Possible deviation in backtest results because of timeshift usage

[majdzadeh]

Problem description

The backtesting broker includes a process_pending_orders() method which looks at previous day's data (according to the commented code) when the data source is PANDAS. I have noticed a couple of issues when backtesting with minute-level data:

• The timeshift parameter for get_historical_prices() has been set to -2 however, the returned data is 2 days ahead of the current bar's timestamp.
• The code comment says; This is a hack to get around the fact that we need to get the previous day's date to prevent lookahead bias. whereas, at least with minute-level data, the data is already available and there would be no lookahead bias.
• The use of timeshift causes unexpected backtest results (at least with minute-level data). For example, the price that's used to fill orders with is different from the price that's seen during the trading iteration for the asset.

[majdzadeh]

My suggestion is to introduce a new timeshift argument to process_pending_orders() which defaults to None and use it to set the timeshift argument when calling strategy.get_historical_prices(); with proper validation of course. This has two advantages:

• No more hard-coded -2 as the value for the timeshift argument when calling strategy.get_historical_prices() from BacktestingBroker
• Give more control to the strategy developer and/or backtester on how to fetch prices to process pending orders.

@grzesir kindly share your thoughts.

[majdzadeh]

The other issue that I observe with process_pending_orders() method on the backtesting broker is that there exist two invocations of strategy.get_historical_prices() where for the first one, a timedelta value is used as the value for the timeshift argument but for the second one, an integer value (-2) is used for the same argument.