支援向量機(SVM)是什麼意思？

正好最近自己學習機器學習，看到reddit上 Please explain Support Vector Machines (SVM) like I am a 5 year old 的帖子，一個字贊！於是整理一下和大家分享。(如有錯歡迎指教！)

什麼是SVM?

支援向量機/support vector machine (SVM)。當然首先看一下wiki.

Support Vector Machines are learning models used for classification: which individuals in a population belong where? So… how do SVM and the mysterious “kernel” work?
 

好吧，故事是這樣子的：

在很久以前的情人節，大俠要去救他的愛人，但魔鬼和他玩了一個遊戲。魔鬼在桌子上似乎有規律放了兩種顏色的球，說：“你用一根棍分開它們？要求：儘量在放更多球之後，仍然適用。”
… 文章 詳細內容 地址：https://www.botvs.com/bbs-topic/680

• 我們來用Python 實現一個 SVM 分類器 預測 買賣

程式是 基於發明者量化平臺的，標的物選擇為電子貨幣，因為電子貨幣適合回測。Python機器學習之SVM 預測買賣，Python入門簡單策略 sklearn 機器學習庫的使用，
回測系統自帶的庫有：

numpy pandas TA-Lib scipy statsmodels sklearn cvxopt hmmlearn pykalman arch matplotlib
 

實盤需要在託管者所在機器安裝策略需要的庫，策略原始碼地址： https://www.botvs.com/strategy/21370

from sklearn import svm
import numpy as np

def main():
    preTime = 0
    n = 0
    success = 0
    predict = None
    pTime = None
    marketPosition = 0
    initAccount = exchange.GetAccount()
    Log("Running...")
    while True:
        r = exchange.GetRecords()
        if len(r) < 60:
            continue
        bar = r[len(r)-1]
        if bar.Time > preTime:
            preTime = bar.Time
            if pTime is not None and r[len(r)-2].Time == pTime:
                diff = r[len(r)-2].Close - r[len(r)-3].Close
                if diff > SpreadVal:
                    success += 1 if predict == 0 else 0
                elif diff < -SpreadVal:
                    success += 1 if predict == 1 else 0
                else:
                    success += 1 if predict == 2 else 0
                pTime = None
                LogStatus("預測次數", n, "成功次數", success, "準確率:", '%.3f %%' % round(float(success) * 100 / n, 2))
        else:
            Sleep(1000)
            continue
        inputs_X, output_Y = [], []
        sets = [None, None, None]
        for i in xrange(1, len(r)-2, 1):
            inputs_X.append([r[i].Open, r[i].Close])
            Y = 0
            diff = r[i+1].Close - r[i].Close
            if diff > SpreadVal:
                Y = 0
                sets[0] = True
            elif diff < -SpreadVal:
                Y = 1
                sets[1] = True
            else:
                Y = 2
                sets[2] = True
            output_Y.append(Y)
        if None in sets:
            Log("樣本不足, 無法預測 ...")
            continue
        n += 1
        clf = svm.LinearSVC()
        clf.fit(inputs_X, output_Y)
        predict = clf.predict(np.array([bar.Open, bar.Close]).reshape((1, -1)))
        pTime = bar.Time
        
        Log("預測當前Bar結束:", bar.Time, ['漲', '跌', '橫'][predict])
        if marketPosition == 0:
            if predict == 0:
                exchange.Buy(initAccount.Balance/2)
                marketPosition = 1
            elif predict == 1:
                exchange.Sell(initAccount.Stocks/2)
                marketPosition = -1
        else:
            nowAccount = exchange.GetAccount()
            if marketPosition > 0 and predict != 0:
                exchange.Sell(nowAccount.Stocks - initAccount.Stocks)
                nowAccount = exchange.GetAccount()
                marketPosition = 0
            elif marketPosition < 0 and predict != 1:
                while True:
                    dif = initAccount.Stocks - nowAccount.Stocks
                    if dif < 0.01:
                        break
                    ticker = exchange.GetTicker()
                    exchange.Buy(ticker.Sell + (ticker.Sell-ticker.Buy)*2, dif)
                    while True:
                        Sleep(1000)
                        orders = exchange.GetOrders()
                        for order in orders:
                            exchange.CancelOrder(order.Id)
                        if len(orders) == 0:
                            break
                    nowAccount = exchange.GetAccount()
                marketPosition = 0
            if marketPosition == 0:
                LogProfit(_N(nowAccount.Balance - initAccount.Balance, 4), nowAccount)
 

閱讀原文