Fundamental concept I am missing with indicator development: (Adaptive LRSI Filter)

[RandyT]

I've spent a good part of the day trying to get this working and seem to continue to miss some fundamental concepts about indicator development. When something needs to be in __init__() vs next() or something.

I have the following bit of code that does not quite accomplish the goal. Help would be very much appreciated.

class AdaptiveLaguerreFilter(btind.PeriodN):
    '''
    // Adaptive Laguerre Filter, from John Ehlers
    // Link : http://www.mesasoftware.com/Papers/Time%20Warp%20Without%20Space%20Travel.exe
    // Another work from Ehlers : http://www.mesasoftware.com/technicalpapers.htm
    //
    // Description :
    // Laguerre Filtering, (alpha is automaticaly adapted depending the error of filtering)
    function ALFilter(price, length, medianlong) {
        result=price;
        L0 = price;
        L1 = price;
        L2 = price;
        L3 = price;
        coef=0.5;
        Diff=0;
        HH=0.1;
        LL=0;
        alpha=0.5;

        for(i = 1+length; i < BarCount; i++) {
            Diff[i] = abs(price[i] - result[i-1]);
            HH[i] = Diff[i];
            LL[i] = Diff[i];

            for(j = 0; j < (length-1); j++) {
                if (Diff[i-j] > HH[i]) HH[i] = Diff[i-j];
                if (Diff[i-j] < LL[i]) LL[i] = Diff[i-j];
            }

            if ( (i > length) AND (HH[i] - LL[i] != 0) ) {
                coeftemp=(Diff - LL) / (HH - LL);
                mlen = medianlong;
                for(k = mlen - 1; k >= 0; k--) temparray[k] = coeftemp[i + k - (mlen - 1)];
                temp=0;
                for(k = mlen - 1; k > 0; k--) {
                    for (j = mlen - 1; j > 0; j--) {
                        if (temparray[j-1] > temparray[j]) {
                            temp = temparray[j-1];
                            temparray[j-1] = temparray[j];
                            temparray[j] = temp;
                        }
                    }
                }
                coef[i] = temparray[(mlen/2)-0.5];
                //----- End median calculation
            } // end main IF

            alpha=coef[i];
            L0[i] = alpha*price[i] + (1 - alpha)*L0[i-1];
            L1[i] = -(1 - alpha)*L0[i] + L0[i-1] + (1 - alpha)*L1[i-1];
            L2[i] = -(1 - alpha)*L1[i] + L1[i-1] + (1 - alpha)*L2[i-1];
            L3[i] = -(1 - alpha)*L2[i] + L2[i-1] + (1 - alpha)*L3[i-1];
            result[i] = (L0[i] + 2*L1[i] + 2*L2[i] + L3[i]) / 6;
        }// end main  FOR
        return result;
    }
    '''
    alias = ('ALF', 'ALRSI', 'AdaptiveLRSI')
    lines = ('alf',)
    params = (('period', 20),
              ('medianperiod', 5),)
    plotinfo = dict(subplot=False)

    def __init__(self):
        self.l0, self.l1, self.l2, self.l3 = 0.0, 0.0, 0.0, 0.0
        self.diff = abs(self.data - self.l.alf(-1))
        super(AdaptiveLaguerreFilter, self).__init__()

    def next(self):
        l0_1 = self.l0  # cache previous intermediate values
        l1_1 = self.l1
        l2_1 = self.l2
        alpha = 0.5
        llv, hhv = 0.0, 0.0

        llv = btind.Lowest(self.diff, period=self.p.period)
        hhv = btind.Highest(self.diff, period=self.p.period)
        alpha = btind.Mean((self.diff[0] - llv) / (hhv - llv),
                           period=self.p.medianperiod)

        self.l0 = l0 = (alpha * self.data[0]) + ((1.0 - alpha) * l0_1)
        self.l1 = l1 = (-(1 - alpha) * l0) + l0_1 + (1 - alpha) * l1_1
        self.l2 = l2 = (-(1 - alpha) * l1) + l1_1 + (1 - alpha) * l2_1
        self.l3 = l3 = (-(1 - alpha) * l2) + l2_1 + (1 - alpha) * self.l3
        self.l.alf[0] = (l0 + (2 * l1) + (2 * l2) + l3) / 6

[RandyT]

