📘
连接 NeuroSolutions 神经网络·综合运用
把训练好的神经网络塞进EA里跑通
接好神经网络的EA，文件落位要先对齐：WeekPattern.dll 和 WeekPattern.nsw 丢进 MQL5\Files\NeuroSolutions\，通用适配层 NeuroSolutionsAdapter.dll 放 MQL5\Libraries\。路径错一处，EA加载就直接哑火。 验证连接是否成功，最实在的办法是开 MT5 策略测试器，用和训练神经网络完全相同的周期跑一遍。如果该周期正是模型所拟合的，盈利曲线倾向走出一段稳定向上——原文示例里那张图就是漂亮的单边上扬，说明 DLL 与权重文件握手正常。 但别被这条曲线骗了。某周期上训练出的净收益，只反映它对那段数据模式的识别力，不预示其他周期也能盈利；外汇与贵金属市场高风险，周期切换后主导形态可能完全变脸。 代码里 CNeuroSolutionsNeuralNet 类把神经网络相关逻辑全收进去了。构造函数拼出 terminal 路径下的 dll 与 nsw 绝对地址；Calc() 取最近6根K线（0号作基准 zlevel，1~5号拆成 open/high/low/close 共20个输入），再经 CalcNeuralNet 调 DLL 算输出。下面这段是类定义与输入填充的核心：
MQL5 / C++
<span class="keyword">class="kw">input</span> <span class="keyword">class="type">class="kw">double</span> Lots = <span class="number">class="num">0.1</span>; <span class="comment">class=class="str">"cmt">//+------------------------------------------------------------------+</span> <span class="comment">class=class="str">"cmt">// Connect the DLL adapter, class="kw">using which we are going to use the DLL neuronet created in NeuroSolutions</span> <span class="preprocessor">class="macro">#class="kw">import <span class="class="type">class="kw">string">"NeuroSolutionsAdapter.dll"</span></span> <span class="keyword">class="type">int</span> CalcNeuralNet(<span class="keyword">class="type">class="kw">string</span> dllPath, <span class="keyword">class="type">class="kw">string</span> weightsPath, <span class="keyword">class="type">class="kw">double</span>& inputs[], <span class="keyword">class="type">class="kw">double</span>& outputs[]); <span class="preprocessor">class="macro">#class="kw">import </span> <span class="comment">class=class="str">"cmt">//+------------------------------------------------------------------+</span> <span class="keyword">class</span> CNeuroSolutionsNeuralNet { <span class="keyword">class="kw">private</span>: <span class="keyword">class="type">class="kw">string</span> dllPath; <span class="comment">class=class="str">"cmt">// Path to a DLL neuronet created in NeuroSolutions</span> <span class="keyword">class="type">class="kw">string</span> weightsPath; <span class="comment">class=class="str">"cmt">// Path to a file of the neuronet balances</span> <span class="keyword">class="kw">public</span>: <span class="keyword">class="type">class="kw">double</span> in[<span class="number">class="num">20</span>]; <span class="comment">class=class="str">"cmt">// Neuronet inputs - OHLC of class="num">5 bars</span> <span class="keyword">class="type">class="kw">double</span> out[<span class="number">class="num">1</span>]; <span class="comment">class=class="str">"cmt">// Neuronet outputs - Close of a current bar</span> CNeuroSolutionsNeuralNet(); <span class="keyword">class="type">bool</span> Calc(); }; <span class="comment">class=class="str">"cmt">//+------------------------------------------------------------------+</span> <span class="keyword">class="type">void</span> CNeuroSolutionsNeuralNet::CNeuroSolutionsNeuralNet() { <span class="keyword">class="type">class="kw">string</span> terminal = <span class="functions">TerminalInfoString</span>(<span class="keyword">TERMINAL_PATH</span>); dllPath = terminal + <span class="class="type">class="kw">string">"\\MQL5\\Files\\NeuroSolutions\\WeekPattern.dll"</span>; weightsPath = terminal + <span class="class="type">class="kw">string">"\\MQL5\\Files\\NeuroSolutions\\WeekPattern.nsw"</span>; } <span class="comment">class=class="str">"cmt">//+------------------------------------------------------------------+</span> <span class="keyword">class="type">bool</span> CNeuroSolutionsNeuralNet::Calc() { <span class="comment">class=class="str">"cmt">// Get current quotes for the neuronet</span> <span class="predefines">class="type">MqlRates</span> rates[], rate; <span class="functions">CopyRates</span>(<span class="functions">Symbol</span>(), <span class="functions">Period</span>(), <span class="number">class="num">0</span>, <span class="number">class="num">6</span>, rates); <span class="functions">ArraySetAsSeries</span>(rates, true); <span class="comment">class=class="str">"cmt">// Fill the array of class="kw">input data of the neuronet</span> <span class="keyword">class="type">class="kw">double</span> zlevel=<span class="number">class="num">0</span>; <span class="keyword">for</span> (<span class="keyword">class="type">int</span> bar=<span class="number">class="num">0</span>; bar<=<span