深度神经网络 (第七部分)。 神经网络的融合: 堆叠·综合运用
◍ 用R+Keras搭一个带噪声抑制的二元DNN分类器
在R环境里通过reticulate调用Keras，可以把MT5导出的特征矩阵直接喂给深度学习模型做二分类。下面这套脚本把训练集、测试集从env环境取出，标签用to_categorical做one-hot，类别数锁死为2，批次32，学习率1e-4，最多跑100个epoch。 模型主体是一条Sequential堆叠：输入先过一层高斯噪声(stddev=0.001)做正则，再连续交替BatchNorm、ELU全连接、Dropout。隐层节点数100→50→10，丢弃率前半段0.5、末段0.2，最后softmax出二类概率。优化器用RMSprop并带衰减，损失函数binary_crossentropy。 早停回调盯着val_acc，min_delta设1e-5、patience=20，避免白烧显卡。fit时validation_split=0.2，verbose=0关掉刷屏，但view_metrics=TRUE可在后台看曲线。 在测试集上跑出的混淆矩阵是：实际0类预测对184、错判46；实际1类错判78、预测对193。汇总指标Accuracy 0.752，Kappa 0.5068，No Information Rate仅0.523，P值<2.2e-16说明该模型显著优于瞎猜基线。外汇与贵金属行情具有高杠杆高风险，此类信号仅作概率参考，实盘须自担盈亏。 直接复制这段R代码到你的RStudio，把env$res替换成MT5导出的xts特征，就能复现这套75%出头的样本外准确率；若想压过NIR基线，重点调的是GN层stddev和dp1/dp2的丢弃率。
library(reticulate) library(keras) py_set_seed(class="num">12345) num_classes <- 2L batch_size <- 32L learning_rate <- class="num">0.0001 epochs <- 100L #--------- x_train <- env$res$InputTrain y_train <- env$Ytest %>% to_categorical() x_test <- env$res$InputTest y_test <- env$Ytest1 %>% to_categorical() x_test1 <- env$res$InputTest1 y_test1 <- env$Ytest2 %>% to_categorical() #---------------------------------------- early_stopping <- callback_early_stopping(monitor = "val_acc", min_delta = class="num">1e-5, patience = class="num">20, verbose = class="num">0, mode = "auto") ##----modelDNN--keras------------------------- # define model modDNN <- keras_model_sequential() # add layers and compile modDNN %>% layer_gaussian_noise(stddev = class="num">0.001, input_shape = dim(x_train)[class="num">2], name = "GN") %>% layer_batch_normalization() %>% layer_dense(units = class="num">100, activation = "elu", name = "dense1") %>% layer_dropout(rate = class="num">0.5, name = "dp1") %>% layer_batch_normalization() %>% layer_dense(units = class="num">50, activation = "elu", name = "dense2") %>% layer_batch_normalization() %>% layer_dropout(rate = class="num">0.5, name = "dp2") %>% layer_dense(units = class="num">10, activation = "elu", name = "dense3") %>% layer_batch_normalization() %>% layer_dropout(rate = class="num">0.2, name = "dp3") %>% layer_dense(units = num_classes, activation = "softmax", name = "soft") %>% compile( loss = &class="macro">#x27;binary_crossentropy&class="macro">#x27;, optimizer = optimizer_rmsprop(lr = learning_rate, decay = class="num">0.0001), metrics = &class="macro">#x27;accuracy&class="macro">#x27; ) ## Training & Evaluation --------------------------- # Fit model to data modDNN %>% fit( x_train, y_train, batch_size = batch_size, epochs = epochs, verbose = class="num">0, view_metrics = TRUE , shuffle = TRUE, validation_split = class="num">0.2, class="macro">#validation_data = list(x_test, y_test), callbacks = list(early_stopping)) -> history #--model--test------------------------- predict(modDNN, x_test) -> Ypr.test Ypr.test %>% apply(class="num">1, function(x) which.max(x)) %>% subtract(class="num">1) -> y_pr_test evalq(res_mod_test <- Eval(Ytest1, y_pr_test), env)
