📊

借助成交量精准洞悉交易动态：超越传统OHLC图表·进阶篇

Q: 成交量一阶导数和二阶导数在MT5里具体怎么算？

一阶导数是相邻周期成交量的差分，近似变化率；二阶导数是一阶导的差分，代表加速度。EA中通过指标缓冲区逐根计算并绘制在独立窗口。

Q: LSTM预测的成交量就一定能预示价格拐点吗？

不一定。模型输出是概率性前瞻，在算法市中参考价值较高，但仍需结合价格结构和背离验证，外汇贵金属杠杆高、滑点突发行情下可能失效。

Q: 小布盯盘能直接加载这套成交量衍生指标吗？

可以，小布盯盘品种页已集成分离窗口的三缓冲显示，含导数与LSTM预测，不用自己写EA也能看同类信号。

Q: 为什么传统OHLC技术指标在算法驱动市场不够用？

算法单常在价格显形前就改变成交量结构，纯OHLC滞后于参与度的微妙变化，微积分+ML能提前暴露参与度位移。

（2/3）·靠OHLC看盘总慢半拍？把成交量拆成一阶、二阶导数再喂给LSTM，拐点可能更早浮现

实战向第 2/3 篇

不少交易者把成交量只当副图里的柱线扫一眼，忽略了它变化率和加速度里藏着的早期信号。等价格破位才反应，往往已经错过最优入场区。用微积分视角重新拆成交量，配合轻量神经网络，盘口语言会清楚很多。

本章目录

持仓台账与当日盈亏的底层维护
从持仓止损到 OnTick 调度的实现切片
把ADX与LSTM权重塞进MT5结构体
LSTM 推下一根量的归一与反归一细节

持仓台账与当日盈亏的底层维护

这段逻辑解决一个实盘常见痛点：EA 重启后怎么知道当前挂着的单子状态、以及今天到底赚了多少。核心是用一个 openTrades 结构数组在内存里记账，开仓时写 openTime 和 barsOpen=0，平仓则调 RemoveTrade 把对应槽位清零。 RemoveTrade 按 ticket 线性扫描数组，命中后把 ticket、openTime、barsOpen 全部置 0 并 break，避免无谓循环。注意这里没有压缩数组，只是逻辑删除，长时间运行数组尾部会有空洞，但 ArraySize 不变。 GetDailyProfit 是当日盈亏的真实来源。它先用 TimeToString+TIME_DATE 拿到当天 0 点，跨天后重置 dailyProfit 并 HistorySelect 拉取今日起的历史成交，只累加 _Symbol 的 DEAL_PROFIT；再遍历 PositionsTotal 把同品种浮动盈亏加进去。 ApplyTrailingStop 的入口已经暴露了关键字段：openPrice 取 POSITION_PRICE_OPEN，currentPrice 取 POSITION_PRICE_CURRENT，后续追踪止损的距离计算就基于这两者的差。外汇与贵金属杠杆高，追踪止损参数设错可能瞬间扫掉盈利单，建议在 MT5 策略测试器用点差 20 点的 XAUUSD 先验证。

MQL5 / C++

openTrades[i].openTime = TimeCurrent();
openTrades[i].barsOpen = class="num">0;
break;
}
}
}
class="type">void RemoveTrade(class="type">ulong ticket) {
   for(class="type">int i = class="num">0; i < ArraySize(openTrades); i++) {
      if(openTrades[i].ticket == ticket) {
         openTrades[i].ticket = class="num">0;
         openTrades[i].openTime = class="num">0;
         openTrades[i].barsOpen = class="num">0;
         break;
      }
   }
}
class="type">class="kw">double GetDailyProfit() {
   class="type">class="kw">datetime today = StringToTime(TimeToString(TimeCurrent(), TIME_DATE));
   if(lastDayChecked != today) {
      lastDayChecked = today;
      dailyProfit = class="num">0;
      HistorySelect(today, TimeCurrent());
      class="type">int deals = HistoryDealsTotal();
      for(class="type">int i = class="num">0; i < deals; i++) {
         class="type">ulong ticket = HistoryDealGetTicket(i);
         if(ticket > class="num">0) {
            if(HistoryDealGetString(ticket, DEAL_SYMBOL) == _Symbol)
               dailyProfit += HistoryDealGetDouble(ticket, DEAL_PROFIT);
         }
      }
   }
   
