資源の読み込みに... 荷物...

ニューラルネットワークスーパートレンド戦略

作者: リン・ハーンチャオチャン, 日時: 2023-09-14 16:49:38
タグ:

戦略の論理

この戦略は,取引のための神経ネットワークモデル,RSI指標,スーパートレンド指標を組み合わせています.

論理的には

  1. ボリューム変化,ボリンジャー帯,RSIなどを含むインプットでニューラルネットワークモデルを構築します

  2. ネットワークは将来の価格変動率を予測します

  3. RSI値を計算し,予測価格変化と組み合わせる

  4. RSI をベースに動的なストップ・ロスのラインを生成する

  5. 価格がダウンストップを下回るとショート,ストップ・ロストを下回るとロング

  6. フィルタリングのためにスーパートレンドトレンド判断を使用

この戦略は,複雑なデータをモデル化する神経ネットワークの能力を活用し,リスクを制御しながら正確性を向上させるために,RSIやスーパートレンドなどの指標からの追加の信号検証をします.

利点

  • ニューラルネットワークは多次元データをモデル化して 傾向を特定します

  • RSIは利益を守る 超トレンドは判断を助ける

  • 複数の指標が結合して信号の質を向上させる

リスク

  • ニューラルネットワークの訓練のために大きなデータセットが必要です

  • RSI と Super Trend パラメータの精細調整が必要

  • 性能はモデル予測に依存し,不確実性がある

概要

この戦略は,機械学習と,リスク制御の効率化のための従来の技術を組み合わせています.しかし,パラメータとモデル解釈性は改善する必要があります.


/*backtest
start: 2023-08-14 00:00:00
end: 2023-09-13 00:00:00
period: 2h
basePeriod: 15m
exchanges: [{"eid":"Futures_Binance","currency":"BTC_USDT"}]
*/

//@version=4
//ANN taken from https://www.tradingview.com/script/Eq4zZsTI-ANN-MACD-BTC/
//it only work for BTC as the ANN is trained for this data only
//super trend https://www.tradingview.com/script/VLWVV7tH-SuperTrend/
// Strategy version created for @che_trader
strategy ("ANN RSI SUPER TREND STRATEGY BY che_trader", overlay = true)
qty = input(10000, "Buy quantity")

testStartYear = input(2019, "Backtest Start Year")
testStartMonth = input(1, "Backtest Start Month")
testStartDay = input(1, "Backtest Start Day")
testStartHour = input(0, "Backtest Start Hour")
testStartMin = input(0, "Backtest Start Minute")
testPeriodStart = timestamp(testStartYear,testStartMonth,testStartDay,testStartHour,testStartMin)
testStopYear = input(2099, "Backtest Stop Year")
testStopMonth = input(1, "Backtest Stop Month")
testStopDay = input(30, "Backtest Stop Day")
testPeriodStop = timestamp(testStopYear,testStopMonth,testStopDay,0,0)
testPeriod() => true

max_bars_back = (21)
src = close[0]

// Essential Functions

// Highest - Lowest Functions ( All efforts goes to RicardoSantos )

f_highest(_src, _length)=>
    _adjusted_length = _length < 1 ? 1 : _length
    _value = _src
    for _i = 0 to (_adjusted_length-1)
        _value := _src[_i] >= _value ? _src[_i] : _value
    _return = _value

f_lowest(_src, _length)=>
    _adjusted_length = _length < 1 ? 1 : _length
    _value = _src
    for _i = 0 to (_adjusted_length-1)
        _value := _src[_i] <= _value ? _src[_i] : _value
    _return = _value

// Function Sum  

f_sum(_src , _length) => 

    _output  = 0.00 
    
    _length_adjusted = _length < 1 ? 1 : _length
    
    for i = 0 to _length_adjusted-1
        _output := _output + _src[i]


// Unlocked Exponential Moving Average Function

f_ema(_src, _length)=>
    _length_adjusted = _length < 1 ? 1 : _length
    _multiplier = 2 / (_length_adjusted + 1)
    _return  = 0.00
    _return := na(_return[1]) ? _src : ((_src - _return[1]) * _multiplier) + _return[1]