「分类混淆矩阵里的几个关键比率」
上面这组指标来自一个二分类模型的评估输出，直接反映了样本在正负类上的分布与判别质量。Prevalence 为 0.5230，说明标注为「Positive」类（此处代码里 Positive Class 定义为 0）的样本占比略过一半，类别本身不算极端失衡。 Neg Pred Value 读到 0.7122，意味着模型判为负类的结果里，约有 71% 确实为负；Detection Rate 0.3673 则表示真正命中正类的样本只占全体的 36.73%，而 Detection Prevalence 0.4591 是模型总共打出正类标签的比例。 Balanced Accuracy 0.7549 是两类召回率的均值，比单纯看准确率更抗类别倾斜。外汇与贵金属行情里用这类模型做状态识别，仍属高风险，信号只是概率倾向，不能直接当入场依据。 末尾的 plot(history) 是把训练过程历史画出来，开 MT5 接好数据后跑一遍，就能在终端看图核对收敛节奏。
plot(history)五个模型误差分解后谁更抗噪
把均值融合、多数表决、逻辑回归变体、DNN(7,2) 和 DNN500 这五个模型的测试品质摊开看，varb（500个融合输出合并）拿下了最高 Accuracy 0.7804，DNN(7,2) 紧随其后 0.7565。两者在测试样本上的最小误差 PredErr 分别为 0.2199 与 0.2460，且平方偏差 sqBias 比其余模型低一个数量级（varb 为 0.000398，DNN(7,2) 为 0.000195）。 所有模型的误差方差 PredVar 几乎都贴在 0.25 附近，融合没有如预期那样压低方差，反而换来更低的乖离，这点值得在 MT5 里复算时留意。predictionTargetCov 最佳的是 varb（0.13964），说明它与真实响应的联动最紧。 DNN500 各项分数全部垫底，证明在简单任务上堆复杂度反而拖累表现。实战里更划算的组合是：调好基础融合参数，再叠 varb + DNN(7,2) 做合并器。外汇与贵金属信号建模高风险，以上结论仅来自回测分解，实盘概率仍会漂移。 下面这段 R 侧脚本负责把 EnsAver 与 EnsVot 的分类误差拆成噪声、平方偏差、 estimator 方差和协方差，逻辑和其它模型一致，可直接照抄换 predictions 字段跑。
<span class="preprocessor">#---bias--test------------------------------- </span>import_fun(randomUniformForest, biasVarCov, BiasVar) evalq({ target = Ytest1 biasAver <- BiasVar(predictions = ScoreAver$clAver, target = target, regression = <span class="macro">FALSE</span>, idx = <span class="number">class="num">1</span>:length(target)) biasVot <- BiasVar(predictions = ScoreVot$ClVot, target = target, regression = <span class="macro">FALSE</span>, idx = <span class="number">class="num">1</span>:length(target)) }, env) ----------------------------- Noise: <span class="number">class="num">0.2488224</span> Squared bias: <span class="number">class="num">0.002107561</span> Variance of estimator: <span class="number">class="num">0.250475</span> Covariance of estimator and target: <span class="number">class="num">0.1257046</span> Assuming binary classification with classes {<span class="number">class="num">0</span>,<span class="number">class="num">1</span>}, where <span class="class="type">class="kw">string">&class="macro">#x27;class="num">0&class="macro">#x27;</span> is the majority <span class="keyword">class</span>. Misclassification rate = P(Y = <span class="number">class="num">1</span>)P(Y = <span class="number">class="num">0</span>) + {P(Y = <span class="number">class="num">1</span>) - P(Y_hat = <span class="number">class="num">1</span>)}^<span class="number">class="num">2</span> + P(Y_hat = <span class="number">class="num">0</span>)P(Y_hat = <span class="number">class="num">1</span>) - <span class="number">class="num">2</span>*Cov(Y, Y_hat) Misclassification rate = P(Y = <span class="number">class="num">1</span>) + P(Y_hat = <span class="number">class="num">1</span>) - <span class="number">class="num">2</span>*E(Y*Y_hat) = <span