数据科学和机器学习（第 12 部分）：自训练神经网络能否帮助您跑赢股市？·综合运用
把特征压进 [0,1] 的最小-最大缩放
最小-最大缩放把任意数值特征线性映射到闭区间 [0,1]，公式是 x_norm = (x - x_min) / (x_max - x_min)。x_min、x_max 取该特征在样本里的最小和最大值，映射后极值必定落在 0 和 1，不会破坏原始分布的单调性。 在 MT5 的回归神经网络类里，规范化方法由 norm_technique 枚举控制，传 NORM_MIN_MAX_SCALER 即启用本方法。下面这段构造调用里，学习率 0.01、训练 1000 轮、激活用 ReLU、损失用 MSE，并显式指定了最小-最大缩放。 [CODE] CRegNeuralNets::CRegNeuralNets(matrix &xmatrix, vector &yvector,double alpha, uint epochs, activation ACTIVATION_FUNCTION, loss LOSS_FUNCTION, norm_technique NORM_METHOD) enum norm_technique { NORM_MIN_MAX_SCALER, //Min max scaler NORM_MEAN_NORM, //Mean normalization NORM_STANDARDIZATION, //standardization NORM_NONE //Do not normalize. }; nn = new CRegNeuralNets(x_matrix,y_vector,0.01,1000, AF_RELU_, LOSS_MSE_,NORM_MIN_MAX_SCALER); CS 0 22:40:56.457 Self Trained NN EA (NAS100,M30) [ 1/1000 ] Loss = 0.19379434 | accuracy -0.581 CS 0 22:40:56.457 Self Trained NN EA (NAS100,M30) [ 2/1000 ] Loss = 0.07735744 | accuracy 0.369 CS 0 22:40:56.458 Self Trained NN EA (NAS100,M30) [ 3/1000 ] Loss = 0.04761891 | accuracy 0.611 CS 0 22:40:56.458 Self Trained NN EA (NAS100,M30) [ 4/1000 ] Loss = 0.03559318 | accuracy 0.710 CS 0 22:40:56.458 Self Trained NN EA (NAS100,M30) [ 5/1000 ] Loss = 0.02937830 | accuracy 0.760 CS 0 22:40:56.458 Self Trained NN EA (NAS100,M30) [ 6/1000 ] Loss = 0.02582918 | accuracy 0.789 CS 0 22:40:56.459 Self Trained NN EA (NAS100,M30) [ 7/1000 ] Loss = 0.02372224 | accuracy 0.806 CS 0 22:40:56.459 Self Trained NN EA (NAS100,M30) [ 8/1000 ] Loss = 0.02245222 | accuracy 0.817 CS 0 22:40:56.460 Self Trained NN EA (NAS100,M30) [ 9/1000 ] Loss = 0.02168207 | accuracy 0.823 CS 0 22:40:56.460 Self Trained NN EA (NAS100,M30 [/CODE] 上面日志是 NAS100 的 M30 图表上跑出来的前 9 轮：损失从 0.1938 降到 0.0217，准确率从 -0.581 爬到 0.823。最终训练集准确率 82.3%，在 M30 上倾向给出可用输出；但日线测出 -77%、H4 甚至返回 -11234 这类离谱值，说明单层增量规则网络对不同周期形态捕捉能力不一致。 别把正态当圣经 最小-最大缩放不是万能药。它在 M30 上表现稳，换日线或 H4 就崩，外汇和贵金属这类高波动品种尤其容易放大样本极值偏差，实盘前务必自己切周期复测。
CRegNeuralNets::CRegNeuralNets(matrix &xmatrix, vector &yvector,class="type">class="kw">double alpha, class="type">uint epochs, activation ACTIVATION_FUNCTION, loss LOSS_FUNCTION, norm_technique NORM_METHOD) enum norm_technique { NORM_MIN_MAX_SCALER, class=class="str">"cmt">//Min max scaler NORM_MEAN_NORM, class=class="str">"cmt">//Mean normalization NORM_STANDARDIZATION, class=class="str">"cmt">//standardization NORM_NONE class=class="str">"cmt">//Do not normalize. }; nn = new CRegNeuralNets(x_matrix,y_vector,class="num">0.01,class="num">1000, AF_RELU_, LOSS_MSE_,NORM_MIN_MAX_SCALER); CS class="num">0 class="num">22:class="num">40:class="num">56.457 Self Trained NN EA(NAS100,M30) [ class="num">1/class="num">1000 ] Loss = class="num">0.19379434 | accuracy -class="num">0.581 CS class="num">0 class="num">22:class="num">40:class="num">56.457 Self Trained NN EA(NAS100,M30) [ class="num">2/class="num">1000 ] Loss = class="num">0.07735744 | accuracy class="num">0.369 CS class="num">0 class="num">22:class="num">40:class="num">56.458 Self Trained NN EA(NAS100,M30) [ class="num">3/class="num">1000 ] Loss = class="num">0.04761891 | accuracy class="num">0.611 CS class="num">0 class="num">22:class="num">40:class="num">56.458 Self Trained NN EA(NAS100,M30) [ class="num">4/class="num">1000 ] Loss = class="num">0.03559318 | accuracy class="num">0.710 CS class="num">0 class="num">22:class="num">40:class="num">56.458 Self Trained NN EA(NAS100,M30) [ class="num">5/class="num">1000 ] Loss = class="num">0.02937830 | accuracy class="num">0.760 CS class="num">0 class="num">22:class="num">40:class="num">56.458 Self Trained NN EA(NAS100,M30) [ class="num">6/class="num">1000 ] Loss = class="num">0.02582918 | accuracy class="num">0.789 CS class="num">0 class="num">22:class="num">40:class="num">56.459 Self Trained NN EA(NAS100,M30) [ class="num">7/class="num">1000 ] Loss = class="num">0.02372224 | accuracy class="num">0.806 CS class="num">0 class="num">22:class="num">40:class="num">56.459 Self Trained NN EA(NAS100,M30) [ class="num">8/class="num">1000 ] Loss = class="num">0.02245222 | accuracy class="num">0.817 CS class="num">0 class="num">22:class="num">40:class="num">56.460 Self Trained NN EA(NAS100,M30) [ class="num">9/class="num">1000 ] Loss = class="num">0.02168207 | accuracy class="num">0.823 CS class="num">0 class="num">22:class="num">40:class="num">56.460 Self Trained NN EA(NAS100,M30
◍ 自训练神经网络只在首 tick 跑一次
一段真实 EA 日志能说明问题：NAS100 的 M30 图上，网络用 1000/1000 样本训完，Loss 落到 0.02046533，训练集准确率 0.833。这种自训练 NN 模型在外汇与贵金属上属于高风险尝试，样本内表现好不等于样本外能持续。 控制训练频数是关键。OnTick 里用 train_nn 布尔锁，第一次进来才调用 TrainNetwork()，随后立刻置 true，后续 tick 不再重训，避免每笔报价都烧 CPU 重算梯度。 TrainNetwork 的拼装方式值得照抄：x_matrix 取两个品种的同周期收盘价各填一列，y_vector 取当前品种收盘价作标签，学习率 0.01、迭代 1000 次、ReLU 激活、MSE 损失、最小最大归一化。开 MT5 把 symbol_x2 换成和你主品种有领先关系的标的，就能复现这套输入结构。
class="type">void OnTick() { class=class="str">"cmt">//--- if (!train_nn) TrainNetwork(); class=class="str">"cmt">//Train the network only once train_nn = true; } class=class="str">"cmt">//+------------------------------------------------------------------+ class=class="str">"cmt">//| | class=class="str">"cmt">//+------------------------------------------------------------------+ class="type">void TrainNetwork() { matrix x_matrix(n_samples,class="num">2); vector y_vector; vector x_vector; x_vector.CopyRates(symbol_x,PERIOD_CURRENT,copy_rates_x,class="num">0,n_samples); x_matrix.Col(x_vector, class="num">0); x_vector.CopyRates(symbol_x2, PERIOD_CURRENT,copy_rates_x,class="num">0,n_samples); x_matrix.Col(x_vector, class="num">1); y_vector.CopyRates(Symbol(), PERIOD_CURRENT,COPY_RATES_CLOSE,class="num">0,n_samples); nn = new CRegNeuralNets(x_matrix,y_vector,class="num">0.01,class="num">1000, AF_RELU_, LOSS_MSE_,NORM_MIN_MAX_SCALER); }
