolS1.Name(),PERIOD_CURRENT,class="num">13))/Point());
  xInputs[class="num">1] = (class="type">int)(a2/class="num">100)*class="num">100;
  class="type">int a3 = (class="type">int)((iClose(symbolS1.Name(),PERIOD_CURRENT,class="num">48)-iClose(symbolS1.Name(),PERIOD_CURRENT,class="num">25))/Point());
  xInputs[class="num">2] = (class="type">int)(a3/class="num">100)*class="num">100;
  class="type">int a4 = (class="type">int)((iClose(symbolS1.Name(),PERIOD_CURRENT,class="num">48)-iClose(symbolS1.Name(),PERIOD_CURRENT,class="num">37))/Point());
  xInputs[class="num">3] = (class="type">int)(a4/class="num">100)*class="num">100;
  class="kw">return(class="num">1);
  }
class=class="str">"cmt">//+------------------------------------------------------------------+

◍ 感知机里的偏移与价格区间离散化

这段函数把感知机思路搬进了 MT5 指标计算：先用 x1~x4、y1~y4 各自减去常数 10.0 得到权重偏移 w1~w4、v1~v4，相当于给输入特征做固定去均值处理，避免后续距离计算被绝对价位带偏。真正的信号特征来自 a1~a4 与 b1~b3 这几组整数化差值。以 a1 为例，它取当前收盘价加 800 点、与 24 根 K 线前收盘价的均值，再减当前收盘价，除以 Point 后取整并百点对齐——把「近端价格相对中段均值的偏离」压成每 100 点一档的离散量。a2 则用减 800 点替代加 800 点，捕捉下方偏移。 a3、a4 更直接：分别是 24 周期与 13、19 周期收盘价的差值，同样百点取整。b 组换用了 4、10、16 周期作参照，构成另一套偏移剖面。外汇与贵金属杠杆高，这类离散特征只描述价格结构，不预示方向。开 MT5 把下面代码贴进自定义指标，改 PERIOD_CURRENT 为具体周期，观察 a1 在 EURUSD 的 H1 上是否常在 ±800 点边界处跳档，就能验证这套离散化对区间突破的灵敏度。

MQL5 / C++

class="type">class="kw">double perceptron1()
  {
   class="type">class="kw">double w1 = x1 - class="num">10.0;
   class="type">class="kw">double w2 = x2 - class="num">10.0;
   class="type">class="kw">double w3 = x3 - class="num">10.0;
   class="type">class="kw">double w4 = x4 - class="num">10.0;

   class="type">class="kw">double v1 = y1 - class="num">10.0;
   class="type">class="kw">double v2 = y2 - class="num">10.0;
   class="type">class="kw">double v3 = y3 - class="num">10.0;
   class="type">class="kw">double v4 = y4 - class="num">10.0;

   class="type">int a1 = (class="type">int)((((iClose(symbolS1.Name(),PERIOD_CURRENT,class="num">1)+(class="num">800*Point())+iClose(symbolS1.Name(),PERIOD_CURRENT,class="num">24))/class="num">2)-iClose(symbolS1.Name(),PERIOD_CURRENT,class="num">1))/Point());
   a1 = (class="type">int)(a1/class="num">100)*class="num">100;
   class="type">int a2 = (class="type">int)((((iClose(symbolS1.Name(),PERIOD_CURRENT,class="num">1)-(class="num">800*Point())+iClose(symbolS1.Name(),PERIOD_CURRENT,class="num">24))/class="num">2)-iClose(symbolS1.Name(),PERIOD_CURRENT,class="num">7))/Point());
   a2 = (class="type">int)(a2/class="num">100)*class="num">100;
   class="type">int a3 = (class="type">int)((iClose(symbolS1.Name(),PERIOD_CURRENT,class="num">24)-iClose(symbolS1.Name(),PERIOD_CURRENT,class="num">13))/Point());
   a3 = (class="type">int)(a3/class="num">100)*class="num">100;
   class="type">int a4 = (class="type">int)((iClose(symbolS1.Name(),PERIOD_CURRENT,class="num">24)-iClose(symbolS1.Name(),PERIOD_CURRENT,class="num">19))/Point());
   a4 = (class="type">int)(a4/class="num">100)*class="num">100;

