Incorrect volume with consolidator

Hi Keith,

Did you make sure that the same bars are considered in both cases? Like making sure the last one is not considered in both or is considered in both…

Thinking out loud trying to find a reason for your observation…

Fred

Hi Fred,

Thank you for your response. I do not think that is the problem. I have logged the last 5 min bar and it does not seem to related to other bars. I believe it could be something about having the symbol data in a class and having the consolidator in this class. However, this approach appears to the typical way of dealing with consolidator with multiple symbols. I noticed this problem when I have been working on a more complex alpha model and I was able to reproduce this error with this simple backtest. Any help would be greatly appreciated. 

Thanks,

Keith

“I believe it could be something about having the symbol data in a class and having the consolidator in this class.”

I haven't done testing myself yet on this but it's not the first time I discuss this. With Mak, for instance, we've been kinda leaning towards that conclusion too. There's either a bug around that in LEAN or there's something we don't understand/see.

I don't have an answer right now. But I will eventually find the time to do some small test, designed to pinpoint only that and thus see what's wrong if anything is.

Fred

Hi Fred, 

I think I have found a major issue when calling history and consolidator. In the backtest below, I call the past history and populate a rolling window of 5 min bars in the symbol data class. I also used the consolidator during the backtest. I would think that 5 min bars in the symbol data class would be the same as the 5 min bars in the backtest with bars being indexed on the endtime of the bar. However, the history and consolidator appears to be indexed on the start time of the bar and not endtime.   For example in the SymbolData class history, the 1 min bar for TSLA at 2020-08-18 with an endtime of 12:52 is the following: 

1 min bar:  TSLA: O: 378.532 H: 379.074 L: 378.524 C: 378.972 V: 78795 start of bar: 2020-08-18 12:51:00 EndTime of bar: 2020-08-18 12:52:00

However when history is called during the backtest, this bar is labeled as 12:51 and not 12:52 (where history data is indexed by endtime and not start) 

                    askclose  askhigh   asklow  askopen  asksize  bidclose  bidhigh   bidlow  bidopen  bidsize    close     high      low     open    volume
time                                                                                                                                                         
2020-08-18 12:50:00   378.536  379.534  378.438  379.400   1500.0   378.400  379.390  378.210  379.084    500.0  378.490  379.520  378.256  379.300  133940.0
2020-08-18 12:51:00   378.972  379.090  378.536  378.536    500.0   378.800  378.980  378.400  378.400   1500.0  378.972  379.074  378.524  378.532   78795.0
2020-08-18 12:52:00   379.004  379.200  378.886  378.972   1000.0   378.700  379.012  378.520  378.800    500.0  378.716  379.180  378.710  378.900   35705.0
 

I had to run the backtest twice and change the start date to get the data above. However this does not resolve the volume issue identify above.  How can the 5 min bars align so that 5 min bars are the same in both the symbolData class and backtest consolidator? 

Thank you, 
Keith

Yep. I do think that the start time vs end time indexing is an issue. I don't know if this is only for Tesla - it can be.

Please report the Tesla issue here: 

I will also reference this thread to the devs (in case they don't see it).

I still don't have an answer re the original issue with consolidators/universe as I did not do the test I was writing about. I'll do it asap and report back here.

Fred

Hi Fred, 
I think it is a systemic issue: It is the same for HES, for example 1 min bar with endtime of 12:52 

bar from symbolData class:

2020-08-20 00:00:00 1 min bar:  HES: O: 48.00951 H: 48.00951 L: 47.96554 C: 47.96554 V: 2572 start of bar: 2020-08-19 12:51:00 EndTime of bar: 2020-08-19 12:52:00

Called history during backtest: 
                    askclose    askhigh     asklow   askopen  asksize  bidclose    bidhigh     bidlow   bidopen  bidsize      close       high        low       open  volume
time                                                                                                                                                                         
2020-08-19 12:50:00  48.03882  48.077899  48.019280  48.04859    307.0  48.01928  48.058359  48.009510  48.03882    205.0  48.038820  48.058359  48.038820  48.043705  3903.0
2020-08-19 12:51:00  47.97043  48.038820  47.970430  48.03882    205.0  47.95089  48.019280  47.950890  48.01928    102.0  47.965545  48.009510  47.965545  48.009510  2572.0
 

For spy, for example 1 min with endtime 12:51:

2020-08-20 00:00:00 1 min bar:  SPY: O: 331.8704 H: 331.9194 L: 331.8704 C: 331.9145 V: 57464 start of bar: 2020-08-19 12:50:00 EndTime of bar: 2020-08-19 12:51:00

               askclose     askhigh      asklow     askopen  asksize    bidclose     bidhigh      bidlow     bidopen  bidsize       close        high         low        open   volume
time                                                                                                                                                                                          
2020-08-19 12:50:00  331.919384  331.919384  331.870447  331.880235    102.0  331.909597  331.909597  331.860660  331.860660   1635.0  331.914490  331.919384  331.870447  331.870447  57464.0
2020-08-19 12:51:00  331.909597  331.929171  331.899809  331.919384    102.0  331.899809  331.919384  331.890022  331.909597   2963.0  331.909597  331.919384  331.899809  331.919384  17486.0
2020-08-19 12:52:00  331.890022  331.919384  331.880235  331.909597   1124.0  331.880235  331.909597  331.870447  331.899809    409.0  331.880235  331.914490  331.880235  331.909597  32483.0

Thanks, 
Keith

Hi Fred, 

I think I realized part of the problem with the indexing. The issue is associated with the for loop in the symbolData class (not the history call). 

        for time, row in self.df_min.iterrows():
           tradebarbar = TradeBar(time, self.Symbol, row.open, row.high, row.low, row.close, row.volume)
           self.rolling_windowMin.Add(tradebarbar)
           self.consolidatorMins.Update(tradebarbar)

When the tradebarbar is constructed, it is considering the time as the start time of the bar instead of the endtime as it is in the DataFrame Index. Is there a way to have TrabeBar consider time as the end time and not the start?

Thank you,

Keith

Hi Keith,

TradeBar's first param is the start time. And the last param is the period. TradeBar calculates EndTime by doing start time + period.

When you construct your TradeBar there, you do not pass the last param (period). So you are using some default value. Not sure what it is. Probably 1 minute as that is the default resolution pretty much everywhere.

Yeah, just looked at the code and it is 1 minute by default. So if the start time is given as first param, since you are at minute res, that in particular is fine.

Fred

Hi again,

Not sure the implication was clear so here I am again: if you have the end time and you want to create a TradeBar, you need to substract the period to the end time to get the start time and pass that as first param…

Fred

Hi Fred, 

Great thank. Yes, I modified the loop to be:

        for time, row in self.df_min.iterrows():
           tradebarbar = TradeBar(time-timedelta(minutes= 1), self.Symbol, row.open, row.high, row.low, row.close, row.volume )
           self.rolling_windowMin.Add(tradebarbar)
           self.consolidatorMins.Update(tradebarbar) 

This resolved the indexing issues. Unfortunately, this did not resolve the volume consolidator issues when running the backtest. Does the consolidator need to be updated when the backtest is running as it is the symbolData class?  

Thank you, 

Keith

Hi Keith,

Once the consolidator is registered, it should be handled by LEAN no matter where in the code it was registered. So no, it should not need to be updated because it is in SymbolData. LEAN is not aware of SymbolData.

BUT

Still did not have time to do my testing. I do have my little finger telling me there's a problem with consolidators and universe selection. Either a bug OR some explanation on how to do it right.

Fred