获取市场优势的秘诀（第二部分）：预测技术指标·进阶篇
「预测均线变化比猜收盘价靠谱」
直接拿模型去猜收盘价，结果惨不忍睹：逻辑回归 0.5220、LDA 0.5192、XGB 0.5120、简单神经网络 0.5235、大型神经网络 0.5187，全挤在 50% 出头，基本等于抛硬币。
换目标去预测移动平均的变化，准确率立刻抬到 69% 附近：逻辑回归 0.6927、LDA 0.6964、XGB 0.6933、简单神经网络 0.6948、大型神经网络 0.6965。大型神经网络数字最高，但性能图里有两个远低于均值的离群点，生产环境不敢用。
结论很直接——在现有数据上，理想模型应当比逻辑回归复杂、比大型神经网络简单。我们挑了小型神经网络（隐藏层 20×10 做训练对比，实盘用 5×2）落地信号，因为它最稳。
下面这段是把对比实验和 MT5 交易骨架拼起来的原型代码，注意 VOLUME 初始为 0，必须在 symbol 循环里被 volume_min 覆盖，否则发单必失败。外汇与贵金属波动剧烈，迷你模型信号仅作辅助，实盘前请在模拟盘验证稳定性。
for i in enumerate(np.arange(class="num">0,error_close_df.shape[class="num">1])): print(error_close_df.columns[i[class="num">0]]," ", error_close_df.iloc[:,i[class="num">0]].max()) class="macro">#Training each model to predict changes in a technical indicator(in this example simple moving average) instead of close price. for i,(train,test) in enumerate(tscv.split(csv)): model= MLPClassifier(solver=&class="macro">#x27;lbfgs&class="macro">#x27;,alpha=class="num">1e-5,hidden_layer_sizes=(class="num">20, class="num">10), random_state=class="num">1) model.fit(csv_reduced.loc[train[class="num">0]:train[-class="num">1],:],csv.loc[train[class="num">0]:train[-class="num">1],"Target MA"]) error_ma_df.iloc[i,class="num">4] = accuracy_score(csv.loc[test[class="num">0]:test[-class="num">1],"Target MA"],model.predict(csv_reduced.loc[test[class="num">0]:test[-class="num">1],:])) for i in enumerate(np.arrange(class="num">0,error_ma_df.shape[class="num">1])): print(error_ma_df.columns[i[class="num">0]]," ", error_ma_df.iloc[:,i[class="num">0]].max()) class="macro">#Import the libraries we need class="kw">import MetaTrader5 as mt5 class="kw">import pandas_ta as ta class="kw">import pandas as pd class="macro">#Trading global variables MARKET_SYMBOL = &class="macro">#x27;Volatility class="num">75 Index&class="macro">#x27; class="macro">#This data frame will store the most recent price update last_close = pd.DataFrame() class="macro">#We may not always enter at the price we want, how much deviation can we tolerate? DEVIATION = class="num">10000 class="macro">#We will always enter at the minimum volume VOLUME = class="num">0 class="macro">#How many times the minimum volume should our positions be LOT_MUTLIPLE = class="num">1 class="macro">#What timeframe are we working on? TIMEFRAME = mt5.TIMEFRAME_M1 class="macro">#Which model have we decided to work with? neural_network_model= MLPClassifier(solver=&class="macro">#x27;lbfgs&class="macro">#x27;,alpha=class="num">1e-5,hidden_layer_sizes=(class="num">5, class="num">2), random_state=class="num">1) class="macro">#Determine the minimum volume for index,symbol in enumerate(symbols): if symbol.name == MARKET_SYMBOL: print(f"{symbol.name} has minimum volume: {symbol.volume_min}") VOLUME = symbol.volume_min * LOT_MULTIPLE # function to send a market order def market_order(symbol, volume, order_type, **kwargs): class="macro">#Fetching the current bid and ask prices tick = mt5.symbol_info_tick(symbol) class="macro">#Creating a dictionary to keep track of order direction order_dict = {&class="macro">#x27;buy&class="macro">#x27;: class="num">0, &class="macro">#x27;sell&class="macro">#x27;: class="num">1} price_dict = {&class="macro">#x27;buy&class="macro">#x27;: