QuantConnect Community

Hi Karel, 

the ExecutionModel have one required method to implement Execute(algorithm, targets). How many times Execute() is called depends on your PortfolioConstructionModel, i.e. Execute() is called when it receives new/modified targets from your PortfolioConstructionModel. 

It appears to me that your approach would effect a violation of the separation of concerns principle. The timing aspect of your signals/orders should be located in your AlphaModel. 

Regarding the other questions: I am not sure but I guess the algorithm will keep the open orders in a pending status. Maybe someone of the QC staff can tell us more on this. 

Hi Arthur,

Well, the ExecutionModel I want to build is based on some machine learning principles whereby it will consume Minute bar data to choose the ideal moment to execute a trade, wheres the AlphaModel will consume Day bar data on the same Symbol to decide when to enter a trade. The ExecutionModel does not have to provide an Insight, just execute wat was decided before, but it must be able to do so using its own insights. I more or less consider Alpha and Execution to be 2 different agents, as in unsupervided/re-inforced machine learning speak.

You can consider AlphaModel and ExecutionModel as two different agents, but they're not independent here as the data will be piped sequentially through the Algorithm Framework modules, i.e. :

UniverseSelection --> Alpha --> PortfolioConstruction --> RiskManagement --> Execution

>> The ExecutionModel does not have to provide an Insight, just execute wat was decided before, but it must be able to do so using its own insights. <<

What do you mean by "using its own inisghts"? Can you provide an example? 

>> the ExecutionModel I want to build is based on some machine learning principles whereby it will consume Minute bar data to choose the ideal moment to execute a trade, wheres the AlphaModel will consume Day bar data on the same Symbol [...] <<

If you need Minute Resolution data in your algorithm then you have to subscribe to Minute Resolution anyway for the corresponding symbols. Accordingly, the method Update() in your AlphaModel will be called every minute. On this point at least, it shouldn't be a problem to move the code to the AlphaModel. 

Nevertheless, there might be cases which are not suited for the Algorithm Framework, and, if so, you should not hesitate to use the classic style (with OnData() as primary entry point). 

Well, my plan is to have Alpha (or QCA) to call a new method ExecutionModel.Update() or alike, everytime Update() (onData()) is called. 

ExecutionModel would contain a unsupervised agent with a neural net that takes a few inputs, but also the Minute Bars and do some time-series related forecasting/calculation so that it the agent will be rewarded for execution for example a sell order at the high of a day/..., and penalised if it is not able to execute the trade before the close of the market