「随机打散才不会被K线顺序带偏」
做过 Python 机器学习的人对 sklearn 的 train_test_split 不陌生。拆数据不光是分训练集和测试集，核心是把样本顺序随机化，否则模型会把「数据按时间先后排列」这种结构噪声当成规律学进去。 神经网络抓的是变量之间的形态，不是样本的组织方式。若直接拿原始顺序的行情矩阵去拟合，模型可能只是在拟合排列惯性，换一段行情就失效。外汇与贵金属波动受宏观与流动性冲击，顺序相关伪形态尤其多，这类过拟合风险很高。 MQL5 里收集完特征后，用 TrainTestSplitMatrices 在随机状态 42 下做切分，70% 训练、30% 测试。随机状态固定成 42 是为了可复现，你改这个数，训练集样本就换一批，回测结果也可能跟着飘。 下面这段是实际调用方式，注意 Matrix 前三列分别是两个符号的某速率特征和当前品种收盘价，切分后丢进 CRegNeuralNets 做回归，最后用 r_squared 打测试集分数。
class="type">void TrainNetwork() { class=class="str">"cmt">//--- collecting the data matrix Matrix(n_samples,class="num">3); vector y_vector; vector x_vector; x_vector.CopyRates(symbol_x,PERIOD_CURRENT,copy_rates_x,class="num">0,n_samples); Matrix.Col(x_vector, class="num">0); x_vector.CopyRates(symbol_x2, PERIOD_CURRENT,copy_rates_x,class="num">0,n_samples); Matrix.Col(x_vector, class="num">1); y_vector.CopyRates(Symbol(), PERIOD_CURRENT,COPY_RATES_CLOSE,class="num">0,n_samples); Matrix.Col(y_vector, class="num">2); class=class="str">"cmt">//--- matrix x_train, x_test; vector y_train, y_test; matrix_utils.TrainTestSplitMatrices(Matrix, x_train, y_train, x_test, y_test, class="num">0.7, class="num">42); nn = new CRegNeuralNets(x_train,y_train,class="num">0.01,class="num">1000, AF_RELU_, LOSS_MSE_,NORM_MIN_MAX_SCALER); vector test_pred = nn.ForwardPass(x_test); printf("Testing Accuracy =%.3f",metrics.r_squared(y_test, test_pred)); } class="type">void TrainTestSplitMatrices(matrix &matrix_,matrix &x_train,vector &y_train,matrix &x_test, vector &y_test,class="type">class="kw">double train_size=class="num">0.7,class="type">int random_state=-class="num">1)
在 OnTick 里喂数据给神经网络
实时预测的关键，是把每个 tick 到来的最新行情塞进已经训练好的网络输入向量，而不是等一根柱子走完才动手。上面这段逻辑直接在 OnTick 内完成采集与前向推算，省掉了任何中间缓存。 代码里先用 CopyRates 只取当前蜡烛的 x1、x2 两个标的数值（示例为 Apple 与 Tesla），拼成长度为 2 的 inputs 向量，再丢给 nn.ForwardPass 得到纳斯达克预测值。注意 train_nn 标志位保证 TrainNetwork 只跑一次，后续 tick 不再重训，避免重复消耗。 你在 MT5 里验证时，把 symbol_x / symbol_x2 换成自己盯的跨市场品种，copy_rates_x 对应特征索引，Comment 输出的 pred 就是该 tick 的网络输出。外汇与贵金属叠加此类跨资产模型，杠杆与滑点会放大误差，信号仅作概率参考，实盘前务必用历史 tick 回测。
class="type">void OnTick() { class=class="str">"cmt">//--- if (!train_nn) TrainNetwork(); class=class="str">"cmt">//Train the network only once train_nn = true; vector x1, x2; x1.CopyRates(symbol_x,PERIOD_CURRENT,copy_rates_x,class="num">0,class="num">1); class=class="str">"cmt">//only the current candle x2.CopyRates(symbol_x2,PERIOD_CURRENT,copy_rates_x,class="num">0,class="num">1); class=class="str">"cmt">//only the current candle vector inputs = {x1[class="num">0], x2[class="num">0]}; class=class="str">"cmt">//current values of x1 and x2 instruments | Apple & Tesla matrix OUT = nn.ForwardPass(inputs); class=class="str">"cmt">//Predicted Nasdaq value class="type">class="kw">double pred = OUT[class="num">0][class="num">0]; Comment("pred ",OUT); }