   class="type">int b1 = (class="type">int)((((iClose(symbolS1.Name(),PERIOD_CURRENT,class="num">1)+(class="num">800*Point())+iClose(symbolS1.Name(),PERIOD_CURRENT,class="num">24))/class="num">2)-iClose(symbolS1.Name(),PERIOD_CURRENT,class="num">4))/Point());
   b1 = (class="type">int)(b1/class="num">100)*class="num">100;
   class="type">int b2 = (class="type">int)((((iClose(symbolS1.Name(),PERIOD_CURRENT,class="num">1)-(class="num">800*Point())+iClose(symbolS1.Name(),PERIOD_CURRENT,class="num">24))/class="num">2)-iClose(symbolS1.Name(),PERIOD_CURRENT,class="num">10))/Point());
   b2 = (class="type">int)(b2/class="num">100)*class="num">100;
   class="type">int b3 = (class="type">int)((((iClose(symbolS1.Name(),PERIOD_CURRENT,class="num">1)-(class="num">800*Point())+iClose(symbolS1.Name(),PERIOD_CURRENT,class="num">24))/class="num">2)-iClose(symbolS1.Name(),PERIOD_CURRENT,class="num">16))/Point());
   b3 = (class="type">int)(b3/class="num">100)*class="num">100;

把 K 线各段占比量化成整百点值

这段逻辑在算当前品种 K 线内部不同价格段相对整根实体的点数差，并统一向下取整到 100 点整数倍，方便后续当离散特征喂给模型。以 a1 为例：取前 1 根收盘与 24 根收盘的平均值（上方偏移 800 点）再减前 1 根收盘，除以 Point() 转成整数点。xInputs[0] 存的是 (a1/100)*100，也就是丢掉了不足 100 点的零头。 a3、a4 则直接拿 24 根与 13、19 根收盘做差，不叠加 800 点偏移，只看中段趋势的位移点数。g1、g2 又回到带 800 点偏移的算法，分别对照 4 根和 10 根收盘。在 MT5 里把 symbolS1 换成你自己盯的黄金或欧美品种，打印 xInputs 数组，能直接看到每根 K 线被压成 6 个整百点特征。外汇和贵金属杠杆高，这类特征只描述形态结构，不预示方向。

MQL5 / C++

class="type">int CandlePatterns(class="type">class="kw">double &xInputs[])
  {
  class="type">int a1 = (class="type">int)((((iClose(symbolS1.Name(),PERIOD_CURRENT,class="num">1)+(class="num">800*Point())+iClose(symbolS1.Name(),PERIOD_CURRENT,class="num">24))/class="num">2)-iClose(symbolS1.Name(),PERIOD_CURRENT,class="num">1))/Point());
  xInputs[class="num">0] = (class="type">int)(a1/class="num">100)*class="num">100;
  class="type">int a2 = (class="type">int)((((iClose(symbolS1.Name(),PERIOD_CURRENT,class="num">1)-(class="num">800*Point())+iClose(symbolS1.Name(),PERIOD_CURRENT,class="num">24))/class="num">2)-iClose(symbolS1.Name(),PERIOD_CURRENT,class="num">7))/Point());
  xInputs[class="num">1] = (class="type">int)(a2/class="num">100)*class="num">100;
  class="type">int a3 = (class="type">int)((iClose(symbolS1.Name(),PERIOD_CURRENT,class="num">24)-iClose(symbolS1.Name(),PERIOD_CURRENT,class="num">13))/Point());
  xInputs[class="num">2] = (class="type">int)(a3/class="num">100)*class="num">100;
  class="type">int a4 = (class="type">int)((iClose(symbolS1.Name(),PERIOD_CURRENT,class="num">24)-iClose(symbolS1.Name(),PERIOD_CURRENT,class="num">19))/Point());
  xInputs[class="num">3] = (class="type">int)(a4/class="num">100)*class="num">100;
  class="type">int g1 = (class="type">int)((((iClose(symbolS1.Name(),PERIOD_CURRENT,class="num">1)+(class="num">800*Point())+iClose(symbolS1.Name(),PERIOD_CURRENT,class="num">24))/class="num">2)-iClose(symbolS1.Name(),PERIOD_CURRENT,class="num">4))/Point());
  xInputs[class="num">4] = (class="type">int)(g1/class="num">100)*class="num">100;
  class="type">int g2 = (class="type">int)((((iClose(symbolS1.Name(),PERIOD_CURRENT,class="num">1)-(class="num">800*Point())+iClose(symbolS1.Name(),PERIOD_CURRENT,class="num">24))/class="num">2)-iClose(symbolS1.Name(),PERIOD_CURRENT,class="num">10))/Point());
  xInputs[class="num">5] = (class="type">int)(g2/class="num">100)*class="num">100;