// Unlocked Moving Average Function 

f_sma(_src, _length)=>
    
    _output = 0.00
    _length_adjusted = _length < 0 ? 0 : _length
    w = cum(_src)

    _output:= (w - w[_length_adjusted]) / _length_adjusted
   
    _output    


// Definition : Function Bollinger Bands

Multiplier = 2 
_length_bb = 20


e_r = f_sma(src,_length_bb)


// Function Standard Deviation : 

f_stdev(_src,_length) =>

    float _output = na 
    _length_adjusted = _length < 2 ? 2 : _length
    _avg  = f_ema(_src , _length_adjusted)
    evar  = (_src - _avg) * (_src - _avg)
    evar2 = ((f_sum(evar,_length_adjusted))/_length_adjusted)
    
    _output := sqrt(evar2)


std_r = f_stdev(src , _length_bb )


upband = e_r + (Multiplier * std_r)  // Upband
dnband = e_r - (Multiplier * std_r)  // Lowband
basis  = e_r                         // Midband

// Function : RSI


length = input(14, minval=1) // 


f_rma(_src, _length) =>
    _length_adjusted = _length < 1 ? 1 : _length
    alpha = _length_adjusted
    sum = 0.0
    sum := (_src + (alpha - 1) * nz(sum[1])) / alpha



f_rsi(_src, _length) => 

    _output = 0.00 
    _length_adjusted = _length < 0 ? 0 : _length

    u = _length_adjusted < 1 ? max(_src - _src[_length_adjusted], 0) : max(_src - _src[1] , 0) // upward change
    d = _length_adjusted < 1 ? max(_src[_length_adjusted] - _src, 0) : max(_src[1] - _src , 0) // downward change
    rs = f_rma(u, _length) / f_rma(d, _length)
    res = 100 - 100 / (1 + rs)
    res


_rsi = f_rsi(src, length)


// MACD 

_fastLength   = input(12 , title = "MACD Fast Length")
_slowlength   = input(26 , title = "MACD Slow Length")
_signalLength = input(9  , title = "MACD Signal Length")


_macd   = f_ema(close, _fastLength) - f_ema(close, _slowlength)
_signal = f_ema(_macd, _signalLength)
	   
_macdhist = _macd - _signal


// Inputs on Tangent Function : 

tangentdiff(_src) => nz((_src - _src[1]) / _src[1] ) 


// Deep Learning Activation Function (Tanh) : 

ActivationFunctionTanh(v) => (1 - exp(-2 * v))/( 1 + exp(-2 * v))


// DEEP LEARNING 

// INPUTS : 

input_1 = tangentdiff(volume)
input_2 = tangentdiff(dnband)
input_3 = tangentdiff(e_r)
input_4 = tangentdiff(upband)
input_5 = tangentdiff(_rsi)
input_6 = tangentdiff(_macdhist)

// LAYERS : 

// Input Layers 

n_0 = ActivationFunctionTanh(input_1 + 0)   
n_1 = ActivationFunctionTanh(input_2 + 0) 
n_2 = ActivationFunctionTanh(input_3 + 0) 
n_3 = ActivationFunctionTanh(input_4 + 0) 
n_4 = ActivationFunctionTanh(input_5 + 0)
n_5 = ActivationFunctionTanh(input_6 + 0)


