How do you make the renko charts immune to this problem? Any takers?
Machine Learning with algoTraderJo
- #864
- Edited 11:44am Mar 12, 2017 10:33am | Edited 11:44am
- Joined Dec 2014 | Status: Trader | 412 Posts
{quote} As far as I can see my Renko charts do look the same regardless on when they start. I think the point is that Renko charts build their bricks on levels measured from the zero level rather than from the starting tick (if this was your objection). The objection about testing I did not get. IMHO there is no difference in fidelity comparing to the standard charts (if both constructed from ticks).
Even with these problem I still favor this solution of establishing a fixed reference point. It is much less complicated than something like the default implementation in the F4 framework which frankly I don't see as even remotely useful. In the default F4 code for renko bars they are always constructed using the first price in the buffer as reference. This makes it practically useless as the renko between two system calls is always different since the buffer is a moving window of price data.
For those guys who are interested in using renkos built with zero as the base reference I coded a function for the F4 framework in C++. The function is based on the addNewRenko included with the framework but does not have the problems I talked about:
Inserted Code
AsirikuyReturnCode EasyTrade::addNewRenkoRatesFromZero(int originalRatesIndex, int ratesIndex, double renkoSize)
{
int arraySize, i, newCandle;
double cumulativeVolume;
double currentOpen, lowest, highest, lastOpen;
double upperLimit, lowerLimit;
double originalLow, originalHigh;
time_t currentTime;
arraySize = pParams->ratesBuffers->rates[originalRatesIndex].info.arraySize;
pParams->ratesBuffers->rates[ratesIndex].info.arraySize = 0 ;
newCandle = TRUE;
cumulativeVolume = 0;
for(i=0;i<(arraySize-1);i++)
{
if (newCandle == TRUE)
{
if (pParams->ratesBuffers->rates[ratesIndex].info.arraySize == 0){
currentOpen = pParams->ratesBuffers->rates[originalRatesIndex].open[i] ;
} else {
currentOpen = lastOpen;
}
upperLimit = renkoSize*ceil(currentOpen/renkoSize)+renkoSize ;
lowerLimit = renkoSize*floor(currentOpen/renkoSize)-renkoSize ;
currentTime = pParams->ratesBuffers->rates[originalRatesIndex].time[i];
lowest = currentOpen;
highest = currentOpen;
cumulativeVolume = 0;
newCandle = FALSE;
}
cumulativeVolume += pParams->ratesBuffers->rates[originalRatesIndex].volume[i];
originalLow = pParams->ratesBuffers->rates[originalRatesIndex].low[i];
originalHigh = pParams->ratesBuffers->rates[originalRatesIndex].high[i];
if (originalLow < lowest) lowest = originalLow;
if (originalHigh > highest) highest = originalHigh;
if ((originalLow <= lowerLimit) || (originalHigh >= upperLimit) || (i == arraySize-2))
{
pParams->ratesBuffers->rates[ratesIndex].info.arraySize += 1 ;
if (pParams->ratesBuffers->rates[ratesIndex].info.arraySize > 1)
{
pParams->ratesBuffers->rates[ratesIndex].time = (time_t*)realloc(pParams->ratesBuffers->rates[ratesIndex].time, pParams->ratesBuffers->rates[ratesIndex].info.arraySize * sizeof(time_t));
pParams->ratesBuffers->rates[ratesIndex].open = (double*)realloc(pParams->ratesBuffers->rates[ratesIndex].open, pParams->ratesBuffers->rates[ratesIndex].info.arraySize * sizeof(double));
pParams->ratesBuffers->rates[ratesIndex].high = (double*)realloc(pParams->ratesBuffers->rates[ratesIndex].high, pParams->ratesBuffers->rates[ratesIndex].info.arraySize * sizeof(double));
pParams->ratesBuffers->rates[ratesIndex].low = (double*)realloc(pParams->ratesBuffers->rates[ratesIndex].low, pParams->ratesBuffers->rates[ratesIndex].info.arraySize * sizeof(double));
pParams->ratesBuffers->rates[ratesIndex].close = (double*)realloc(pParams->ratesBuffers->rates[ratesIndex].close, pParams->ratesBuffers->rates[ratesIndex].info.arraySize * sizeof(double));
pParams->ratesBuffers->rates[ratesIndex].volume = (double*)realloc(pParams->ratesBuffers->rates[ratesIndex].volume, pParams->ratesBuffers->rates[ratesIndex].info.arraySize * sizeof(double));
}
else
{
pParams->ratesBuffers->rates[ratesIndex].time = (time_t*)malloc(pParams->ratesBuffers->rates[ratesIndex].info.arraySize * sizeof(time_t));
pParams->ratesBuffers->rates[ratesIndex].open = (double*)malloc(pParams->ratesBuffers->rates[ratesIndex].info.arraySize * sizeof(double));
pParams->ratesBuffers->rates[ratesIndex].high = (double*)malloc(pParams->ratesBuffers->rates[ratesIndex].info.arraySize * sizeof(double));
pParams->ratesBuffers->rates[ratesIndex].low = (double*)malloc(pParams->ratesBuffers->rates[ratesIndex].info.arraySize * sizeof(double));
pParams->ratesBuffers->rates[ratesIndex].close = (double*)malloc(pParams->ratesBuffers->rates[ratesIndex].info.arraySize * sizeof(double));
pParams->ratesBuffers->rates[ratesIndex].volume = (double*)malloc(pParams->ratesBuffers->rates[ratesIndex].info.arraySize * sizeof(double));
}
if (originalLow <= lowerLimit)
{
pParams->ratesBuffers->rates[ratesIndex].time[pParams->ratesBuffers->rates[ratesIndex].info.arraySize-1] = currentTime;
pParams->ratesBuffers->rates[ratesIndex].open[pParams->ratesBuffers->rates[ratesIndex].info.arraySize-1] = currentOpen;
pParams->ratesBuffers->rates[ratesIndex].high[pParams->ratesBuffers->rates[ratesIndex].info.arraySize-1] = highest;