◍ 神经网络预测价驱动的开平仓逻辑
这套自训练 EA 的核心判断只有一条：模型输出的预测价 pred 高于实时卖价 ticks.bid，就倾向做多；低于实时买价 ticks.ask，就倾向做空。止盈与止损不是固定点数，而是用预测价与当前价的差 target_gap 乘以外部输入系数再换算成 Point 小数位。 开仓前有一道硬门槛：MathAbs(pred - ticks.ask) 加上点差 spread 必须大于经纪商规定的止损最小距离 stops_level，否则整段逻辑直接跳过。这一步能避免订单因距离不足被拒，外汇与贵金属品种在高波动时 stops_level 可能瞬间拉大，需留意。 多单的止损挂在 ticks.bid 减去 target_gap*stop_loss 个 Point，止盈挂在 ticks.bid 加上 target_gap*take_profit 个 Point；空单则对称反向。PosExist 检查确保同方向不重复加仓。 下面这段代码就是 MetaEditor 里的实际写法，可直接复制到 EA 的 OnTick 里验证：先取 stops_level、最小手数、点差和最新 tick，再走上面的判断分支。
stops_level = (class="type">int)SymbolInfoInteger(Symbol(),SYMBOL_TRADE_STOPS_LEVEL); Lots = SymbolInfoDouble(Symbol(),SYMBOL_VOLUME_MIN); spread = (class="type">class="kw">double)SymbolInfoInteger(Symbol(), SYMBOL_SPREAD); class="type">MqlTick ticks; SymbolInfoTick(Symbol(), ticks); if (MathAbs(pred - ticks.ask) + spread > stops_level) { if (pred > ticks.ask && !PosExist(POSITION_TYPE_BUY)) { target_gap = pred - ticks.bid; m_trade.Buy(Lots, Symbol(), ticks.ask, ticks.bid - ((target_gap*stop_loss) * Point()) , ticks.bid + ((target_gap*take_profit) * Point()),"Self Train NN | Buy"); } if (pred < ticks.bid && !PosExist(POSITION_TYPE_SELL)) { target_gap = ticks.ask - pred; m_trade.Sell(Lots, Symbol(), ticks.bid, ticks.ask + ((target_gap*stop_loss) * Point()), ticks.ask - ((target_gap*take_profit) * Point()), "Self Train NN | Sell"); } }
「回测里的训练与测试准确率陷阱」
在 MT5 策略测试器里跑机器学习 EA，最容易被忽略的是训练集与测试集的质量差异。我曾用 NAS100 的 4 小时图，基于真实跳价回测 2023.01.01 至 2023.02.23，把模型训练函数挂在测试生命周期第一根跳价上，训练和推断几乎瞬时完成。 第一轮结果很反直觉：训练准确率仅 37.9%，测试准确率却冲到 71.7%。这种倒挂通常指向训练数据有洞——每一段缺失的行情都可能让模型学偏。该 EA 最终 78.27% 的平仓交易以亏损收场，在没动止损止盈前，这概率说明模型本身扛不住实盘噪声。 随后做了一轮短优化，挑出这组参数：copy_rates_x 用 COPY_RATES_LOW、n_samples=2950、滑点=1、止损=7.4、止盈=5.0。重跑后训练准确率 61.5%、测试准确率 63.5%，两者贴近才像能用的模型。外汇与贵金属杠杆高，回测顺滑不代表实盘能复制，开 MT5 用真实跳价复一遍再信。 下面这段测试器日志是初始烂参数的现场：copy_rates_x=1、n_samples=200、滑点=100、止损/止盈都=2.0，还出现了单日 1379 根分钟棒里缺 2 分钟真实跳价。这种数据洞正是准确率倒挂的温床。
CS class="num">0 class="num">15:class="num">50:class="num">47.676 Tester NAS100,H4(Pepperstone-Demo): generating based on real ticks CS class="num">0 class="num">15:class="num">50:class="num">47.677 Tester NAS100,H4: testing of Experts\Advisors\Self Trained NN EA.ex5 from class="num">2023.01.class="num">01 class="num">00:class="num">00 to class="num">2023.02.class="num">23 class="num">00:class="num">00 started with inputs: CS class="num">0 class="num">15:class="num">50:class="num">47.677 Tester symbol_x=Apple_Inc_(AAPL.O) CS class="num">0 class="num">15:class="num">50:class="num">47.677 Tester symbol_x2=Tesco_(TSCO.L) CS class="num">0 class="num">15:class="num">50:class="num">47.677 Tester copy_rates_x=class="num">1 CS class="num">0 class="num">15:class="num">50:class="num">47.677 Tester n_samples=class="num">200 CS class="num">0 class="num">15:class="num">50:class="num">47.677 