SUM Function Overflow?

[dwrety]

I am designing an adaptive lookback period for Kaufman's Efficiency Ratio. The Efficiency Ratio logic is copied from the built-in KAMA indicator. Here is the logic.
Basically I want a the lookback period to be able to change on every bar defined by this integer p, p is between _MinERPeriod and _BaseERPeriod. However, when too many bars are loaded. The variable

noise

is getting weird. It can sometimes become negative or very close to 0. This issue goes away if I load less than 1000 candle bars. I am wondering if there is a way to get around this issue.

protected override void OnBarUpdate()
        {
            er[0] = 0;
            diffSeries[0] = 0;
            diffSeries[0] = CurrentBar > 0 ? Math.Abs(Input[0] - Input[1]): 0;
            double DCWidthBase = DonchianChannel(_BaseERPeriod).Upper[0] - DonchianChannel(_BaseERPeriod).Lower[0];
            int p = _MinERPeriod;
            for (int i = _MinERPeriod; i<= _BaseERPeriod; i++)
            {
                p = i;
                double DCWidth = DonchianChannel(i).Upper[0] - DonchianChannel(i).Lower[0];
                if (DCWidth >= _PctThreshold * DCWidthBase)
                {
                    break;
                }
            }
            if (CurrentBar < DCWidthBase)
            {
                return;
            }
            
            double signal = Math.Abs(Input[0] - Input[p]);
            double noise = SUM(diffSeries, p)[0];
            
            if (noise == 0)
            {
                er[0] = er[1];
                return;
            }
            if (noise < 0)
            {
                Print(p);
                return;
            }
            er[0] = signal / noise;
            ER[0] = er[0];​
}

[Reycastro]

protected override void Initialize()
{
diffSeries = new DataSeries(this);
noiseSums = new Sum[_BaseERPeriod - _MinERPeriod + 1];
for(int i = _MinERPeriod; i <= _BaseERPeriod; i++)
noiseSums[i - _MinERPeriod] = SUM(diffSeries, i);
}

protected override void OnBarUpdate()
{
// need at least one prior bar for diff
if (CurrentBar == 0) return;
diffSeries[0] = Math.Abs(Input[0] - Input[1]);

// pick p based on Donchian widths
double baseWidth = DonchianChannel(_BaseERPeriod).Upper[0]
- DonchianChannel(_BaseERPeriod).Lower[0];
int p = _BaseERPeriod;
for(int i = _MinERPeriod; i <= _BaseERPeriod; i++)
{
double w = DonchianChannel(i).Upper[0] - DonchianChannel(i).Lower[0];
if (w >= _PctThreshold * baseWidth)
{
p = i;
break;
}
}

// make sure we have at least p bars
if (CurrentBar < p)
{
er[0] = er[1];
return;
}

// compute ER safely
double signal = Math.Abs(Input[0] - Input[p]);
double noise = noiseSums[p - _MinERPeriod][0];
noise = Math.Max(noise, 1e-8);

er[0] = signal / noise;
ER[0] = er[0];
}


With that in place:

• You never call SUM with a period you don’t yet have bars for.
• You never let noise go to zero or negative.
• NinjaTrader’s internal SUM buffers stay in lockstep with your DataSeries.

That will eliminate the weird negative/near-zero spikes when you load thousands of bars.