   class=class="str">"cmt">// Add floating P/L
   class="type">class="kw">double floatingProfit = class="num">0;
   for(class="type">int i = PositionsTotal() - class="num">1; i >= class="num">0; i--) {
      class="type">ulong ticket = PositionGetTicket(i);
      if(ticket > class="num">0 && PositionSelectByTicket(ticket)) {
         if(PositionGetString(POSITION_SYMBOL) == _Symbol)
            floatingProfit += PositionGetDouble(POSITION_PROFIT);
      }
   }
   class="kw">return dailyProfit + floatingProfit;
}
class="type">void ApplyTrailingStop(class="type">class="kw">double point) {
   for(class="type">int i = PositionsTotal() - class="num">1; i >= class="num">0; i--) {
      class="type">ulong ticket = PositionGetTicket(i);
      if(ticket <= class="num">0) class="kw">continue;
      if(PositionSelectByTicket(ticket)) {
         class="type">class="kw">double openPrice = PositionGetDouble(POSITION_PRICE_OPEN);
         class="type">class="kw">double currentPrice = PositionGetDouble(POSITION_PRICE_CURRENT);

「从持仓止损到 OnTick 调度的实现切片」

下面这段 MT5 代码把移动止损和每根 K 线一次的调度逻辑拼在了一起。先在持仓循环里取出现有 SL 和持仓方向，仅当浮盈点数越过 TrailingStart 才重算止损位，避免无谓的修改请求。 double currentSL = PositionGetDouble(POSITION_SL); long posType = PositionGetInteger(POSITION_TYPE); // Calculate and update trailing stop double profitPoints; if(posType == POSITION_TYPE_BUY) { profitPoints = (currentPrice - openPrice) / point; if(profitPoints >= TrailingStart) { double newSL = openPrice + (profitPoints - TrailingStep) * point;

if(currentSL < newSL	currentSL == 0) {

trade.PositionModify(ticket, newSL, PositionGetDouble(POSITION_TP)); } } } // Similar logic for SELL positions 逐行拆解：第1行用 PositionGetDouble(POSITION_SL) 读出当前止损价；第2行取持仓类型（买或卖）。profitPoints 是浮盈换算成点的数值。买入仓里，若 profitPoints 达到 TrailingStart（例如设 30 点），newSL 会在开仓价基础上保留 TrailingStep（例如 15 点）的回撤空间。只有当新止损严于原止损或原止损为 0 时才调用 PositionModify，卖仓逻辑对称。 OnTick 里用静态变量 lastBar 记住上一根 K 线时间，iTime(_Symbol, PERIOD_CURRENT, 0) 取到当前柱时间，相等就直接 return，保证一天内每根柱只跑一次主体。若当日累计利润低于 MaxDailyLoss 阈值则平仓并退出，这是硬风控。 void OnTick() { static datetime lastBar = 0; datetime currentBar = iTime(_Symbol, PERIOD_CURRENT, 0); if(lastBar == currentBar) return; lastBar = currentBar; if(GetDailyProfit() < MaxDailyLoss) { CloseAllPositions(); return; } // Update and calculate UpdateTradesInfo(); CheckTimeBasedClose(); if(!CanTrade()) return; // Volume analysis and LSTM prediction // ... volume calculations ... if(UseLSTM && volumePredictor != NULL) { // LSTM prediction logic // ... prediction calculations ... } CalculateDerivatives(); if(consecutiveAccel >= AccelBars) {

if((!UseADX && !UseOBV)	ConfirmLongSignal()) {

// Open buy position } } else if(consecutiveAccel <= -AccelBars) {

if(!UseADX && !UseOBV	ConfirmShortSignal()) {

// Open sell position } } } 开仓判定依赖 consecutiveAccel 与 AccelBars 的比较：连续加速大于等于阈值且过滤条件通过才考虑买，反向则考虑卖。外汇与贵金属杠杆高，这类 EA 在滑点扩大时可能漏单或追错价，上 MT5 策略测试器用 2023 年 XAUUSD 的 M15 数据跑一遍更能看清实际表现。 ADX 模块开头就判 UseADX 开关，关掉直接 return 不占资源。三个缓冲区 adx_buffer、plusdi_buffer、minusdi_buffer 都用 ArraySetAsSeries(..., true) 设成时间序列，最新值落在索引 0，方便后面直接读当前柱。 void CalculateADX() { if(!UseADX) return; double adx_buffer[]; double plusdi_buffer[]; double minusdi_buffer[]; ArraySetAsSeries(adx_buffer, true); ArraySetAsSeries(plusdi_buffer, true); ArraySetAsSeries(minusdi_buffer, true);