// Hidden Layers 

n_6   = ActivationFunctionTanh( -2.580743 * n_0 + -1.883627 * n_1 + -3.512462 * n_2 + -0.891063 * n_3 + -0.767728 * n_4 + -0.542699 * n_5 +  0.221093) 
n_7   = ActivationFunctionTanh( -0.131977 * n_0 + -1.543499 * n_1 +  0.019450 * n_2 +  0.041301 * n_3 + -0.926690 * n_4 + -0.797512 * n_5 + -1.804061) 
n_8   = ActivationFunctionTanh( -0.587905 * n_0 + -7.528007 * n_1 + -5.273207 * n_2 +  1.633836 * n_3 +  6.099666 * n_4 +  3.509443 * n_5 + -4.384254) 
n_9   = ActivationFunctionTanh( -1.026331 * n_0 + -1.289491 * n_1 + -1.702887 * n_2 + -1.052681 * n_3 + -1.031452 * n_4 + -0.597999 * n_5 + -1.178839) 
n_10  = ActivationFunctionTanh( -5.393730 * n_0 + -2.486204 * n_1 +  3.655614 * n_2 +  1.051512 * n_3 + -2.763198 * n_4 +  6.062295 * n_5 + -6.367982) 
n_11  = ActivationFunctionTanh(  1.246882 * n_0 + -1.993206 * n_1 +  1.599518 * n_2 +  1.871801 * n_3 +  0.294797 * n_4 + -0.607512 * n_5 + -3.092821) 
n_12  = ActivationFunctionTanh( -2.325161 * n_0 + -1.433500 * n_1 + -2.928094 * n_2 + -0.715416 * n_3 + -0.914663 * n_4 + -0.485397 * n_5 + -0.411227) 
n_13  = ActivationFunctionTanh( -0.350585 * n_0 + -0.810108 * n_1 + -1.756149 * n_2 + -0.567176 * n_3 + -0.954021 * n_4 + -1.027830 * n_5 + -1.349766) 


// Output Layer 

_output  = ActivationFunctionTanh(2.588784 * n_6  + 0.100819 * n_7  + -5.305373 * n_8  + 1.167093 * n_9  + 
                                  3.770143 * n_10 + 1.269190 * n_11 +  2.090862 * n_12 + 0.839791 * n_13 + -0.196165)

_chg_src = tangentdiff(src) * 100

_seed = (_output - _chg_src)
// BEGIN ACTUAL STRATEGY
length1 = input(title="RSI Period", type=input.integer, defval=21)
mult = input(title="RSI Multiplier", type=input.float, step=0.1, defval=4.0)
wicks = input(title="Take Wicks into Account ?", type=input.bool, defval=false)
showLabels = input(title="Show Buy/Sell Labels ?", type=input.bool, defval=true)

srsi = mult* rsi(_seed ,length1)

longStop = hl2 - srsi
longStopPrev = nz(longStop[1], longStop)
longStop := (wicks ? low[1] : close[1]) > longStopPrev ? max(longStop, longStopPrev) : longStop

shortStop = hl2 + srsi
shortStopPrev = nz(shortStop[1], shortStop)
shortStop := (wicks ? high[1] : close[1]) < shortStopPrev ? min(shortStop, shortStopPrev) : shortStop

dir = 1
dir := nz(dir[1], dir)
dir := dir == -1 and (wicks ? high : close) > shortStopPrev ? 1 : dir == 1 and (wicks ? low : close) < longStopPrev ? -1 : dir

longColor = color.green
shortColor = color.red

plot(dir == 1 ? longStop : na, title="Long Stop", style=plot.style_linebr, linewidth=2, color=longColor)
buySignal = dir == 1 and dir[1] == -1
plotshape(buySignal ? longStop : na, title="Long Stop Start", location=location.absolute, style=shape.circle, size=size.tiny, color=longColor, transp=0)
plotshape(buySignal and showLabels ? longStop : na, title="Buy Label", text="Buy", location=location.absolute, style=shape.labelup, size=size.tiny, color=longColor, textcolor=color.white, transp=0)

plot(dir == 1 ? na : shortStop, title="Short Stop", style=plot.style_linebr, linewidth=2, color=shortColor)
sellSignal = dir == -1 and dir[1] == 1
plotshape(sellSignal ? shortStop : na, title="Short Stop Start", location=location.absolute, style=shape.circle, size=size.tiny, color=shortColor, transp=0)
plotshape(sellSignal and showLabels ? shortStop : na, title="Sell Label", text="Sell", location=location.absolute, style=shape.labeldown, size=size.tiny, color=shortColor, textcolor=color.white, transp=0)





if testPeriod() and buySignal
    strategy.entry("Long",strategy.long)

if testPeriod() and sellSignal
    strategy.entry("Short",strategy.short)

もっと