Debugging the Missing Time Spread Backtrades (Part 1)

Because the current debugging effort is challenging, today I want to put my thoughts to computer screen and push forward to find out why the backtester is missing trades (last discussed here).

Need some juice today? Let’s get in the mood with this (thanks Frida!).

The first thing I should do is run the backtest a couple more times to make sure the values from the table in the previous post are repeatable. I would be more concerned if trades get missed randomly than regularly.

Next, I want to know the values of all relevant variables when trade skipping takes place. I can then trace what I expect should be the logic through the program to see if the results actually match.

Relevant variables can be monitored in one of two ways.

First, I can create a variable containing current_date minus previous trade ending date and raise an exception > 5. This would limit study to one case at a time, which should be sufficient if the cause of missing trades is uniform. Going this route, I need to write the dataframes to .csv and file.close() before raising an exception, which will immediately halt the program. The former will allow me to view the entire dataframe in Excel, which is easier than viewing with Jupyter Notebook. file.close() will avoid the annoying “file in use” warning when I try to open the file after an exception has been raised.

The second way to monitor relevant variables is to let the program run all the way through and collect data on the variables for every trading day without a position. Going this route, I can add code to write to btstats per usual. Skipped days would be evident as consecutive rows labeled ‘INCEPTION.’ I can also scroll down to the particular dates I anticipate this to occur, where I would expect to see mostly zeros with all variables reset.

I’m not sure btstats contains all the relevant variables I want to monitor, which I need to think carefully about to determine. Should this be true, I can always create another dataframe with relevant variables as column names.

The next thing I need to figure out is where in the program I need to call for the relevant variables to be logged. I feel confident to say it should come when control_flag is ‘find_spread.’ That is, after all, when the program is iterating through days without any positions. I also think it makes sense to log relevant variables when spread legs are not encoded.

I will continue next time.