class="number">class="num">0.2499958</span> --------------------- Noise: <span class="number">class="num">0.2488224</span> Squared bias: <span class="number">class="num">0.004079665</span> Variance of estimator: <span class="number">class="num">0.2499721</span> Covariance of estimator and target: <span class="number">class="num">0.1274411</span> Assuming binary classification with classes {<span class="number">class="num">0</span>,<span class="number">class="num">1</span>}, where <span class="class="type">class="kw">string">&class="macro">#x27;class="num">0&class="macro">#x27;</span> is the majority <span class="keyword">class</span>. Misclassification rate = P(Y = <span class="number">class="num">1</span>)P(Y = <span class="number">class="num">0</span>) + {P(Y = <span class="number">class="num">1</span>) - P(Y_hat = <span class="number">class="num">1</span>)}^<span class="number">class="num">2</span> + P(Y_hat = <span class="number">class="num">0</span>)P(Y_hat = <span class="number">class="num">1</span>) - <span class="number">class="num">2</span>*Cov(Y, Y_hat) Misclassification rate = P(Y = <span class="number">class="num">1</span>) + P(Y_hat = <span class="number">class="num">1</span>) - <span class="number">class="num">2</span>*E(Y*Y_hat) = <span class="number">class="num">0.2479918</span> > env$biasAver $predError [<span class="number">class="num">1</span>] <span class="number">class="num">0.2499958</span> $squaredBias [<span class="number">class="num">1</span>] <span class="number">class="num">0.002107561</span> $predictionsVar [<span class="number">class="num">1</span>] <span class="number">class="num">0.250475</span> $predictionsTargetCov [<span class="number">class="num">1</span>] <span class="number">class="num">0.1257046</span>
◍ 融合之外还能怎么榨分类精度
ELM 神经网络分类器用平均或简单多数表决做融合，在计算速度极高时分类品质尚可；但把连续输出先调阈值转成标称变量、求平均前先做校正，品质还能再抬一点，误差方差没看到明显缩减。 把融合输出的平均值换成简单神经网络的 softmax，乖离能压低一个数量级，且方差没有明显恶化；反过来用更复杂的网络去替掉修剪和平均，结果并不好。 用贝叶斯变量选择（varbvs 包）跑出来的逻辑回归模型表现非常好，它挑出的最佳输出可以直接拿去喂神经网络。自预处理后就一直消不掉的 24% 噪声，暗示这些样本可能该在某个环节重标成独立一类。 接 keras 在时间序列上直接训，是后续提高分类品质的可行路径，外汇与贵金属行情高风险，模型信号仅作概率参考。
「随附的 R 与神经网络脚本清单」
这一节的实质是一份随附工程包的目录说明，作者把第八部分用到的 R 语言脚本和记录文件打进了 PartVII.zip（14.75 KB），在 MT5 社群文章页可直接下载。包内不是 MQL5 代码，而是配合 keras/TensorFlow 做深度神经网络融合的 R 侧工具链。 具体文件分工很明确：Importar.R 负责软件包导入，Library.R 是通用函数库，FunPrepareData_VII.R 做初始数据准备，FunStacking.R 创建并测试堆叠集合，Prepare.R 为可训练合并器备数据，Varb.R 跑 varb 基础模型，model_DNN7_2.R 与 model_DNN_500.R 分别是 DNN(7-2) 和 DNN500 网络脚本，SessionInfo_VII.txt 列了全部依赖包版本。 如果你要在本地复现文中的分类融合实验，先下这个 zip，用 R 3.x 环境装好对应包再顺序跑 Prepare.R → FunStacking.R 即可；外汇与贵金属相关的模型验证属高风险尝试，回测结论仅代表历史样本倾向，不等于实盘概率。
一点提醒
R 语言接 Keras 做交易建模，社区里更推荐用 RStudio 团队维护的接口，而不是早期第三方 kerasR——后者在 tensorflow 后端冲到 1.8 版时跟不上节奏，跑起来容易踩坑。 有开发者在 2019 年初提过把强化学习（DQN、DDPG、A3C、TRPO）接进交易训练，这类思路在 MT5+Python 的桥接项目里已有开源实现，比如 PyTrader 连接器，可直连实时行情。 外汇和贵金属杠杆高、滑点跳空频繁，神经网络信号只是概率倾向，实盘前务必用历史数据回测并小仓验证，别把实验室准确率当真金白银。