Warm up period

[tubular2000]

hi, is it possible to add a warm up period when calling run()? if my indicator needs 500 bars to start generating an output, i don't want these bars to be included in the backtest range.

[kernc]

We do skip the first couple bars, precisely to avoid testing on NaN indicators:

backtesting.py/backtesting/backtesting.py

Lines 1144 to 1147 in 0b2325f

	# Skip first few candles where indicators are still "warming up"
	# +1 to have at least two entries available
	start = 1 + max((np.isnan(indicator.astype(float)).argmin(axis=-1).max()
	for _, indicator in indicator_attrs), default=0)

By backtest range, do you mean stats['End'] and stats['Duration']?

[kernc]

Even if you include this range, it won't influence the stats if your indicators aren't forward-looking (as they shouldn't be). For instance, with SMA(df.Close, 20), the first 19 result values will be equal to NaN and thus skipped by the code snippet above.

[tubular2000]

thank you for your help. i managed to make it work by precomputing my indicators in the main dataframe and backtesting a subset like you proposed.

thanks for making this library available to us!

[mikeFyngers]

Thanks kernc and all for the great package!

I implemented what I understand is the suggested solution but I have a major drawback now. I cannot use the class parameters anymore for optimization. Can I circumvent this?

I implemented a long term version of the SMA cross over in the first tutorial, see https://kernc.github.io/backtesting.py/doc/examples/Quick%20Start%20User%20Guide.html. DataFrame prices has the OHLC data from which I precompute the indicator data. I then prune the prices to remove the warm-up period because in reality I will have the indicator available already on day one. I define a PassThrough function to ingest the precomputed data via the suggested indicator definition via self.I(). However, the class variables n1 and n2 are then not used at all. Is there an alternative to including a second DataFrameon top of data.self with additional 200 days prepended (in this case). I understand that this opens the door for nasty issues of misaligning data but my current solution isnt satisfactory either. Thanks!

import pandas as pd

def SMA(values, n):
    """
    Return simple moving average of `values`, at
    each step taking into account `n` previous values.
    """
    return pd.Series(values).rolling(n).mean()

def PassThrough(values):
    """
    Return unchanged `values`, required for call to Indicator creation on Strategy.
    """
    return values

prices=prices.assign(sma200c=SMA(prices.Close, 200))
prices=prices.assign(sma50c=SMA(prices.Close, 50))

# Remove the first 200 rows such that we have a data set without warm-up
prices=prices.iloc[slice(199, len(prices))]

from backtesting import Strategy
from backtesting.lib import crossover

class SmaCross(Strategy):
    # Define the two MA lags as *class variables*
    # for later optimization
    n1 = 50 # Fast
    n2 = 200 # Slow
    
    def init(self):
        # Get precomputed data
        self.sma1 = self.I(PassThrough, self.data.sma50c, name='SMA 50c', overlay=True)
        self.sma2 = self.I(PassThrough, self.data.sma200c, name='SMA 200c', overlay=True)
    
    def next(self):
        #print(self.sma1)
        # If sma1 (fast) crosses above sma2 (slow), close any existing
        # short trades, and buy the asset
        if crossover(self.sma1, self.sma2):
            self.position.close()
            self.buy()

        # Else, if sma1 (fast) crosses below sma2 (slow) , close any existing
        # long trades, but dont sell the asset
        # See also https://github.com/kernc/backtesting.py/discussions/353        
        elif crossover(self.sma2, self.sma1):
            self.position.close()
            
        elif not self.position and self.sma1[-1]>self.sma2[-1]:
            # When strategy starts and there is no position yet, buy 
            # if today's sma1 > sma2 (instead of waiting for first crossing)
            self.buy()

from backtesting import Backtest

bt = Backtest(prices, SmaCross, cash=10_000, commission=.002)
stats = bt.run()
stats['_trades']

[kernc]

@mikeFyngers If you simply did:

        self.sma1 = self.I(SMA, self.data.Close, 50, overlay=True)
        self.sma2 = self.I(SMA, self.data.Close, 200, overlay=True)

the initial 200 bars with undefined "SMA200c" will get skipped automatically. Why not do it like that?

[mikeFyngers]

Yeah, that's what I took originally from the tutorial.

The reason is that I want to prescribe the backtest start date and (less important) the "buy and hold return stat" to be meaningful. I guess for the former I'll start the backtest 200 days earlier (if my indicator needs 200 days warm-up) and ignore the latter. This is also a fair solution, and better than the precomputation marked as preferred answer by tubular2000.

[zillionare]

the warm up period should be excluded from Buy & Hold Return calculation also.

[AGG2017]

We must have optional parameters date_begin and date_end to run the backtesting. Now I'm trying to implement this because I badly need it as well as the trading orders over the options prices provided with a custom function from a database. I see that statistics use first and last value for the closing price from the prices dataframe to calculate the duration of the trades and the corresponding Buy & Hold Return, etc. This way if you have let say using Wave C for the John Carter's Squeeze system then the backtesting will skip first ~400 bars for the trading but they will be included for the Buy & Hold Return which is wrong. We must define the starting date for everything, otherwise we will never create a strategy close enough to the wrongly calculated Buy & Hold Return with 2 years (in my case) of advance calculation.

[kernc]

This is already tracked as issue #327. The acceptable solution would be to automatically clip Buy&Hold in stats to the active period.

[AGG2017]

I didn't see this issue but I mention my observation just to accent how important is to have clear start and end for everything to avoid such mysteries how good is just to buy and hold.
I did some initial implementation and the easiest way for me was to add begin and end index parameters for the stat function. By default they are 0 and -1 to keep the function compatible. The run function accept date_begin and date_end as dates, skip the strategy processing outside of this date range and convert these dates as begin/end index that are transferred to stat for proper calculations. Everything looks good, clean and easy so far but I have to do more testing to be sure I don't miss anything.