Here is a hack to get this working, but I must be missing something. Looking at some of the other adaptive indicators, it seems it should be easier.

    alias = ('ALF', 'ALRSI', 'AdaptiveLRSI')
    lines = ('alf',)
    params = (('period', 20),
              ('medianperiod', 5),)
    plotinfo = dict(subplot=False)

    def __init__(self):
        self.l0, self.l1, self.l2, self.l3 = 0.0, 0.0, 0.0, 0.0
        self.diff = abs(self.data - self.l.alf(-1))
        self.llv = btind.Lowest(self.diff, period=self.p.period)
        self.hhv = btind.Highest(self.diff, period=self.p.period)
        self.alpha = btind.Mean((self.diff - self.llv) / (self.hhv - self.llv),
                                period=self.p.medianperiod)
        super(AdaptiveLaguerreFilter, self).__init__()

    def next(self):
        l0_1 = self.l0  # cache previous intermediate values
        l1_1 = self.l1
        l2_1 = self.l2

        alpha = self.alpha.av[0] if not math.isnan(self.alpha.av[0]) else 0.5

        self.l0 = l0 = (alpha * self.data) + ((1.0 - alpha) * l0_1)
        self.l1 = l1 = (-(1 - alpha) * l0) + l0_1 + (1 - alpha) * l1_1
        self.l2 = l2 = (-(1 - alpha) * l1) + l1_1 + (1 - alpha) * l2_1
        self.l3 = l3 = (-(1 - alpha) * l2) + l2_1 + (1 - alpha) * self.l3
        self.l.alf[0] = (l0 + (2 * l1) + (2 * l2) + l3) / 6

[backtrader]

• In __init__

  Operations like self.x = (self.data.high + self.data.low) / 2.0 generate a lazily evaluated lines object.

  Even this generates that: self.y = self.data.close > self.data.open

• In next

  Here the [] operator is used to access the values provided by lines objects like the generated self.x or self.y or standard lines like self.data.high

  if self.data.close[0] > self.data.open[0]:
      do_something()
  

  Thanks to operator overloading the following is equivalent also in next

  if self.data.close > self.data.open:
      do_something()
  

  But unlike in __init__, this comparison generates a bool.

One can generate a complete indicator just by using operations and logic in __init__. See for example an old friend of everybody like MACD (removing documentation and plotting preparation boilerplate)

class MACD(bt.Indicator):
    lines = ('macd', 'signal',)
    params = (
        ('period_me1', 12),
        ('period_me2', 26), ('period_signal', 9),
        ('movav', MovAv.Exponential),
    )

    def __init__(self):
        super(MACD, self).__init__()
        me1 = self.p.movav(self.data, period=self.p.period_me1)
        me2 = self.p.movav(self.data, period=self.p.period_me2)
        self.lines.macd = me1 - me2
        self.lines.signal = self.p.movav(self.lines.macd, period=self.p.period_signal)

No need to use scalar operations/logic in next. Everything is defined in terms of lines objects.

Some indicators may need next. Some help from things defined in __init__ can be used. For example the ZeroLagIndicator

class ZeroLagIndicator(MovingAverageBase):
    alias = ('ZLIndicator', 'ZLInd', 'EC', 'ErrorCorrecting',)
    lines = ('ec',)
    params = (
        ('gainlimit', 50),
        ('_movav', MovAv.EMA),
    )

    def __init__(self):
        self.ema = MovAv.EMA(period=self.p.period)
        self.limits = [-self.p.gainlimit, self.p.gainlimit + 1]

        # To make mixins work - super at the end for cooperative inheritance
        super(ZeroLagIndicator, self).__init__()

    def next(self):
        leasterror = MAXINT  # 1000000 in original code
        bestec = ema = self.ema[0]  # seed value 1st time for ec
        price = self.data[0]
        ec1 = self.lines.ec[-1]
        alpha, alpha1 = self.ema.alpha, self.ema.alpha1

        for value1 in range(*self.limits):
            gain = value1 / 10
            ec = alpha * (ema + gain * (price - ec1)) + alpha1 * ec1
            error = abs(price - ec)
            if error < leasterror:
                leasterror = error
                bestec = ec

        self.lines.ec[0] = bestec

Here the base ema for the calculations is defined in __init__ but the core of the operations which is a loop over several values is best done in next.