BrokenProcessPool when doing backtest.optimize

[sacharbit]

Expected Behavior

Returning the stats on the best run for backtest.run()

Actual Behavior

raised a BrokenProcessPool error : A process in the process pool was terminated abruptly while the future was running or pending.

Steps to Reproduce

bt = Backtest(df, Strat,, cash=starting_money, commission=.002)
bt.run()
stats = bt.optimize(tresholdVol=range(5,100, 5), maximize='Equity Final [$]')

Additional info

• Backtesting version:

[kernc]

Is it possible your strategy raises an unhandled exception somewhere? Did you maybe run out of RAM? Can you attach a whole reproducible MWE example here?

[sacharbit]

def var(values) :
    return (pd.Series(values) - pd.Series(values).shift(1)) / pd.Series(values).shift(1)

def vol(high, low, close) :
    return (pd.Series(high) - pd.Series(low)) / pd.Series(close)

def ma(values, n):
    return pd.Series(values).rolling(n).mean()

class MyStrat(Strategy) :
    n_small = 20
    n_big = 9
    thresholdVol = 50
    
    def init(self) :
        self.volumeVar = self.I(var, self.data.Volume)
        self.priceVar = self.I(var, self.data.Close)
        self.ma_small = self.I(ma, self.data.Close, self.n_small)
        self.ma_big = self.I(ma, self.data.Close, self.n_big)
        
    def next(self) :
        if self.priceVar[-1] > 0 and self.ma_small[-1] < self.ma_big[-1] :
            self.buy()
        elif self.priceVar[-1] < 0 and self.ma_small[-1] > self.ma_big[-1] :
            self.sell()
                
#%%
                
def backtest(name) :
    df = pd.read_csv("stock_data/{}.csv".format(name))
    df["t"] = pd.to_datetime(df["t"])
    df = df.set_index("t")
    df = df[df.Volume != 0]
    bt = Backtest(df, MyStrat, cash=starting_money, commission=.002)
    return bt

#%%
bt = backtest("Apple")
#result = bt.run()
stats = bt.optimize(thresholdVol=range(5,7), maximize='Equity Final [$]')

It's definitely not an issue with the RAM since it's a pretty small dataset and not that many combinaisons that I ask.

[kernc]

Does it work if your wrap bt.optimize() call into if __name__ == '__main__': block?

[sacharbit]

No it doesn't. I run this on Spyder so it's already on main anyway. Does the script that I posted works for you?

[kernc]

It does, but I'm on GNU/Linux. I assumed the problem is spawn process execution method on Windows which, without the main guard, may recurse (a common multiprocessing/futures problem).

What about if you change in backtesting code the line:

backtesting.py/backtesting/backtesting.py

Line 807 in a7d1d4f

with ProcessPoolExecutor() as executor:

to use concurrent.futures.ThreadPoolExecutor instead of ProcessPoolExecutor?

[sacharbit]

it says : name 'concurrent' is not defined

[kernc]

from concurrent.futures import ThreadPoolExecutor  # like ProcessPoolExecutor is

# or

import concurrent.futures  # somewhere at the top

Thanks!

[sacharbit]

Oh Yeah! It works! Awesome!

Thank you very much for the VERY quick response.

[kernc]

Successful diagnosis does not make the issue fixed. 😄

Note to self: Apparently, we should opt for ThreadPoolExecutor (with a performance degradation warning) on platforms where ProcessPoolExecutor fails.

Thanks for investigating this!

[kernc]

@sacharbit Are you maybe willing to test if the workaround proposed in #6 works? 😅

[kernc]

I've been trying to reproduce this without much success. What version of Windows and Spyder were you running this on?

---

Note to self: Switching to ThreadPoolExecutor won't really do with unguarded contexts. It'll just run the computation multiple times.

[sacharbit]

I was using Python 3.7 and windows 10. The spyder version is pretty irrelevant I think.

I'm sorry I didn't reply earlier, I'll try to test it this weekend, I have a lot of work right now.