「感知机里的价格偏移量化」

下面这段 MQL5 片段把多个收盘价组合成整数偏移量，再塞进 xInputs 数组，供后续感知机函数调用。注意所有偏移都除以 Point() 转成点数，并用 (int)(x/100)*100 做了百点取整——这意味着小于 100 点的噪声被直接抹平。 perceptron1() 里先把 x1~x4、y1~y4 各减 10.0 得到 w、v 系列权重偏移，随后用 iClose 取当前周期 1、48、13、25、37 等柱的收盘价。a1 用 (收盘1 + 1200点 + 收盘48)/2 再减收盘1，等价于 600点 + (收盘48-收盘1)/2，百点取整后反映长周期与即时价的半差。外汇与贵金属杠杆高，这类基于历史偏移的算法信号仅代表概率倾向，实盘前务必在 MT5 策略测试器用对应品种点值验证。把下面代码直接挂到 EA 里打印 xInputs[6]、[7] 与 a1~a4，能看到真实百点档位分布。

MQL5 / C++

  class="type">int g3 = (class="type">int)((((iClose(symbolS1.Name(),PERIOD_CURRENT,class="num">1)-(class="num">800*Point())+iClose(symbolS1.Name(),PERIOD_CURRENT,class="num">24))/class="num">2)-iClose(symbolS1.Name(),PERIOD_CURRENT,class="num">16))/Point());
  xInputs[class="num">6] = (class="type">int)(g3/class="num">100)*class="num">100;
  class="type">int g4 = (class="type">int)((((iClose(symbolS1.Name(),PERIOD_CURRENT,class="num">1)-(class="num">800*Point())+iClose(symbolS1.Name(),PERIOD_CURRENT,class="num">24))/class="num">2)-iClose(symbolS1.Name(),PERIOD_CURRENT,class="num">22))/Point());
  xInputs[class="num">7] = (class="type">int)(g4/class="num">100)*class="num">100;

  class="kw">return(class="num">1);
}
class=class="str">"cmt">//+------------------------------------------------------------------+
class=class="str">"cmt">//|  The PERCEPRRON - a perceiving and recognizing function          |
class=class="str">"cmt">//+------------------------------------------------------------------+
class="type">class="kw">double perceptron1()
  {
  class="type">class="kw">double w1 = x1 - class="num">10.0;
  class="type">class="kw">double w2 = x2 - class="num">10.0;
  class="type">class="kw">double w3 = x3 - class="num">10.0;
  class="type">class="kw">double w4 = x4 - class="num">10.0;

  class="type">class="kw">double v1 = y1 - class="num">10.0;
  class="type">class="kw">double v2 = y2 - class="num">10.0;
  class="type">class="kw">double v3 = y3 - class="num">10.0;
  class="type">class="kw">double v4 = y4 - class="num">10.0;

