teddykoker / blog Goto Github PK

Hi,

I am using your code and I wonder if it should be dBdR = 2. * R / T.

Thank you.

Hi Teddy, great post. I'm a bit confused with the following code though:

self.rankings.sort(key=lambda d: self.inds[d]["momentum"][0])

and then
# buy stocks with remaining cash for i, d in enumerate(self.rankings[:int(num_stocks * 0.2)]): cash = self.broker.get_cash() value = self.broker.get_value() if cash <= 0: break if not self.getposition(self.data).size: size = value * 0.001 / self.inds[d]["atr20"] self.buy(d, size=size)

Seems to be picking stocks with the lowest momentum? And It's supposed to be picking stocks with the highest momentum in Andreas F. Clenow's book?

Thanks!

I can't seem to be able to reproduce your results with my own back-testing program.

hence I installed the backtester yesterday, trying to run your program, and found some problems,

in below code, rebalance portfolio never executed correctly.
if self.getposition(self.data).size: ==> should be ==> if self.getposition(d).size: huge bug...

after fixing this bug, I still have some problems... hence I'm contacting you to see if you have fixed any other bugs on this program. or there are some issues in Backtester itself. I am not an expert on Backtester. just started to use it.

my own testers seem to be working correctly.
thanks in advance.

def rebalance_portfolio(self):
# only look at data that we can have indicators for
self.rankings = list(filter(lambda d: len(d) > 100, self.stocks))
self.rankings.sort(key=lambda d: self.inds[d]["momentum"][0])
num_stocks = len(self.rankings)

    # sell stocks based on criteria
    for i, d in enumerate(self.rankings):
        if self.getposition(self.data).size:
            if i > num_stocks * 0.2 or d < self.inds[d]["sma100"]:
                self.close(d)
    
    if self.spy < self.spy_sma200:
        return
    
    # buy stocks with remaining cash
    for i, d in enumerate(self.rankings[:int(num_stocks * 0.2)]):
        cash = self.broker.get_cash()
        value = self.broker.get_value()
        if cash <= 0:
            break
        if not self.getposition(self.data).size:
            size = value * 0.001 / self.inds[d]["atr20"]
            self.buy(d, size=size)

HI teddy

I would like to know as to who should I replace the following part ii , if I whant say instead want to maximize for other variant quantities say like calmar ratio

A = np.mean(R)
B = np.mean(np.square(R))
S = A / np.sqrt(B - A ** 2)    # under function gradient  

the above is just for sharpe ratio , that means that I have to write down the equivalent quantity to be maximized , then is it just enough to the following

def calc_calmar(rets):
# Peaks and bottoms indexes in sequence
mins = np.ravel(argrelmin(rets))
maxs = np.ravel(argrelmax(rets))
extrema = np.concatenate((mins, maxs))
extrema.sort()
return -rets.median()/np.diff(rets[extrema]).min()

then the above is just the another definition just like the one you introduced in your article i.e.

def sharpe_ratio(rets):
return rets.mean() / rets.std()

but now shall I add the quantity under gradient as well

mins = np.ravel(argrelmin(rets))
maxs = np.ravel(argrelmax(rets))
extrema = np.concatenate((mins, maxs)) 
extrema.sort()
return -rets.mean()/np.diff(rets[extrema]).min()

then finding derivatives w.r.t. the parameters of the calamr ratio , which are argmin and argmax ? is that right
your input is highly appreciated
best regards

Just a quick note to thank you for a splendid series of posts on Backtrader.

They are a great read!

Splendid code too.

hello, Thanks for your great work. I learned a lot from your posts

I have a question.

On post

trading-with-reinforcement-learning-in-python-part-two-application.ipynb

each time step, the flowing grad is calculated
(dRdF * dFdtheta + dRdFp * dFpdtheta)

my question is: it seems this is only part of grad of each step?

Hi Teddy,

I was trying to implement the code you have written.
Can you provide a sample csv file so that I can cross check with my csv file whether they are same or not. Because with my csv file, I am getting a lot of error.

There is a data leakage in "trading with reinforcement learning part two" notebook. You need to split data before making normalization, and not the other way around.

I admit your project and math are over my head. I'm still struggling to understand your position function and with that I'm trying to come up with a simple

• input format
• that yields a simple output format (that I can interprete as a signal to buy, sell or hold)

How would that be done in your code?

best30 = momentums30.iloc[-1:].max().sort_values(ascending=False).index[:5]
plt.figure(figsize=(12, 9))
plt.xlabel('Days')
plt.ylabel('Stock Price')
for best in best30:
print ("best = ", best, " momentum = ", momentums30[best].iloc[-1 :])
rets = np.log(stocks[best].iloc[-30 :])
x = np.arange(len(rets))
slope, intercept, r_value, p_value, std_err = linregress(x, rets)
plt.plot(np.arange(90), stocks[best][-90:]);
y = np.arange(60, 90)
z = np.e ** (intercept + slope*x)
plt.plot(y, z)
plt.show();

Hello, great work. But, I just read thru here and found some trouble understanding this. Don't you want to find the best 5 momentum stocks of the the current day, not anyday in the past?
hence, I changed the code to
best30 = momentums30.iloc[-1:].max().sort_values(ascending=False).index[:5]
to get the current day best 5 momentum candidates. ( i used 30 days momentum in this example)

Also changed the plot on regression line from the current day back to 30 days ago.

a. I plotted the 90 day stock prices from 90 days ago to today
b. plotted the regression line only for the last 30 days of the stock prices.

I believe this makes much more sense.

Comments? thanks in advance.

Thanks for the bt version of Clenow momentum equity strategy.

QQ: Should the rankings be reverse sorted, as we want higher values first?

The line

self.rankings.sort(key=lambda d: self.inds[d]["momentum"][0])

would become

self.rankings.sort(key=lambda d: self.inds[d]["momentum"][0], reverse=True)