MQL5 / C++

class="type">class="kw">double currentSL = PositionGetDouble(POSITION_SL);
class="type">long posType = PositionGetInteger(POSITION_TYPE);
class=class="str">"cmt">// Calculate and update trailing stop
class="type">class="kw">double profitPoints;
if(posType == POSITION_TYPE_BUY) {
   profitPoints = (currentPrice - openPrice) / point;
   if(profitPoints >= TrailingStart) {
      class="type">class="kw">double newSL = openPrice + (profitPoints - TrailingStep) * point;
      if(currentSL < newSL || currentSL == class="num">0) {
         trade.PositionModify(ticket, newSL, PositionGetDouble(POSITION_TP));
      }
   }
}
class=class="str">"cmt">// Similar logic for SELL positions
}
}
}
class="type">void OnTick() {
   class="kw">static class="type">class="kw">datetime lastBar = class="num">0;
   class="type">class="kw">datetime currentBar = iTime(_Symbol, PERIOD_CURRENT, class="num">0);
   if(lastBar == currentBar) class="kw">return;
   lastBar = currentBar;
   if(GetDailyProfit() < MaxDailyLoss) {
      CloseAllPositions();
      class="kw">return;
   }
   class=class="str">"cmt">// Update and calculate
   UpdateTradesInfo();
   CheckTimeBasedClose();
   if(!CanTrade()) class="kw">return;
   class=class="str">"cmt">// Volume analysis and LSTM prediction
   class=class="str">"cmt">// ... volume calculations ...
   if(UseLSTM && volumePredictor != NULL) {
      class=class="str">"cmt">// LSTM prediction logic
      class=class="str">"cmt">// ... prediction calculations ...
   }
   CalculateDerivatives();
   if(consecutiveAccel >= AccelBars) {
      if((!UseADX && !UseOBV) || ConfirmLongSignal()) {
         class=class="str">"cmt">// Open buy position
      }
   }
   else if(consecutiveAccel <= -AccelBars) {
      if(!UseADX && !UseOBV || ConfirmShortSignal()) {
         class=class="str">"cmt">// Open sell position
      }
   }
}
class="type">void CalculateADX() {
   if(!UseADX) class="kw">return;
   class="type">class="kw">double adx_buffer[];
   class="type">class="kw">double plusdi_buffer[];
   class="type">class="kw">double minusdi_buffer[];
   ArraySetAsSeries(adx_buffer, true);
   ArraySetAsSeries(plusdi_buffer, true);
   ArraySetAsSeries(minusdi_buffer, true);

◍ 把ADX与LSTM权重塞进MT5结构体

在 MT5 里跑 AI 辅助信号，第一步是把指标缓冲和模型参数都装进自定义结构体。下面这段把 ADX 的三条线（主线、+DI、-DI）从指标句柄拷进本地数组，再转存到 adx_main / adx_plus / adx_minus，方便后续做矩阵运算。 CopyBuffer(adx_handle, MAIN_LINE, 0, ArraySize(adx_main), adx_buffer); CopyBuffer(adx_handle, PLUSDI_LINE, 0, ArraySize(adx_plus), plusdi_buffer); CopyBuffer(adx_handle, MINUSDI_LINE, 0, ArraySize(adx_minus), minusdi_buffer); ArrayCopy(adx_main, adx_buffer); ArrayCopy(adx_plus, plusdi_buffer); ArrayCopy(adx_minus, minusdi_buffer); LSTMCell 结构体把遗忘门、输入门、输出门和细胞状态各自的权重矩阵（Matrix2D）与偏置数组分开声明，hidden_size 与 input_size 决定矩阵维度。初始化时用 MathSqrt(2.0/(input_size+hidden_size)) 做 He 缩放，避免深层网络梯度消失——这是 ReLU 类激活常用的经验系数。 double scale = MathSqrt(2.0 / (lstm.input_size + lstm.hidden_size)); lstm.Wf.Init(lstm.hidden_size, lstm.input_size); for(int i=0; i<lstm.hidden_size; i++){ for(int j=0; j<lstm.input_size; j++){ lstm.Wf.Set(i, j, (MathRand()/32768.0 - 0.5) * scale); } } 成交量是价格行为里常被忽视的维度。NormalizeVolumes 先求 historical_volumes 的均值，再准备平方和，典型 Z-score 标准化前半段。外汇与贵金属杠杆高，这类特征若不做归一，LSTM 门控更新可能偏向极端样本，实盘信号失真概率会上升。开 MT5 把上面结构体贴进 EA，先跑通 ADX 拷贝再接权重初始化，能省掉后续大半调试时间。