tick.ask, &class="macro">#x27;sell&class="macro">#x27;: tick.bid} request = { "action": mt5.TRADE_ACTION_DEAL, "symbol": symbol, "volume": volume, "type": order_dict[order_type], "price": price_dict[order_type], "deviation": DEVIATION, "magic": class="num">100, "comment": "Indicator Forecast Market Order", "type_time": mt5.ORDER_TIME_GTC, "type_filling": mt5.ORDER_FILLING_FOK, } order_result = mt5.order_send(request) print(order_result) class="kw">return order_result # Closing our order based on ticket id def close_order(ticket): positions = mt5.positions_get() for pos in positions:
用神经网络信号反手平掉持仓
平仓单的核心是把原持仓方向反过来吃对手价。代码里 type_dict 把 0 映射成 1、1 映射成 0，配合 price_dict 取 tick.bid 或 tick.ask，等于买仓用 ask 价卖、卖仓用 bid 价买，滑点控制在 DEVIATION 常量内。 订单结构体里 type_filling 用了 ORDER_FILLING_FOK，意味着整笔必须在当时报价全成，否则不成交；外汇与贵金属杠杆高，FOK 在流动性薄的时段可能直接废单，实盘前建议在 MT5 策略测试器跑一遍观察 reject 率。 信号端用 60 周期 SMA 做特征，target 定义为下一根 close 高于当前 close 则标 1，模型 predict 出 1 解读为 buy、0 为 sell。回测区间写死从 2024-01-01 到 datetime.now()，换品种时这段日期和 TIMEFRAME 都要手动改，否则拿到的样本量会偏短。 主循环是 while True 死循环轮询 ai_signal()，每轮 print 出方向。真要挂机得加 try/except 和 mt5 连接心跳，不然 MT5 掉线后脚本会空转打印报错。
tick = mt5.symbol_info_tick(pos.symbol) class="macro">#validating that the order is for this symbol type_dict = {class="num">0: class="num">1, class="num">1: class="num">0} # class="num">0 represents buy, class="num">1 represents sell - inverting order_type to close the position price_dict = {class="num">0: tick.ask, class="num">1: tick.bid} class="macro">#bid ask prices if pos.ticket == ticket: request = { "action": mt5.TRADE_ACTION_DEAL, "position": pos.ticket, "symbol": pos.symbol, "volume": pos.volume, "type": type_dict[pos.type], "price": price_dict[pos.type], "deviation": DEVIATION, "magic": class="num">10000, "comment": "Indicator Forecast Market Order", "type_time": mt5.ORDER_TIME_GTC, "type_filling": mt5.ORDER_FILLING_FOK, } order_result = mt5.order_send(request) print(order_result) class="kw">return order_result class="kw">return &class="macro">#x27;Ticket does not exist&class="macro">#x27; class="macro">#Update our date from and date to date_from = class="type">class="kw">datetime(class="num">2024,class="num">1,class="num">1) date_to = class="type">class="kw">datetime.now() class="macro">#Let&class="macro">#x27;s create a function to preprocess our data def preprocess(df): class="macro">#Calculating class="num">60 period Simple Moving Average df.ta.sma(length=class="num">60,append=True) class="macro">#Drop any rows that have missing values df.dropna(axis=class="num">0,inplace=True) class="macro">#Get signals from our model def ai_signal(): class="macro">#Fetch OHLC data df = pd.DataFrame(mt5.copy_rates_range(market_symbol,TIMEFRAME,date_from,date_to)) class="macro">#Process the data df[&class="macro">#x27;time&class="macro">#x27;] = pd.to_datetime(df[&class="macro">#x27;time&class="macro">#x27;],unit=&class="macro">#x27;s&class="macro">#x27;) df[&class="macro">#x27;target&class="macro">#x27;] = (df[&class="macro">#x27;close&class="macro">#x27;].shift(-class="num">1) > df[&class="macro">#x27;close&class="macro">#x27;]).astype(class="type">int) preprocess(df) class="macro">#Select the last row last_close = df.iloc[-class="num">1:,class="num">1:] class="macro">#Remove the target column last_close.pop(&class="macro">#x27;target&class="macro">#x27;) class="macro">#Use the last row to generate a forecast from our moving average forecast model class="macro">#Remember class="num">1 means buy and class="num">0 means sell forecast = neural_network_model.predict(last_close) class="kw">return forecast[class="num">0] class="macro">#if __name__ == &class="macro">#x27;__main__&class="macro">#x27;: class="macro">#We&class="macro">#x27;ll use an infinite loop to keep the program running class="kw">while True: class="macro">#Fetching model prediction signal = ai_signal() class="macro">#Decoding model prediction into an action if signal == class="num">1: direction = &class="macro">#x27;buy&class="macro">#x27; elif signal == class="num">0: direction = &class="macro">#x27;sell&class="macro">#x27; print(f&class="macro">#x27;AI Forecast: {direction}&class="macro">#x27;) class="macro">#Opening A Buy Trade
◍ 同品种反手前的持仓清理逻辑
在 MT5 上跑单边信号时，最容易被忽略的是同符号反向单的残留。上面这段控制流先判断方向，若信号为 buy，就遍历 positions_get() 返回的持仓，把 type==1（即卖出持仓）逐个用 close_order(ticket) 平掉，避免新旧方向相反互相打架。 清理完反向仓后，代码用 mt5.positions_totoal()（原文拼写如此，实应 positions_total）判断当前是否零持仓，确认空仓才调用 market_order 以 VOLUME 手数进场。卖信号分支完全对称：扫 type==0 的买仓平仓，再检查 positions_get() 为空后开卖。 循环末尾 sleep(60) 代表每 60 秒轮询一次信号与持仓，print 出 datetime.now() 方便你直接在终端看每次触发的时间戳。外汇与贵金属杠杆高，这种强制平反手单的机制能降低多空叠加的敞口风险，但是否执行仍取决于你的风控参数。 开 MT5 把这段贴进 Python 环境跑之前，先核对 positions_totoal 这个拼写错误——原样复制会直接抛异常，改回 positions_total 才能验证清理逻辑是否如预期。
if direction == &class="macro">#x27;buy&class="macro">#x27;: class="macro">#Close any sell positions for pos in mt5.positions_get(): if pos.type == class="num">1: class="macro">#This is an open sell order, and we need to close it close_order(pos.ticket) if not mt5.positions_totoal(): class="macro">#We have no open positions market_order(MARKET_SYMBOL,VOLUME,direction) class="macro">#Opening A Sell Trade elif direction == &class="macro">#x27;sell&class="macro">#x27;: class="macro">#Close any buy positions for pos in mt5.positions_get(): if pos.type == class="num">0: class="macro">#This is an open buy order, and we need to close it close_order(pos.ticket) if not mt5.positions_get(): class="macro">#We have no open positions market_order(MARKET_SYMBOL,VOLUME,direction) print(&class="macro">#x27;time: &class="macro">#x27;, class="type">class="kw">datetime.now()) print(&class="macro">#x27;-------\n&class="macro">#x27;) time.sleep(class="num">60)
「把技术指标当目标才绕开伪回归坑」
做预测时，技术指标的变化往往比收盘价更好猜，原因很硬：MFM 这类指标的完整公式我们全知道，输入只涉及收盘价、最高价、最低价，预测它等于解一道已知方程的题。图12 里说的很清楚——技术指标有精确数学描述，收盘价却没有。 反观收盘价，没有任何公式列明哪些变量驱动它。用一串价格加技术指标去反推价格，本质是假设指标能影响价格，这逻辑从根上就站不住，指标永远只是价格的派生量。 机器学习里这种「目标底层函数未知」的活法，容易掉进伪回归。作者举的例子很直白：山顶隔着雾喊「那边有狗」，走近发现是灌木后真有狗。输入（模糊形状）和输出（狗）其实独立，只是凑巧答对，模型什么真实关系都没学到，验证集还可能给个骗人的低错误率。 外汇和贵金属这种高波动品种，伪回归一旦带进 EA，实盘亏穿比回测漂亮更常见。既然没有现成公式框定价格输入，与其硬磕收盘价，不如拿有已知输入关系的指标当学习目标，至少能少造一批「看起来会、实则瞎蒙」的模型。
指标可预测，但滞后始终是硬伤
抛开直接预测收盘价那条路，转向预测技术指标变化在实测中确实更站得住脚。不过别迷信高准确率——移动平均线预测得再准，它本身和价格就存在滞后，可能出现均线向下而价格向上的背离，这在外汇和贵金属这种高杠杆市场里容易误判进出场。 下一步值得在 MT5 里继续挖：除了 MA，震荡类指标是否具备更高预测精度，社区里已经有人用标准 MQL5 功能跑通了 Tick 成交量预测作为例证。 这类方法终究只是概率优势，不是确定性信号；真要落地，建议先把自己写的指标缓冲区预测逻辑在策略测试器里跑一遍，看残差分布再决定参数。