class="number">class="num">5</span>; bar++) { rate = rates[bar]; <span class="comment">class=class="str">"cmt">// class="num">0 bar is not taken for class="kw">input</span> <span class="keyword">if</span> (bar==<span class="number">class="num">0</span>) zlevel=rate.open; <span class="comment">class=class="str">"cmt">// level of price calculation</span> <span class="comment">class=class="str">"cmt">// class="num">1-class="num">5 bars are inputed</span> <span class="keyword">else</span> { <span class="keyword">class="type">int</span> i=(bar-<span class="number">class="num">1</span>)*<span class="number">class="num">4</span>; <span class="comment">class=class="str">"cmt">// class="kw">input number</span> in[i ] = rate.open -zlevel; in[i+<span class="number">class="num">1</span>] = rate.high -zlevel; in[i+<span class="number">class="num">2</span>] = rate.low -zlevel; in[i+<span class="number">class="num">3</span>] = rate.close-zlevel; } } <span class="comment">class=class="str">"cmt">// Calculate the neuronet in the NeuroSolutions DLL(though the DLL adapter)</span> <span class="keyword">class="type">int</span> res = CalcNeuralNet(dllPath, weightsPath, in, out); <span class="keyword">class="kw">switch</span> (res) { <span class="keyword">case</span> <span class="number">class="num">1</span>: <span class="functions">Print</span>(<span class="class="type">class="kw">string">"Error of creating neuronet from DLL \""</span>, dllPath, <span class="class="type">class="kw">string">"\""</span>); <span class="keyword">class="kw">return</span> (class="kw">false);
◍ 把神经网络输出接进开平仓逻辑
上面那段神经网络封装返回 true 时，预测值已经躺在 out[0] 里，调用方不用再做任何加工。OnTick 里先跑 NN.Calc()，成功就把 out[0] 赋给 Prognoze，失败则清零——这个 0 既是错误标记也是「不交易」信号。 Trade() 的核心只有两条规则：持仓方向与预测相反就平，没仓且预测非零就按符号开。具体看，BUY 仓遇到 Prognoze<=0 平，SELL 仓遇到 Prognoze>=0 平；空仓时 Prognoze>0 买、<0 卖。 外汇和贵金属杠杆高，神经网络预测失准会连续触发反向平仓，实盘前务必在 MT5 策略测试器用历史数据跑一遍，确认 Calc() 的失败分支不会把 Prognoze 卡在旧值上。 Lots 变量在贴出的片段里未定义，编译前要在文件头补一个 double Lots=0.01; 之类的赋值，否则 Buy/Sell 调用会报未声明标识符。
MQL5 / C++
case class="num">2: Print("Error of loading balances to neuronet from the file "", weightsPath, """); class="kw">return (class="kw">false); case class="num">3: Print("Error of calculation of neuronet"); class="kw">return (class="kw">false); } class=class="str">"cmt">// Output of the neuronet has appeared in the array out, you shouldn&class="macro">#x27;t do anything with it class="kw">return (true); } class=class="str">"cmt">//+------------------------------------------------------------------+ CNeuroSolutionsNeuralNet NN; class="type">class="kw">double Prognoze; class=class="str">"cmt">//+------------------------------------------------------------------+ class="macro">#include <Trade\Trade.mqh> class=class="str">"cmt">//+------------------------------------------------------------------+ class="type">void OnTick() { class=class="str">"cmt">// Get the price prediction from the neuronet if (NN.Calc()) Prognoze = NN.out[class="num">0]; else Prognoze = class="num">0; class=class="str">"cmt">// Perform necessary trade actions Trade(); } class=class="str">"cmt">//+------------------------------------------------------------------+ class="type">void Trade() { class=class="str">"cmt">// Close an open position if it is opposite to the prediction if(PositionSelect(_Symbol)) { class="type">long type=PositionGetInteger(POSITION_TYPE); class="type">bool close=class="kw">false; if((type == POSITION_TYPE_BUY) && (Prognoze <= class="num">0)) close = true; if((type == POSITION_TYPE_SELL) && (Prognoze >= class="num">0)) close = true; if(close) { CTrade trade; trade.PositionClose(_Symbol); } } class=class="str">"cmt">// If there is no positions, open one according to the prediction if((Prognoze!=class="num">0) && (!PositionSelect(_Symbol))) { CTrade trade; if(Prognoze > class="num">0) trade.Buy(Lots); if(Prognoze < class="num">0) trade.Sell(Lots); } }
「画得少，看得清」
把神经网络接进 MT5 做自动分析，本质仍是对历史样本拟合：weekpattern.mq5 仅 3.78 KB，weekpattern-export.mq5 2.04 KB，配套 dll_nsw.zip 383.37 KB，文件体积本身就说明轻量，但轻量不等于对未来有决定作用。 评论区里有交易者直言，参数优化只是历史盈利最大化的变形，靠这套直接盈利大概率会落空；也有用户卡在 VC6 生成 DLL 的环境上，证明落地门槛不在算法而在工具链。 真要验证，下载那两个 mq5 在策略测试器跑 EURUSD 周线，看样本外曲线是否塌方，比读任何结论都实在。外汇与贵金属波动剧烈，高杠杆下这类实验只该用模拟盘。
常见问题
把模型导出的权重和拓扑结构写成EA里的计算函数，用历史样本归一化后逐根K线调用前向传播，输出值接进交易判断。
设一个阈值区间，输出大于上轨只做多、小于下轨只做空，中间观望；同时用波动率过滤掉极端行情的误触。
小布可加载你的模型输出做实时信号标注，自动把多空概率画在图上，你只看关键区就行。
市场分布漂移导致输入特征失效，建议每季度用近期数据重训，并只保留鲁棒性高的少数特征。
只留一条净方向线和阈值带，其余中间量收进缓冲区，做到画得少、看得清。