Tester = CS class="num">0 class="num">15:class="num">50:class="num">47.677 Tester slippage=class="num">100 CS class="num">0 class="num">15:class="num">50:class="num">47.677 Tester stop_loss=class="num">2.0 CS class="num">0 class="num">15:class="num">50:class="num">47.677 Tester take_profit=class="num">2.0 CS class="num">3 class="num">15:class="num">50:class="num">49.209 Ticks NAS100 : class="num">2023.02.class="num">21 class="num">23:class="num">59 - real ticks absent for class="num">2 minutes out of class="num">1379 total minute bars within a day CS class="num">0 class="num">15:class="num">50:class="num">51.466 History Tesco_(TSCO.L),H4: history begins from class="num">2022.01.class="num">04 class="num">08:class="num">00 CS class="num">0 class="num">15:class="num">50:class="num">51.467 Self Trained NN EA(NAS100,H4) class="num">2023.01.class="num">03 class="num">01:class="num">00:class="num">00 [ class="num">1/class="num">1000 ] Loss = class="num">0.14025037 | accuracy -class="num">1.524
NAS100 自训练神经网络的收敛轨迹
在 MT5 策略测试器里跑自训练神经网络 EA（标的 NAS100、周期 H4、回测起点 2023.01.03 01:00），前 10 个 epoch 的日志已经能看出训练动态。第 2 轮 Loss 为 0.05244676、准确率仅 0.056，到第 10 轮 Loss 降到 0.03498872、准确率升到 0.370。 整轮 1000 次迭代跑完，最终 Loss 停在 0.03452066，准确率 0.379。也就是说，即便迭代满额，模型在样本内也只能拿到约 38% 的判别命中，离随机基准不算遥遥领先。 外汇与贵金属这类高波动品种接同类自训练网络时，样本内准确率偏低往往预示样本外泛化会更弱，实盘前建议先用小资金或模拟账户验证。打开 MT5 的 Experts 日志面板，把上述 epoch 行和你的训练输出对照，就能判断自己的网络是否也卡在 0.38 附近的瓶颈。
◍ 自训练神经网络在NAS100四小时图上的回测表现
把自训练的神经网络EA跑在NAS100的H4周期上，训练集与测试集的命中率会出现明显落差。一次典型日志里，初始测试准确率为0.717，说明模型在未见数据上先给了个偏乐观的印象。 随后完整跑完1000/1000批次，训练过程记录 Loss = 0.05890808，训练集准确率降到0.615；紧接着测试准确率报出0.635。训练集和测试集相差约2个百分点，过拟合不算严重，但绝对水平刚过六成，信号本身偏弱。 这类结果提醒做外汇或贵金属量化的人：NAS100虽非外汇直盘，但指数CFD同样高波动、高杠杆，用六成出头的模型直接跟单风险很大。建议先在MT5策略测试器用相同品种周期复现日志，再决定是否加过滤条件。
CS class="num">0 class="num">15:class="num">50:class="num">51.662 Self Trained NN EA(NAS100,H4) class="num">2023.01.class="num">03 class="num">01:class="num">00:class="num">00 Testing Accuracy =class="num">0.717 CS class="num">0 class="num">17:class="num">11:class="num">52.100 Self Trained NN EA(NAS100,H4) class="num">2023.01.class="num">03 class="num">01:class="num">00:class="num">00 [ class="num">1000/class="num">1000 ] Loss = class="num">0.05890808 | accuracy class="num">0.615 CS class="num">0 class="num">17:class="num">11:class="num">52.101 Self Trained NN EA(NAS100,H4) class="num">2023.01.class="num">03 class="num">01:class="num">00:class="num">00 Testing Accuracy =class="num">0.635
「别急着下结论」
增量规则只服务于单层回归型神经网络，这一点要刻在脑子里。它本质就是若干线性回归模型的组合，比如你可以把它当成 5 个各管一摊的线性回归模型凑在一起解同一个问题。 这类单层结构理解不了变量里的复杂形态，跑不出预期结果时不必意外。它真正价值在于充当通用反向传播算法的积木，后者才撑得起深度学习的复杂网络。 作者故意让自己被训练误差带着走，是想说明：哪怕网络真能学出形态，也得抠住缩放、初始化这些小细节，很多环节做对了它才肯工作。 MALE5 函数库在 GitHub 持续迭代，目前含 metrics.mqh（准确性测量）、preprocessing.mqh（数据缩放）、matrix_utils.mqh（矩阵操控）、selftrain NN.mqh（自训练核心）及测试用 EA；外汇与贵金属杠杆高，实盘前务必在 MT5 策略测试器跑通再说。