pParams->ratesBuffers->rates[ratesIndex].low[pParams->ratesBuffers->rates[ratesIndex].info.arraySize-1] = lowerLimit;
pParams->ratesBuffers->rates[ratesIndex].close[pParams->ratesBuffers->rates[ratesIndex].info.arraySize-1] = lowerLimit;
pParams->ratesBuffers->rates[ratesIndex].volume[pParams->ratesBuffers->rates[ratesIndex].info.arraySize-1] = cumulativeVolume;
lastOpen = lowerLimit;
}
if (originalHigh >= upperLimit)
{
pParams->ratesBuffers->rates[ratesIndex].time[pParams->ratesBuffers->rates[ratesIndex].info.arraySize-1] = currentTime;
pParams->ratesBuffers->rates[ratesIndex].open[pParams->ratesBuffers->rates[ratesIndex].info.arraySize-1] = currentOpen;
pParams->ratesBuffers->rates[ratesIndex].high[pParams->ratesBuffers->rates[ratesIndex].info.arraySize-1] = upperLimit;
pParams->ratesBuffers->rates[ratesIndex].low[pParams->ratesBuffers->rates[ratesIndex].info.arraySize-1] = lowest;
pParams->ratesBuffers->rates[ratesIndex].close[pParams->ratesBuffers->rates[ratesIndex].info.arraySize-1] = upperLimit;
pParams->ratesBuffers->rates[ratesIndex].volume[pParams->ratesBuffers->rates[ratesIndex].info.arraySize-1] = cumulativeVolume;
lastOpen = upperLimit;
}
if ((i == arraySize-2) && (originalLow > lowerLimit) && (originalHigh < upperLimit))
{
pParams->ratesBuffers->rates[ratesIndex].time[pParams->ratesBuffers->rates[ratesIndex].info.arraySize-1] = currentTime;
pParams->ratesBuffers->rates[ratesIndex].open[pParams->ratesBuffers->rates[ratesIndex].info.arraySize-1] = currentOpen;
pParams->ratesBuffers->rates[ratesIndex].high[pParams->ratesBuffers->rates[ratesIndex].info.arraySize-1] = highest;
pParams->ratesBuffers->rates[ratesIndex].low[pParams->ratesBuffers->rates[ratesIndex].info.arraySize-1] = lowest;
pParams->ratesBuffers->rates[ratesIndex].close[pParams->ratesBuffers->rates[ratesIndex].info.arraySize-1] = pParams->ratesBuffers->rates[originalRatesIndex].close[i];
pParams->ratesBuffers->rates[ratesIndex].volume[pParams->ratesBuffers->rates[ratesIndex].info.arraySize-1] = cumulativeVolume;
}
newCandle = TRUE;
}
}
return SUCCESS;
} The beauty of the above is that it can be called in the same way as the other renko function, so you basically do a simple call like:
Inserted Code
addNewRenkoRatesFromZero(PRIMARY_RATES,2,0.0050);
This builds a renko chart for the main pair you used, the last value you pass is the renko size in price units. The above would be 50 pips for the EURUSD. I thought this function to trade and backtest accurately over OHLC, not ticks, it provides a consistent renko chart since it only relies on already closed OHLC bars to create the renko chart. If you trade only at the close of a given OHLC timeframe, say 5 min, this will give you a consistent renko picture. The renko will not be the same as a renko constructed from ticks, especially if you use a renko size that can often be within the range of the OHLC timeframe you choose, but if whatever picture is drawn will always be the same. The price you pay is that you can only trade at the open of a new OHLC bar since information for the renko is not updated before that bar closes. This avoids the accuracy problems when you create renko bars using OHLC instead of ticks and try to trade between bars.
With this I have been able to get some decent machine learning done and my backtests and live trading match as well. Below a visual example of how the renko is created, here a 50 pip renko on the USD/JPY in 2016 data in F4. Renkos built using 5min OHLC. The video shows how it builds as you would expect a consistent renko chart to be built. For each frame F4 rebuilds the renko from scratch using the function I wrote.
Simple example of USDJPY Linear regression algo using 1H charts and 60 pip size renko bars:
The ML algo here uses renko bar returns as inputs and uses the MFE/MAE in a fixed renko bar frontier as the output. We can improve on this by also using inputs related to the time between renko bars. This example is just to show that the renkos themselves can give us something useful to start with and it doesn't have to be a smaller renko size using a more complex lower timeframe either.
Inserted Code
SL_ATR_MULTIPLIER=1.3 BARS_IN_FRONTIER=60 BARS_PER_EXAMPLE=2.0 NUM_OF_EXAMPLES=50 RENKO_SIZE_PIPS=60
The ML algo here uses renko bar returns as inputs and uses the MFE/MAE in a fixed renko bar frontier as the output. We can improve on this by also using inputs related to the time between renko bars. This example is just to show that the renkos themselves can give us something useful to start with and it doesn't have to be a smaller renko size using a more complex lower timeframe either.
Time differences in renko as ML inputs can lead to some improvements in results. Here is a similar result from USDJPY trading on renko bars derived from 1H OHLC bars (trading only on the 1H open).
There is also no reason why we cannot create renko charts using different sizes and learn from all of them. This next.
Inserted Code
SL_ATR_MULTIPLIER=1.2 BARS_IN_FRONTIER=60 BARS_PER_EXAMPLE=5.0 NUM_OF_EXAMPLES=50 RENKO_SIZE_PIPS=60
There is also no reason why we cannot create renko charts using different sizes and learn from all of them. This next.
