Out must have the same dimensions as the input. The best way to exclude securities with missing data (NaNs) is by dropping rows from the output DataFrame, after calling pipeline_output('my_pipeline'), using the dropna() method. Here is the documentation for it: (http://pandas.pydata.org/pandas-docs/stable/generated/pandas.DataFrame.dropna.html).

Thanks Ernesto -

That's kinda what I figured. For the factor I use in the example I attached above, I'll have to drop the securities that don't have full price history windows before running the factor on the data, and then add them back in at the output as NaNs. Presumably alphalens will then ignore the NaNs?

Also, note that I cannot view the entire notebook, when I click View Notebook. Or maybe the entire notebook did not get attached?

The basic operation within the custom factor is:

1. Identify securities that have NaNs in their windows.
   
2. Compute factor values for remaining securities (without any NaNs in their windows).
   
3. Build output vector of all securities, with NaNs and values assigned appropriately.

I think I found some numpy functions that will do the trick here (without writing nasty loops over securities). However, if there are examples out there, or I'm approaching this incorrectly, please let me know.

Here's the factor I'm working on:

class OptRev(CustomFactor):  
    inputs = [USEquityPricing.close]  
    window_length = 5  
    def compute(self, today, assets, out, close):  
        x_tilde = np.nan_to_num(np.mean(close,axis=0)/close[-1,:])  
        m = len(x_tilde)  
        b_t = 1.0*np.ones(m)/m  
        x = cvx.Variable(m)  
        objective = cvx.Minimize(cvx.sum_entries(cvx.square(x-b_t)))  
        constraints = [cvx.sum_entries(x) == 1, cvx.sum_entries(x_tilde*x) >= 1, x >= 0]  
        prob = cvx.Problem(objective, constraints)  
        prob.solve()  
        out[:] = np.squeeze(np.asarray(x.value))  

Here's what I ended up doing:

class OptRev(CustomFactor):  
    inputs = [USEquityPricing.close]  
    window_length = 5  
    def compute(self, today, assets, out, close):  
        x_tilde = np.mean(close,axis=0)/close[-1,:]  
        out_temp = np.copy(x_tilde)  
        n_ok = np.argwhere(np.isfinite(x_tilde))  
        x_tilde = x_tilde[np.isfinite(x_tilde)]  
        m = len(x_tilde)  
        b_t = 1.0*np.ones(m)/m  
        x = cvx.Variable(m)  
        objective = cvx.Minimize(cvx.sum_entries(cvx.square(x-b_t)))  
        constraints = [cvx.sum_entries(x) == 1, cvx.sum_entries(x_tilde*x) >= 1, x >= 0]  
        prob = cvx.Problem(objective, constraints)  
        prob.solve()  
        out[:] = np.put(out_temp,n_ok,np.squeeze(np.asarray(x.value)))