and if i calculate the hourly non-compounded returns manually with:
def prenext(self):
self.last_value = self.broker.getvalue()
def next(self) :
self.value = self.broker.getvalue()
self.hourly_ret += [np.floor((self.value/self.last_value - 1.) * 1e8) / 1e8]
self.last_value = self.value
i get the correct returns and my value at that candle differs from the one the observer calculates.
somehow I always get the difference at the end of a day.. ? probably that should give me a clue.
@alain I actually went through a trade and the error happens at a specific candle. Directly compounded interest (and right result) says there is a small gain, hourly return (indicator in picture) says there's a small loss:
Hi,
To combine 2 strategies (hourly data) I need the non-compounded hourly returns of each.
So far for the compounded returns for every hour I've always used the benchmark observer. More precisely the benchmark.timereturn -part as can be seen in the #ret1 part of the code bellow.
I've now tried to return the hourly returns (non-compounded) by using a specific timeframe in the TimeReturn-observer.
This can be seen in the following code:
#compounded returns
cerebro1.addobserver(bt.observers.Benchmark, timeframe=bt.TimeFrame.NoTimeFrame, data=data, plot=False)
#non-compounded hourly returns
cerebro1.addobserver(bt.observers.TimeReturn, timeframe=bt.TimeFrame.Minutes, compression=60)
#show result
get_trade_rets(thestrat1, startdate, enddate, maxperiod, timeframe)
def get_trade_rets(thestrat, startdate, enddate, maxperiod, timeframe):
dt_hours = pd.date_range(start=startdate, end=enddate, freq=timeframe)
#ret1
ret1_compounded = thestrat.observers.benchmark.timereturn.lines[0].array[maxperiod:(maxperiod + len(dt_hours))]
ret1_compounded = 1 + np.array(ret1_compounded)
print(ret1_compounded)
#ret2
ret2_hourly = thestrat.observers.timereturn.lines[0].array[maxperiod:(maxperiod + len(dt_hours))]
ret2_hourly = np.array(ret2_hourly)
ret2_compounded = np.cumprod(1 + ret2_hourly)
print(ret2_compounded)
I then wanted to be sure that those non-compounded hourly returns I received in the section #ret2 are correct and tried to reproduce the compounded returns from #ret1 (ret1_compounded).
To compound returns I use the standard equation "ret2_compounded = np.cumprod(1 + ret2_hourly)".
What I receive is a new list that starts the same as the directly compounded list from #ret1, but at a certain point the values in the two lists are not the same anymore.
[1. 1. 1. 1. 1. 1. 1. 1. 1. 1. 1. 1. 1. 1. 1. 1. 1. 1. 1. 1. 1. 1. 1. 1. 1.0022292 1.00308236 1.00617001 1.00507308 1.00482932 1.00036036 1.00109164 0.99816651 0.99857278 0.99629767 0.99597266 1.00149791 0.99889779 1.00511371 1.00730756 1.00470744 1.00491057 1.01888623 1.0236802 1.02766164 1.03046489 1.02676785 1.02343644 1.02372083 ...
[1. 1. 1. 1. 1. 1. 1. 1. 1. 1. 1. 1. 1. 1. 1. 1. 1. 1. 1. 1. 1. 1. 1. 1. 1.0022292 1.00308236 1.00617001 1.00507308 1.00482932 1.00036036 1.00109164 0.99816651 0.99857278 0.99629767 0.99597266 1.00149791 0.99889779 1.00511371 1.00730756 1.00470744 1.00491057 1.01888623 1.0236802 1.02766164 1.03046489 1.02676785 1.02343644 1.02039931 ...
I can't explain myself why that is...
Would be great if anyone knows what the issue could be.
Thank you!
how i solved the problem (just used next method to calculate the SMMA explicitly in each step):
class RelativeVolatility(bt.Indicator):
lines = ('rel_vol', 'signal', 'nU', 'nD',)
plotlines = dict(nU=dict(_plotskip=True), nD=dict(_plotskip=True))
params = (('period', 10), ('signal_line_top', 80), ('signal_line_bottom', 20),)
def _plotinit(self):
self.plotinfo.plothylines = [self.p.signal_line_top, self.p.signal_line_bottom]
def __init__(self):
std_dev = bt.ind.StdDev(self.data.close, period=self.p.period)
self.u = bt.If(self.data.close > self.data.close(-1), std_dev, 0)
self.d = bt.If(self.data.close < self.data.close(-1), std_dev, 0)
def nextstart(self):
self.l.nU[0] = 0
self.l.nD[0] = 0
def next(self):
self.l.nU[0] = (13. * self.l.nU[-1] + self.u[0]) / 14.
self.l.nD[0] = (13. * self.l.nD[-1] + self.d[0]) / 14.
self.l.rel_vol[0] = 100. * self.l.nU[0] / (self.l.nU[0] + self.l.nD[0])
I tried again to directly switch SMA to SMMA.. but it really doesn't work for me:
def __init__(self):
std_dev = bt.ind.StdDev(self.data.close, period=self.p.period)
u = bt.If(self.data.close > self.data.close(-1), std_dev, 0)
d = bt.If(self.data.close < self.data.close(-1), std_dev, 0)
nU = btind.SMMA(u, period=14)
nD = btind.SMMA(d, period=14)
self.l.rel_vol = 100. * nU / (nU + nD)
gives me lists full of nan's in nU and nD...
@alain alright, i was able to implement the smoothing moving average inside the next function. it's just a bit less elegant.
Hi,
I tried to implement a relative volatility indicator as follows:
class RelativeVolatility(bt.Indicator):
lines = ('rel_vol',)
params = (('period', 10),)
def __init__(self):
std_dev = bt.ind.StdDev(self.data.close, period=self.p.period)
self.u = bt.If(self.data.close > self.data.close(-1), std_dev, 0)
self.d = bt.If(self.data.close < self.data.close(-1), std_dev, 0)
self.nU = btind.SMA(self.u, period=14)
self.nD = btind.SMA(self.d, period=14)
self.l.rel_vol = 100. * self.nU / (self.nU + self.nD)
The code works perfectly. Actually however i want to replace the simple moving average by a smoothed moving average. But as soon as i change btind.SMA to btind.SmoothedMovingAverage i get lists full or 'nan' values.
Would be great if anyone here knows why that is ?
Thanks!
If I had concluded something I wouldn't be asking here...
Ok, gonna do some reading.
I didn't include any riskfree rate. Or say I set it to zero.
So you're right, there are for sure some differences in my calculations compared to the analyzer. You think if I change these two things (# days/rate) it would give me the "right" sharpe ratio ?
So my guess now is that it's exactly this part of the parameters which causes the issue:
If None, the conversion factor for the riskfree rate from annual to the chosen timeframe will be chosen from a predefined table
Days: 252, Weeks: 52, Months: 12, Years: 1
Else the specified value will be used
I should probably specify that it should use 365 days?