MQL5 / C++

CopyBuffer(adx_handle, MAIN_LINE, class="num">0, ArraySize(adx_main), adx_buffer);
CopyBuffer(adx_handle, PLUSDI_LINE, class="num">0, ArraySize(adx_plus), plusdi_buffer);
CopyBuffer(adx_handle, MINUSDI_LINE, class="num">0, ArraySize(adx_minus), minusdi_buffer);
ArrayCopy(adx_main, adx_buffer);
ArrayCopy(adx_plus, plusdi_buffer);
ArrayCopy(adx_minus, minusdi_buffer);
}
class="kw">struct Matrix2D {
  class="type">class="kw">double values[];
  class="type">int rows;
  class="type">int cols;
  
  class="type">void Init(class="type">int r, class="type">int c) {
    rows = r;
    cols = c;
    ArrayResize(values, rows * cols);
    ArrayInitialize(values, class="num">0);
  }
  class=class="str">"cmt">// ... other methods
}
class="kw">struct LSTMCell {
  class="type">class="kw">double forget_gate[];
  class="type">class="kw">double input_gate[];
  class="type">class="kw">double cell_state[];
  class="type">class="kw">double output_gate[];
  class="type">class="kw">double hidden_state[];
  
  Matrix2D Wf;  class=class="str">"cmt">// Forget gate weights
  Matrix2D Wi;  class=class="str">"cmt">// Input gate weights
  Matrix2D Wc;  class=class="str">"cmt">// Cell state weights
  Matrix2D Wo;  class=class="str">"cmt">// Output gate weights
  
  class="type">class="kw">double bf[];  class=class="str">"cmt">// Forget gate bias
  class="type">class="kw">double bi[];  class=class="str">"cmt">// Input gate bias
  class="type">class="kw">double bc[];  class=class="str">"cmt">// Cell state bias
  class="type">class="kw">double bo[];  class=class="str">"cmt">// Output gate bias
  
  class="type">int hidden_size;
  class="type">int input_size;
};
class="type">class="kw">double Sigmoid(class="type">class="kw">double x) {
  class="kw">return class="num">1.0 / (class="num">1.0 + MathExp(-x));
}
class="type">class="kw">double Tanh(class="type">class="kw">double x) {
  class="kw">return (MathExp(x) - MathExp(-x)) / (MathExp(x) + MathExp(-x));
}
class="type">void InitializeWeights() {
  class="type">class="kw">double scale = MathSqrt(class="num">2.0 / (lstm.input_size + lstm.hidden_size));
  
  class=class="str">"cmt">// Initialize weight matrices
  lstm.Wf.Init(lstm.hidden_size, lstm.input_size);
  lstm.Wi.Init(lstm.hidden_size, lstm.input_size);
  class=class="str">"cmt">// ... other initializations
  class=class="str">"cmt">// Random initialization with scaling
  for(class="type">int i = class="num">0; i < lstm.hidden_size; i++) {
    for(class="type">int j = class="num">0; j < lstm.input_size; j++) {
      lstm.Wf.Set(i, j, (MathRand() / class="num">32768.0 - class="num">0.5) * scale);
      class=class="str">"cmt">// ... other weight initializations
    }
  }
}
class="type">void NormalizeVolumes() {
  volume_mean = class="num">0;
  class="type">class="kw">double sum_squared = class="num">0;
  
  class=class="str">"cmt">// Calculate mean
  for(class="type">int i = class="num">0; i < ArraySize(historical_volumes); i++) {
    volume_mean += historical_volumes[i];
  }

LSTM 推下一根量的归一与反归一细节

这段收尾代码把前面算好的均值和标准差直接用上了。先求历史成交量的标准差：遍历累加每个样本与均值差的平方，再开方；若标准差为 0 则强制置 1，纯粹是为了避免后面归一化时除零崩掉。归一化后，historical_volumes 里存的就不再是原始手数，而是 z-score。PredictNextVolume 拿最新一根的 z-score 当输入，跑一遍 LSTM 前向：三个门（遗忘、输入、输出）加候选细胞态都用 Sigmoid / Tanh 算，hidden_state 由输出门乘细胞态双曲正切得到。最终预测值是所有隐藏单元输出的平均值，再乘回 volume_std 加回 volume_mean 反归一。也就是说，模型在标准化空间里预测，吐出来后再还原成真实量纲——你在 MT5 里改 volume_std 初始值或隐藏层大小，预测的绝对量级会明显漂移。外汇和贵金属成交量受session与平台撮合影响大，这类预测在高波动时段误差可能放大，仅作仓位参考而非方向依据。