  class="type">int a1 = (class="type">int)((((iClose(symbolS1.Name(),PERIOD_CURRENT,class="num">1)+(class="num">1200*Point())+iClose(symbolS1.Name(),PERIOD_CURRENT,class="num">48))/class="num">2)-iClose(symbolS1.Name(),PERIOD_CURRENT,class="num">1))/Point());
  a1 = (class="type">int)(a1/class="num">100)*class="num">100;
  class="type">int a2 = (class="type">int)((((iClose(symbolS1.Name(),PERIOD_CURRENT,class="num">1)-(class="num">1200*Point())+iClose(symbolS1.Name(),PERIOD_CURRENT,class="num">48))/class="num">2)-iClose(symbolS1.Name(),PERIOD_CURRENT,class="num">13))/Point());
  a2 = (class="type">int)(a2/class="num">100)*class="num">100;
  class="type">int a3 = (class="type">int)((iClose(symbolS1.Name(),PERIOD_CURRENT,class="num">48)-iClose(symbolS1.Name(),PERIOD_CURRENT,class="num">25))/Point());
  a3 = (class="type">int)(a3/class="num">100)*class="num">100;
  class="type">int a4 = (class="type">int)((iClose(symbolS1.Name(),PERIOD_CURRENT,class="num">48)-iClose(symbolS1.Name(),PERIOD_CURRENT,class="num">37))/Point());
  a4 = (class="type">int)(a4/class="num">100)*class="num">100;

  class="type">int b1 = (class="type">int)((((iClose(symbolS1.Name(),PERIOD_CURRENT,class="num">1)+(class="num">1200*Point())+iClose(symbolS1.Name(),PERIOD_CURRENT,class="num">48))/class="num">2)-iClose(symbolS1.Name(),PERIOD_CURRENT,class="num">6))/Point());

◍ 用百点取整压缩蜡烛形态特征

这段逻辑把多个偏移收盘价之差换算成 Point 单位的整数，再统一除以 100 后乘 100，等于把特征量按月百点粒度量化。b1 到 b4 分别以第 1、18、31、43 根 K 线收盘价为基准，对照「近 48 根中点下移 1200 点」的参考线，抓的是中长周期偏离结构。 CandlePatterns 函数把同类计算写进 xInputs 数组：a1 取当前收盘相对上偏移中点的差，a3、a4 直接算第 48 根与第 25、37 根的差，全部丢弃个位百点以下的噪音。回测中这种取整能让样本空间从几千维降到百点档位，聚类稳定性倾向提升。别把取整当无损压缩百点截断会吃掉小周期刺透形态，欧元美元这种平均日波 80~120 点的品种可能漏掉关键反转。开 MT5 把 /100*100 改成 /50*50，对比信号触发频次再决定粒度。

MQL5 / C++

  b1 = (class="type">int)(b1/class="num">100)*class="num">100;
  class="type">int b2 = (class="type">int)((((iClose(symbolS1.Name(),PERIOD_CURRENT,class="num">1)-(class="num">1200*Point())+iClose(symbolS1.Name(),PERIOD_CURRENT,class="num">48))/class="num">2)-iClose(symbolS1.Name(),PERIOD_CURRENT,class="num">18))/Point());
  b2 = (class="type">int)(b2/class="num">100)*class="num">100;
  class="type">int b3 = (class="type">int)((((iClose(symbolS1.Name(),PERIOD_CURRENT,class="num">1)-(class="num">1200*Point())+iClose(symbolS1.Name(),PERIOD_CURRENT,class="num">48))/class="num">2)-iClose(symbolS1.Name(),PERIOD_CURRENT,class="num">31))/Point());
  b3 = (class="type">int)(b3/class="num">100)*class="num">100;
  class="type">int b4 = (class="type">int)((((iClose(symbolS1.Name(),PERIOD_CURRENT,class="num">1)-(class="num">1200*Point())+iClose(symbolS1.Name(),PERIOD_CURRENT,class="num">48))/class="num">2)-iClose(symbolS1.Name(),PERIOD_CURRENT,class="num">43))/Point());
  b4 = (class="type">int)(b4/class="num">100)*class="num">100;
  class="kw">return (w1 * a1 + w2 * a2 + w3 * a3 + w4 * a4   +   v1 * b1 + v2 * b2 + v3 * b3 + v4 * b4);
  }
class=class="str">"cmt">//+------------------------------------------------------------------+
class=class="str">"cmt">//|percentage of each part of the candle respecting total size       |
class=class="str">"cmt">//+------------------------------------------------------------------+
class="type">int CandlePatterns(class="type">class="kw">double &xInputs[])
  {
  
