NaN Handling

[pcgm-team]

No description provided.

[enzbus]

Hello! I'm not sure I understand, sorry. Data that is fed to the optimization routines can never have np.nan. Historical market data is managed, in this case, by UserProvidedMarketData. Its default behavior is to look at the returns dataframe and at every time step only include in the back-test (or online execution) names that have non-nan return for the period and have had non-nan returns for at least a min_history time period. The interface is defined in the MarketData base class, so if you want a different behavior you can derive from there (it only has 3 public methods). The default one defines its logic in MarketDataInMemory so you may as well redefine the one or two methods that have to do with the asset selection (I may make it easier to do in the future). MarketSimulator also has logic to decide what to do if an asset is dropped during a back-test, the default behavior is to liquidate (transform in cash) without transaction cost, assuming it's an M&A kind of situation. In general the default behavior are designed around daily trading of stocks in the US, but any other can be achieved by deriving from the default objects.

[enzbus]

Reopening this because I need to improve docs on it.

[enzbus]

You should post the stack trace, there are many pieces that could throw a nan error. values_in_time is the method called by all elements of a policy at each point in time of a back-test. If you're sure the error comes from the ReturnsForecast then it could be a missing or misaligned timestamp, or a missing asset. Do you have constant or varying universe?