MQL5 / C++

   volume_mean /= ArraySize(historical_volumes);
   
   class=class="str">"cmt">// Calculate standard deviation
   for(class="type">int i = class="num">0; i < ArraySize(historical_volumes); i++) {
      sum_squared += MathPow(historical_volumes[i] - volume_mean, class="num">2);
   }
   volume_std = MathSqrt(sum_squared / ArraySize(historical_volumes));
   
   class=class="str">"cmt">// Normalize volumes
   if(volume_std == class="num">0) volume_std = class="num">1;   class=class="str">"cmt">// Prevent division by zero
   for(class="type">int i = class="num">0; i < ArraySize(historical_volumes); i++) {
      historical_volumes[i] = (historical_volumes[i] - volume_mean) / volume_std;
   }
}
class="type">class="kw">double PredictNextVolume() {
   if(!is_initialized) class="kw">return class="num">0;
   
   class="type">class="kw">double input1 = historical_volumes[ArraySize(historical_volumes)-class="num">1];
   
   class=class="str">"cmt">// LSTM forward pass
   for(class="type">int h = class="num">0; h < lstm.hidden_size; h++) {
      class="type">class="kw">double ft = class="num">0, it = class="num">0, ct = class="num">0, ot = class="num">0;
      
      class=class="str">"cmt">// Calculate gates
      for(class="type">int i = class="num">0; i < lstm.input_size; i++) {
         ft += lstm.Wf.Get(h, i) * input1;
         it += lstm.Wi.Get(h, i) * input1;
         ct += lstm.Wc.Get(h, i) * input1;
         ot += lstm.Wo.Get(h, i) * input1;
      }
      
      class=class="str">"cmt">// Apply gates and calculate states
      lstm.forget_gate[h] = Sigmoid(ft + lstm.bf[h]);
      lstm.input_gate[h] = Sigmoid(it + lstm.bi[h]);
      class="type">class="kw">double c_tilde = Tanh(ct + lstm.bc[h]);
      lstm.cell_state[h] = lstm.forget_gate[h] * lstm.cell_state[h] +
                           lstm.input_gate[h] * c_tilde;
      lstm.output_gate[h] = Sigmoid(ot + lstm.bo[h]);
      lstm.hidden_state[h] = lstm.output_gate[h] * Tanh(lstm.cell_state[h]);
   }
   
   class=class="str">"cmt">// Calculate final prediction
   class="type">class="kw">double prediction = class="num">0;
   for(class="type">int h = class="num">0; h < lstm.hidden_size; h++)
      prediction += lstm.hidden_state[h];
   prediction /= lstm.hidden_size;
   
   class="kw">return prediction * volume_std + volume_mean;   class=class="str">"cmt">// Denormalize
}

把导数计算交给小布盯盘

这些成交量一阶、二阶导数的实时计算和LSTM预测可视化，小布盯盘的AIGC已内置，打开对应品种页即可看到分离窗口里的三个缓冲区，你只管读信号做决策。

←

借助成交量精准洞悉交易动态：超越传统OHLC图表（基础篇）

→

借助成交量精准洞悉交易动态：超越传统OHLC图表·综合运用

常见问题

一阶导数是相邻周期成交量的差分，近似变化率；二阶导数是一阶导的差分，代表加速度。EA中通过指标缓冲区逐根计算并绘制在独立窗口。

不一定。模型输出是概率性前瞻，在算法市中参考价值较高，但仍需结合价格结构和背离验证，外汇贵金属杠杆高、滑点突发行情下可能失效。

可以，小布盯盘品种页已集成分离窗口的三缓冲显示，含导数与LSTM预测，不用自己写EA也能看同类信号。

算法单常在价格显形前就改变成交量结构，纯OHLC滞后于参与度的微妙变化，微积分+ML能提前暴露参与度位移。

借助成交量精准洞悉交易动态：超越传统OHLC图表·进阶篇

持仓台账与当日盈亏的底层维护

「从持仓止损到 OnTick 调度的实现切片」

◍ 把ADX与LSTM权重塞进MT5结构体

LSTM 推下一根量的归一与反归一细节

常见问题

相关阅读