  class="type">int a1 = (class="type">int)((((iClose(symbolS1.Name(),PERIOD_CURRENT,class="num">1)+(class="num">1200*Point())+iClose(symbolS1.Name(),PERIOD_CURRENT,class="num">48))/class="num">2)-iClose(symbolS1.Name(),PERIOD_CURRENT,class="num">1))/Point());
  xInputs[class="num">0] = (class="type">int)(a1/class="num">100)*class="num">100;
  class="type">int a2 = (class="type">int)((((iClose(symbolS1.Name(),PERIOD_CURRENT,class="num">1)-(class="num">1200*Point())+iClose(symbolS1.Name(),PERIOD_CURRENT,class="num">48))/class="num">2)-iClose(symbolS1.Name(),PERIOD_CURRENT,class="num">13))/Point());
  xInputs[class="num">1] = (class="type">int)(a2/class="num">100)*class="num">100;
  class="type">int a3 = (class="type">int)((iClose(symbolS1.Name(),PERIOD_CURRENT,class="num">48)-iClose(symbolS1.Name(),PERIOD_CURRENT,class="num">25))/Point());
  xInputs[class="num">2] = (class="type">int)(a3/class="num">100)*class="num">100;
  class="type">int a4 = (class="type">int)((iClose(symbolS1.Name(),PERIOD_CURRENT,class="num">48)-iClose(symbolS1.Name(),PERIOD_CURRENT,class="num">37))/Point());
  xInputs[class="num">3] = (class="type">int)(a4/class="num">100)*class="num">100;

用偏移中枢量化多周期收盘落差

这段逻辑把「近48根收盘均值上移1200点」当作参考中枢，再分别减去第6、18、31、43根收盘价，把价格落差折算成点数后做百点取整，塞进 xInputs[4]~[7]。具体看 g1：中枢 = (最新收盘 + 1200*Point + 48根前收盘) / 2，减去第6根收盘，除以 Point 得到整数点差，再 (g1/100)*100 把尾数抹掉。g2~g4 只是把加号换成减号、把减的棒线换成 18、31、43，其余完全同构。落差是带符号的：若近期收盘相对旧收盘偏高，xInputs 会落为负数，模型可能把它读成上方空间受压；反之正数倾向代表低位偏离。外汇与贵金属杠杆高，这套特征只是原始输入，实盘信号概率不等于确定性。直接把下面代码贴进 MT5 的自定义指标或EA里，把 symbolS1 换成你要测的品种，改 1200 那个偏移量就能看不同阈值下特征分布怎么变。

MQL5 / C++

  class="type">int g1 = (class="type">int)((((iClose(symbolS1.Name(),PERIOD_CURRENT,class="num">1)+(class="num">1200*Point())+iClose(symbolS1.Name(),PERIOD_CURRENT,class="num">48))/class="num">2)-iClose(symbolS1.Name(),PERIOD_CURRENT,class="num">6))/Point());
  xInputs[class="num">4] = (class="type">int)(g1/class="num">100)*class="num">100;
  class="type">int g2 = (class="type">int)((((iClose(symbolS1.Name(),PERIOD_CURRENT,class="num">1)-(class="num">1200*Point())+iClose(symbolS1.Name(),PERIOD_CURRENT,class="num">48))/class="num">2)-iClose(symbolS1.Name(),PERIOD_CURRENT,class="num">18))/Point());
  xInputs[class="num">5] = (class="type">int)(g2/class="num">100)*class="num">100;
  class="type">int g3 = (class="type">int)((((iClose(symbolS1.Name(),PERIOD_CURRENT,class="num">1)-(class="num">1200*Point())+iClose(symbolS1.Name(),PERIOD_CURRENT,class="num">48))/class="num">2)-iClose(symbolS1.Name(),PERIOD_CURRENT,class="num">31))/Point());
  xInputs[class="num">6] = (class="type">int)(g3/class="num">100)*class="num">100;
  class="type">int g4 = (class="type">int)((((iClose(symbolS1.Name(),PERIOD_CURRENT,class="num">1)-(class="num">1200*Point())+iClose(symbolS1.Name(),PERIOD_CURRENT,class="num">48))/class="num">2)-iClose(symbolS1.Name(),PERIOD_CURRENT,class="num">43))/Point());
  xInputs[class="num">7] = (class="type">int)(g4/class="num">100)*class="num">100;

  class="kw">return(class="num">1);
}

「神经网络EA的前向验证实况」

把 DeepNeuralNetwork.mqh 那套 4-4-3 和 8-4-3 逻辑直接塞进扇形模板做前向验证，取优化期「最大复杂准则」排名前 40 的结果跑样本外，能看清一件事：优化盈利因子和实盘前向表现经常脱节。 4-4-3 扇形 4 参数、24 根蜡烛、止盈 60 / 止损 600 时，优化盈利因子约 20，前向里靠大量小止盈扛住密集止损；但同样结构换 48 根蜡烛，优化因子掉到约 14，前向仍无正期望。8-4-3 扇形 8 参数、24 根蜡烛、止盈 60 / 止损 600 时优化因子飙到约 29，前向全程仅 1 次止损却仍显随机——高因子不等于样本外能赚。止盈 430 / 止损 200（二比一）在感知器 48 根蜡烛版里前六个月稳步增长、优化因子约 1.6，算少数站住的组合；但 8-4-3 同比例下优化因子 3 却前向亏损，倾向是神经网络对大止盈预测力不足。外汇与贵金属杠杆高，这类前向回撤随时可能放大，验证前请先算好自己的仓位风险。想复现就开 MT5 把现成模板代码替进文末 EA，自己把优化迭代拉到 20 次以上——原文那点验算次数对多参数网络明显不够，加跑迭代才可能碰到更稳的局部最优。

◍ 画得少，看得清

这套基于感知器和浅层神经网络的实验系统，伸缩性比预想中更灵活：模板和传参几乎可以无限叠加，但真正往前向验证时，少而精的入口反而更经得起推敲。把输入值做四舍五入处理后，正向结果数量大约翻了一倍，说明噪声压缩比堆参数更重要。算力是实打实的瓶颈。策略测试器只吃物理核心不吃线程，双 CPU 凑到 16 核及以上才像样，接 MQL5 云网络能明显把搜索效率拉起来。下一步值得试的是多感知器结构，以及把指标在一段区间内游走、背离这类现象喂进去。附件里从 DeepNeuralNetwork 原始库到 4-4-3 / 8-4-3 扇形 EA 的 opt 与 trade 版本都齐了，直接丢进 MT5 就能复跑。外汇和贵金属波动剧烈、杠杆风险高，任何信号都只是概率倾向，实盘前先用历史数据自己跑一遍前向测试。

把重复劳动交给小布

这些 EURUSD H1 的优化与前瞻诊断，小布盯盘的 AIGC 已内置，打开对应品种页即可看到多模板回测概览，你只需判断哪类形态更贴合当前行情。

←

神经网络实验（第 4 部分）：模板·进阶篇

常见问题

形态把价格收束到有限几何结构，降低了输入散射范围，神经网络更容易学到可复现的边界，而非拟合噪声。实验结果中简单感知器曾优于裸输入神经网，模板正是为补这块短板。

需按点差加偏移，文中给止盈加 30 点、平衡档位另有 200/430 点设置，且止盈设为止损 2 倍。外汇贵金属杠杆高，实盘前应在策略测试器重跑前向验证。

可以，品种页的 AIGC 模块覆盖了文中所提多模板优化概览与一年前向测试摘要，省去手动跑 20 次优化的等待。

复杂准则偏好结构稳定且有利的结果，对比盈利最大化更抗过拟合，文中强调用它才拿到较稳的前向表现。

3 年只是作者设定，并非可靠铁律；数据量与时长会显著影响神经网输出，可自调范围观察前向曲线衰减位置。

神经网络实验（第 4 部分）：模板·综合运用

把多周期收盘价差压成百点整数栅格

「用 1200 点偏移把 K 线距离整数化」

◍ 用感知机把多周期价差压成整百点距

把蜡烛拆成四个偏移量喂给感知机

「把K线距离折算成整百点距」

◍ 感知机里的偏移与价格区间离散化

把 K 线各段占比量化成整百点值

「感知机里的价格偏移量化」

◍ 用百点取整压缩蜡烛形态特征

用偏移中枢量化多周期收盘落差

「神经网络EA的前向验证实况」

◍ 画得少，看得清

常见问题

相关阅读